diff --git a/simulation/ai-polygon/Install-RealtimeWorker.ps1 b/simulation/ai-polygon/Install-RealtimeWorker.ps1
new file mode 100644
index 0000000..37f39ce
--- /dev/null
+++ b/simulation/ai-polygon/Install-RealtimeWorker.ps1
@@ -0,0 +1,141 @@
+param(
+    [Parameter(Mandatory=$true)][ValidatePattern('^[a-f0-9]{16}$')][string]$Release,
+    [switch]$Remove
+)
+$ErrorActionPreference='Stop'
+$ProgressPreference='SilentlyContinue'
+[Console]::OutputEncoding=[System.Text.Encoding]::UTF8
+$root='D:\NDC_MISSIONCORE\runtime\simulation'
+$taskName='ndc-ai-polygon-worker'
+$isaac=Join-Path $root 'isaac-sim-6.1.0'
+$python=Join-Path $isaac 'kit\python\python.exe'
+$worker=Join-Path $root "releases\ai-polygon-$Release\simulation\ai-polygon\realtime_worker.py"
+$firewallBackup=Join-Path $root 'private\stream-firewall-backup.json'
+if (Test-Path (Join-Path $root 'state\active.json')) {
+    throw 'Finish or reconcile the active episode before changing the service.'
+}
+if ($Remove) {
+    if (Get-ScheduledTask -TaskName $taskName -ErrorAction SilentlyContinue) {
+        Stop-ScheduledTask -TaskName $taskName
+        Unregister-ScheduledTask -TaskName $taskName -Confirm:$false
+    }
+    foreach ($name in @('ndc-ai-polygon-signal','ndc-ai-polygon-media')) {
+        Get-NetFirewallRule -Name $name -ErrorAction SilentlyContinue | Remove-NetFirewallRule
+    }
+    if (Test-Path $firewallBackup) {
+        foreach ($rule in (Get-Content $firewallBackup -Raw | ConvertFrom-Json)) {
+            Get-NetFirewallRule -Name $rule.Name -ErrorAction Stop | Set-NetFirewallRule -Action Block -LocalPort $rule.LocalPort
+            Get-NetFirewallRule -Name $rule.ExceptionName -ErrorAction SilentlyContinue | Remove-NetFirewallRule
+        }
+    }
+    return
+}
+foreach ($file in @($python,$worker,(Join-Path $root 'private\worker.token'))) {
+    if (!(Test-Path $file)) { throw "Missing prepared payload: $file" }
+}
+$payloadRoot=Join-Path $root "releases\ai-polygon-$Release"
+$manifestPath=Join-Path $root "releases\ai-polygon-$Release.manifest.json"
+if (!(Get-FileHash $manifestPath -Algorithm SHA256).Hash.ToLower().StartsWith($Release)) {
+    throw 'Manifest identity mismatch'
+}
+$manifest=Get-Content $manifestPath -Raw | ConvertFrom-Json
+$verified=0
+foreach($entry in $manifest.PSObject.Properties) {
+    $file=[IO.Path]::GetFullPath((Join-Path $payloadRoot $entry.Name))
+    if (!$file.StartsWith($payloadRoot+'\',[StringComparison]::OrdinalIgnoreCase) -or
+        $entry.Value -notmatch '^[a-f0-9]{64}$' -or
+        (Get-FileHash $file -Algorithm SHA256).Hash.ToLower() -ne $entry.Value) {
+        throw "Payload identity mismatch: $($entry.Name)"
+    }
+    $verified++
+}
+if($verified -lt 1) { throw 'Empty release manifest' }
+$address=(& 'C:\Program Files\Tailscale\tailscale.exe' ip -4).Trim()
+$clientAddress=($env:SSH_CONNECTION -split ' ')[0]
+foreach ($value in @($address,$clientAddress)) {
+    $ip=[Net.IPAddress]::Parse($value)
+    $bytes=$ip.GetAddressBytes()
+    if ($bytes.Length -ne 4 -or $bytes[0] -ne 100 -or $bytes[1] -lt 64 -or $bytes[1] -gt 127) {
+        throw 'This installer admits only the existing private Tailscale path.'
+    }
+}
+$kit=Join-Path $isaac 'kit\python\kit.exe'
+# Windows' user-prompt Block rules override scoped Allow rules. Keep those rules
+# blocking every other port; for the two admitted ports retain an explicit Block
+# for every remote address EXCEPT the current private operator. Preserve rollback.
+$oldRules=@()
+if (Test-Path $firewallBackup) { $oldRules=@(Get-Content $firewallBackup -Raw | ConvertFrom-Json) }
+$clientBytes=([Net.IPAddress]::Parse($clientAddress)).GetAddressBytes()
+# Windows rejects prefix /0 here. These disjoint CIDRs are the exact IPv4
+# complement of the operator /32, plus both IPv6 halves. No address is exposed.
+$number=[uint64]$clientBytes[0]*16777216+[uint64]$clientBytes[1]*65536+[uint64]$clientBytes[2]*256+$clientBytes[3]
+$excluded=@('::/1','8000::/1')
+for($prefix=1;$prefix -le 32;$prefix++) {
+    $bit=[uint64][math]::Pow(2,32-$prefix)
+    $network=([uint64]([math]::Floor($number/$bit)*$bit)) -bxor $bit
+    $networkBytes=[byte[]]@((($network -shr 24) -band 255),(($network -shr 16) -band 255),(($network -shr 8) -band 255),($network -band 255))
+    $excluded+="$([Net.IPAddress]::new($networkBytes))/$prefix"
+}
+try {
+$blocked=Get-NetFirewallRule | Where-Object {
+    $_.Enabled -eq 'True' -and $_.Direction -eq 'Inbound' -and
+    ($_.Action -eq 'Block' -or $_.Name -in @($oldRules.Name)) -and
+    $_.Name -notlike 'ndc-*' -and (($_ | Get-NetFirewallApplicationFilter).Program -ieq $kit)
+}
+foreach($rule in $blocked) {
+    $filter=$rule | Get-NetFirewallPortFilter
+    $port=if($filter.Protocol -eq 'TCP'){49100}elseif($filter.Protocol -eq 'UDP'){47998}else{throw 'Unexpected Isaac block protocol'}
+    $exceptionName="ndc-ai-polygon-block-other-$($filter.Protocol.ToLower())"
+    $backup=$oldRules | Where-Object {$_.Name -eq $rule.Name}
+    if(!$backup) {
+        if($filter.LocalPort -ne 'Any' -or ($rule | Get-NetFirewallAddressFilter).RemoteAddress -ne 'Any') {
+            throw 'Inspect a nonstandard Isaac firewall policy before migrating it.'
+        }
+        $oldRules+=@{Name=$rule.Name;LocalPort='Any';ExceptionName=$exceptionName}
+        $oldRules | ConvertTo-Json -Depth 4 | Set-Content -Encoding UTF8 $firewallBackup
+    }
+    Get-NetFirewallRule -Name $exceptionName -ErrorAction SilentlyContinue | Remove-NetFirewallRule
+    New-NetFirewallRule -Name $exceptionName -DisplayName $exceptionName -Group 'ndc-ai-polygon' `
+        -Direction Inbound -Action Block -Program $kit -Protocol $filter.Protocol -LocalPort $port `
+        -RemoteAddress $excluded -Profile $rule.Profile | Out-Null
+    $rule | Set-NetFirewallRule -Action Block -LocalPort @("1-$($port-1)","$($port+1)-65535")
+}
+foreach ($spec in @(@('ndc-ai-polygon-signal','TCP',49100),@('ndc-ai-polygon-media','UDP',47998))) {
+    $existing=Get-NetFirewallRule -Name $spec[0] -ErrorAction SilentlyContinue
+    if ($existing) { $existing | Remove-NetFirewallRule }
+    New-NetFirewallRule -Name $spec[0] -DisplayName $spec[0] -Group 'ndc-ai-polygon' `
+        -Direction Inbound -Action Allow -Program $kit -Protocol $spec[1] -LocalPort $spec[2] `
+        -LocalAddress $address -RemoteAddress $clientAddress -Profile Any | Out-Null
+}
+} catch {
+    foreach($old in $oldRules) {
+        Get-NetFirewallRule -Name $old.Name -ErrorAction Stop | Set-NetFirewallRule -Action Block -LocalPort $old.LocalPort
+        Get-NetFirewallRule -Name $old.ExceptionName -ErrorAction SilentlyContinue | Remove-NetFirewallRule
+    }
+    foreach($name in @('ndc-ai-polygon-signal','ndc-ai-polygon-media')) {
+        Get-NetFirewallRule -Name $name -ErrorAction SilentlyContinue | Remove-NetFirewallRule
+    }
+    throw
+}
+if (Get-ScheduledTask -TaskName $taskName -ErrorAction SilentlyContinue) {
+    Stop-ScheduledTask -TaskName $taskName
+}
+$arguments="-u `"$worker`" --core http://127.0.0.1:18080 --token-file `"$root\private\worker.token`" --state `"$root\state`" --isaac `"$isaac`" --stream-address $address"
+$action=New-ScheduledTaskAction -Execute $python -Argument $arguments -WorkingDirectory $root
+$user=[Security.Principal.WindowsIdentity]::GetCurrent().Name
+$principal=New-ScheduledTaskPrincipal -UserId $user -LogonType Interactive -RunLevel Limited
+$trigger=New-ScheduledTaskTrigger -AtLogOn -User $user
+$settings=New-ScheduledTaskSettingsSet -MultipleInstances IgnoreNew -RestartCount 3 `
+    -RestartInterval (New-TimeSpan -Minutes 1) -ExecutionTimeLimit ([TimeSpan]::Zero) `
+    -AllowStartIfOnBatteries -DontStopIfGoingOnBatteries -StartWhenAvailable
+Register-ScheduledTask -TaskName $taskName -Action $action -Principal $principal `
+    -Trigger $trigger -Settings $settings -Force | Out-Null
+Start-ScheduledTask -TaskName $taskName
+[ordered]@{schema_version='missioncore.ai-polygon-service/v1'; release=$Release;
+    verified_files=$verified;
+    installed_at=[DateTime]::UtcNow.ToString('o'); task=$taskName;
+    lifecycle='Windows task, at user logon; independent of SSH/viewer';
+    control='existing operator tunnel 18080'; video='private Tailscale only';
+    rollback='Install-RealtimeWorker.ps1 -Release <previous> or -Remove; requires no active episode'
+} | ConvertTo-Json | Set-Content -Encoding UTF8 (Join-Path $root 'realtime-service-installation.json')
+Get-Content (Join-Path $root 'realtime-service-installation.json')
diff --git a/simulation/ai-polygon/Prepare-IsaacRuntime.ps1 b/simulation/ai-polygon/Prepare-IsaacRuntime.ps1
new file mode 100644
index 0000000..ee2b497
--- /dev/null
+++ b/simulation/ai-polygon/Prepare-IsaacRuntime.ps1
@@ -0,0 +1,38 @@
+param([switch]$Extract)
+$ErrorActionPreference = 'Stop'
+$ProgressPreference = 'SilentlyContinue'
+# Additive workstation package; never changes Docker, drivers, PATH or services.
+$root = 'D:\NDC_MISSIONCORE\runtime\simulation'
+$archive = Join-Path $root 'isaac-sim-standalone-6.1.0-windows-x86_64.zip'
+$target = Join-Path $root 'isaac-sim-6.1.0'
+$url = 'https://downloads.isaacsim.nvidia.com/isaac-sim-standalone-6.1.0-windows-x86_64.zip'
+$vendorMd5 = 'a07968e980072c9ca27b2166443e2d89'
+New-Item -ItemType Directory -Force $root | Out-Null
+if ((Get-PSDrive D).Free -lt 60GB) { throw 'Isaac preparation requires 60 GiB free on D:' }
+if (!(Test-Path $archive)) {
+    & curl.exe --fail --location --retry 3 --connect-timeout 20 --max-time 7200 --continue-at - --output "$archive.part" $url
+    if ($LASTEXITCODE -ne 0) { throw 'Download interrupted; partial archive retained for resume.' }
+    if ((Get-FileHash "$archive.part" -Algorithm MD5).Hash.ToLower() -ne $vendorMd5) {
+        throw 'Archive does not match the NVIDIA download-page checksum.'
+    }
+    Move-Item "$archive.part" $archive
+}
+if ((Get-FileHash $archive -Algorithm MD5).Hash.ToLower() -ne $vendorMd5) { throw 'Archive checksum changed.' }
+if ($Extract -and !(Test-Path (Join-Path $target 'python.bat'))) {
+    if (Test-Path $target) { throw 'Incomplete target exists; inspect it before resuming extraction.' }
+    New-Item -ItemType Directory $target | Out-Null
+    & tar.exe -xf $archive -C $target
+    if ($LASTEXITCODE -ne 0) { throw 'Extraction failed; files retained for diagnosis.' }
+}
+[ordered]@{
+    schema_version = 'missioncore.ai-polygon-isaac-preparation/v1'
+    prepared_at = [DateTime]::UtcNow.ToString('o')
+    source_url = $url
+    vendor_md5 = $vendorMd5
+    sha256 = (Get-FileHash $archive -Algorithm SHA256).Hash.ToLower()
+    byte_length = (Get-Item $archive).Length
+    target = $target
+    extracted = (Test-Path (Join-Path $target 'python.bat'))
+    launched = $false
+} | ConvertTo-Json | Set-Content -Encoding UTF8 (Join-Path $root 'isaac-6.1.0-preparation.json')
+Get-Content (Join-Path $root 'isaac-6.1.0-preparation.json')
diff --git a/simulation/ai-polygon/Prepare-RobotAssets.ps1 b/simulation/ai-polygon/Prepare-RobotAssets.ps1
new file mode 100644
index 0000000..badd5b3
--- /dev/null
+++ b/simulation/ai-polygon/Prepare-RobotAssets.ps1
@@ -0,0 +1,33 @@
+$ErrorActionPreference = 'Stop'
+$ProgressPreference = 'SilentlyContinue'
+$root = 'D:\NDC_MISSIONCORE\runtime\simulation\assets\jetbot-6.1-v1'
+$origin = 'https://omniverse-content-production.s3-us-west-2.amazonaws.com/Assets/Isaac/6.1/Isaac/Robots_Multiphysics/NVIDIA/Jetbot/'
+$files = @('jetbot.usda', 'payloads/base.usda', 'payloads/instances.usda', 'payloads/robot.usda', 'payloads/materials.usda', 'payloads/geometries.usd', 'payloads/Physics/physics.usda', 'payloads/Physics/physx.usda')
+$manifest = Join-Path $root 'asset-manifest.json'
+if (Test-Path $manifest) {
+    $existing = Get-Content $manifest -Raw | ConvertFrom-Json
+    foreach ($file in $existing.files) {
+        if ((Get-FileHash (Join-Path $root $file.path) -Algorithm SHA256).Hash.ToLower() -ne $file.sha256) { throw 'Prepared Jetbot asset changed' }
+    }
+    Get-Content $manifest
+    exit 0
+}
+New-Item -ItemType Directory $root -Force | Out-Null
+$rows = @()
+$total = 0
+foreach ($relative in $files) {
+    $target = Join-Path $root $relative
+    New-Item -ItemType Directory (Split-Path $target) -Force | Out-Null
+    if (!(Test-Path $target)) {
+        & curl.exe --fail --silent --show-error --location --connect-timeout 10 --max-time 180 --max-filesize 536870912 --retry 2 --continue-at - --output "$target.part" ($origin + $relative)
+        if ($LASTEXITCODE -ne 0) { throw "Jetbot download failed: $relative" }
+        Move-Item "$target.part" $target
+    }
+    $length = (Get-Item $target).Length
+    $total += $length
+    if ($total -gt 536870912) { throw 'Jetbot package exceeds 512 MiB budget' }
+    $rows += [ordered]@{path=$relative; source_url=($origin+$relative); byte_length=$length; sha256=(Get-FileHash $target -Algorithm SHA256).Hash.ToLower()}
+}
+# OmniPBR.mdl and OmniGlass.mdl are bundled renderer materials, not remote assets.
+[ordered]@{schema_version='missioncore.ai-polygon-robot-assets/v1'; version='isaac-6.1-jetbot'; files=$rows; byte_length=$total; prepared_at=[DateTime]::UtcNow.ToString('o')} | ConvertTo-Json -Depth 5 | Set-Content -Encoding UTF8 $manifest
+Get-Content $manifest
diff --git a/simulation/ai-polygon/Prepare-Scene.ps1 b/simulation/ai-polygon/Prepare-Scene.ps1
new file mode 100644
index 0000000..1a12f63
--- /dev/null
+++ b/simulation/ai-polygon/Prepare-Scene.ps1
@@ -0,0 +1,16 @@
+param([Parameter(Mandatory=$true)][string]$InputManifest,
+      [Parameter(Mandatory=$true)][string]$OutputPly,
+      [int]$Lod = -1)
+$ErrorActionPreference = 'Stop'
+$ProgressPreference = 'SilentlyContinue'
+$tool = 'D:\NDC_MISSIONCORE\runtime\simulation\tools\splat-transform-3.4.0\node_modules\@playcanvas\splat-transform\bin\cli.mjs'
+if (!(Test-Path $InputManifest)) { throw 'Source archive missing' }
+if (Test-Path $OutputPly) { throw 'Do not overwrite a prepared scene' }
+$tmp = $OutputPly -replace '\.ply$', '.part.ply'
+if ($tmp -eq $OutputPly -or (Test-Path $tmp)) { throw 'Invalid or unfinished output' }
+$options = @('--memory', '--no-tty')
+if ($Lod -ge 0) { $options += @('--select-lod', [string]$Lod) }
+& node.exe $tool @options $InputManifest $tmp
+if ($LASTEXITCODE -ne 0) { throw 'Scene conversion failed' }
+Move-Item $tmp $OutputPly
+@{file=$OutputPly;bytes=(Get-Item $OutputPly).Length;sha256=(Get-FileHash $OutputPly -Algorithm SHA256).Hash.ToLower()} | ConvertTo-Json
diff --git a/simulation/ai-polygon/Prepare-SplatTools.ps1 b/simulation/ai-polygon/Prepare-SplatTools.ps1
new file mode 100644
index 0000000..439b588
--- /dev/null
+++ b/simulation/ai-polygon/Prepare-SplatTools.ps1
@@ -0,0 +1,19 @@
+$ErrorActionPreference = 'Stop'
+$ProgressPreference = 'SilentlyContinue'
+$root = 'D:\NDC_MISSIONCORE\runtime\simulation\tools\splat-transform-3.4.0'
+$source = Join-Path $PSScriptRoot 'splat-tools'
+if ([int]((& node.exe --version).TrimStart('v').Split('.')[0]) -lt 22) { throw 'Node 22 or newer required' }
+New-Item -ItemType Directory $root -Force | Out-Null
+foreach ($name in @('package.json', 'package-lock.json')) {
+    $target = Join-Path $root $name
+    if (Test-Path $target) {
+        if ((Get-FileHash $target).Hash -ne (Get-FileHash (Join-Path $source $name)).Hash) { throw 'Splat tools definition changed' }
+    } else { Copy-Item (Join-Path $source $name) $target }
+}
+Push-Location $root
+try {
+    & npm.cmd ci --ignore-scripts --no-audit --no-fund --cache "$root\npm-cache"
+    if ($LASTEXITCODE -ne 0) { throw 'Splat tools preparation failed' }
+    & node.exe "$root\node_modules\@playcanvas\splat-transform\bin\cli.mjs" --version
+    if ($LASTEXITCODE -ne 0) { throw 'Splat tools validation failed' }
+} finally { Pop-Location }
diff --git a/simulation/ai-polygon/build_bundle.py b/simulation/ai-polygon/build_bundle.py
new file mode 100644
index 0000000..6de5f60
--- /dev/null
+++ b/simulation/ai-polygon/build_bundle.py
@@ -0,0 +1,40 @@
+"""Package current source plus simulation adapters into a digest-addressed archive."""
+
+import argparse
+import hashlib
+import json
+import tarfile
+from pathlib import Path
+
+root = Path(__file__).resolve().parents[2]
+parser = argparse.ArgumentParser()
+parser.add_argument("--output", type=Path, required=True)
+args = parser.parse_args()
+args.output.mkdir(parents=True, exist_ok=True)
+files = sorted(
+    p
+    for folder in (root / "src", root / "simulation/ai-polygon")
+    for p in folder.rglob("*")
+    if p.is_file() and not p.is_symlink() and "__pycache__" not in p.parts
+)
+manifest = {
+    p.relative_to(root).as_posix(): hashlib.sha256(p.read_bytes()).hexdigest() for p in files
+}
+encoded = json.dumps(manifest, sort_keys=True, indent=2).encode()
+identity = hashlib.sha256(encoded).hexdigest()
+path = args.output / ("ai-polygon-" + identity[:16] + ".tgz")
+with tarfile.open(path, "w:gz") as archive:
+    for p in files:
+        archive.add(p, arcname=p.relative_to(root).as_posix(), recursive=False)
+manifest_path = path.with_suffix(".manifest.json")
+manifest_path.write_bytes(encoded)
+print(
+    json.dumps(
+        {
+            "bundle": str(path),
+            "identity": identity,
+            "sha256": hashlib.sha256(path.read_bytes()).hexdigest(),
+            "files": len(files),
+        }
+    )
+)
diff --git a/simulation/ai-polygon/compose.models.yaml b/simulation/ai-polygon/compose.models.yaml
new file mode 100644
index 0000000..f28bfd6
--- /dev/null
+++ b/simulation/ai-polygon/compose.models.yaml
@@ -0,0 +1,83 @@
+name: ndc-ai-polygon-models
+services:
+  segformer:
+    image: sha256:f37681beb9feecee991e71a60a68259f15b56b168c939db3c3cb259c322e9536
+    pull_policy: never
+    container_name: ndc-ai-polygon-segformer
+    restart: "no"
+    gpus: all
+    cpus: 3
+    mem_limit: 4g
+    ports: ["127.0.0.1:18091:8010"]
+    volumes:
+      - type: bind
+        source: ./segformer
+        target: /adapter/segformer
+        read_only: true
+      - type: bind
+        source: D:/NDC_MISSIONCORE/runtime/simulation/assets/segformer-b2-ade/de01bae28967510f9ddd496c60a969357195400c
+        target: /assets
+        read_only: true
+    labels:
+      com.nodedc.product: mission-core
+      com.nodedc.stack: ai-polygon
+      com.nodedc.role: segmentation
+      com.nodedc.managed-by: ai-polygon-worker
+  ddrnet:
+    image: sha256:a3b7d22f5d3bfdf2d84444b936c8b01abf8243be652387d7e2024ba7bda587f5
+    pull_policy: never
+    container_name: ndc-ai-polygon-ddrnet
+    restart: "no"
+    gpus: all
+    shm_size: 1gb
+    ports: ["127.0.0.1:18091:8010"]
+    entrypoint: ["/opt/conda/envs/goose/bin/python", "-B", "/adapter/ddrnet_server.py"]
+    volumes:
+      - type: bind
+        source: ./ddrnet_server.py
+        target: /adapter/ddrnet_server.py
+        read_only: true
+      - type: bind
+        source: D:/NDC_MISSIONCORE/runtime/assets/observatory-portable/lab-v1-static-files-v1
+        target: /assets
+        read_only: true
+    labels:
+      com.nodedc.product: mission-core
+      com.nodedc.stack: ai-polygon
+      com.nodedc.role: segmentation
+      com.nodedc.managed-by: ai-polygon-worker
+  detector:
+    image: sha256:58df7489c3f2276f9591d500a012dee03e23d35543ce3c390b4c001e6bf90794
+    pull_policy: never
+    container_name: ndc-ai-polygon-rfdetr
+    restart: "no"
+    gpus: all
+    shm_size: 1gb
+    ports: ["127.0.0.1:18092:8000"]
+    entrypoint: ["tritonserver"]
+    command:
+      - --model-repository=/models
+      - --model-control-mode=explicit
+      - --load-model=rf_detr_large
+      - --allow-grpc=false
+      - --allow-metrics=false
+      - --pinned-memory-pool-byte-size=16777216
+      - --cuda-memory-pool-byte-size=0:16777216
+    volumes:
+      - type: bind
+        source: D:/NDC_MISSIONCORE/runtime/experiments/m48t-fixed-detector-20260825T095425Z/triton-models/rf_detr_large
+        target: /models/rf_detr_large
+        read_only: true
+    labels:
+      com.nodedc.product: mission-core
+      com.nodedc.stack: ai-polygon
+      com.nodedc.role: detector
+      com.nodedc.managed-by: ai-polygon-worker
+networks:
+  default:
+    name: ndc-ai-polygon-models
+    labels:
+      com.nodedc.product: mission-core
+      com.nodedc.stack: ai-polygon
+      com.nodedc.role: inference
+      com.nodedc.managed-by: ai-polygon-worker
diff --git a/simulation/ai-polygon/connect-worker.sh b/simulation/ai-polygon/connect-worker.sh
new file mode 100644
index 0000000..b7ff5be
--- /dev/null
+++ b/simulation/ai-polygon/connect-worker.sh
@@ -0,0 +1,6 @@
+#!/bin/sh
+set -eu
+# Dedicated simulation channel. Existing Observatory port 18080 is untouched.
+exec ssh -N -T -o BatchMode=yes -o ExitOnForwardFailure=yes \
+  -o ServerAliveInterval=5 -o ServerAliveCountMax=3 \
+  -R 127.0.0.1:18081:127.0.0.1:8000 mission-gpu
diff --git a/simulation/ai-polygon/core_client.py b/simulation/ai-polygon/core_client.py
new file mode 100644
index 0000000..b1fb46a
--- /dev/null
+++ b/simulation/ai-polygon/core_client.py
@@ -0,0 +1,97 @@
+"""Private loopback transport over the operator's dedicated SSH reverse tunnel."""
+
+import hashlib
+import http.client
+import json
+import threading
+import time
+from pathlib import Path
+from urllib.parse import urlsplit
+
+PREFIX = "/api/v1/ai-polygon"
+
+
+class CoreClient:
+    def __init__(self, origin: str, token_file: Path, instance: str):
+        url = urlsplit(origin)
+        if (
+            url.scheme != "http"
+            or url.hostname != "127.0.0.1"
+            or url.path
+            or url.query
+            or url.fragment
+            or url.username
+            or not url.port
+        ):
+            raise ValueError("Core must use an explicit loopback SSH tunnel origin")
+        self.port = url.port
+        self.token = token_file.read_text().strip()
+        if not 32 <= len(self.token) <= 512:
+            raise ValueError("Invalid simulation credential")
+        self.instance = instance
+        self._connections = threading.local()
+
+    def request(self, path: str, body=None):
+        connection = getattr(self._connections, "connection", None)
+        if (
+            connection is not None
+            and time.monotonic() - getattr(self._connections, "last_used", 0) > 2
+        ):
+            connection.close()
+            connection = None
+        if connection is None:
+            connection = http.client.HTTPConnection("127.0.0.1", self.port, timeout=10)
+            self._connections.connection = connection
+        try:
+            connection.request(
+                "GET" if body is None else "POST",
+                PREFIX + path,
+                body=None if body is None else json.dumps(body, allow_nan=False).encode(),
+                headers={
+                    "Authorization": "Bearer " + self.token,
+                    "Worker-Instance": self.instance,
+                    "Content-Type": "application/json",
+                },
+            )
+            response = connection.getresponse()
+            raw = response.read(4 * 1024**2 + 1)
+            if response.status not in (200, 201) or len(raw) > 4 * 1024**2:
+                raise RuntimeError("Core rejected simulation operation: " + str(response.status))
+            self._connections.last_used = time.monotonic()
+            return json.loads(raw)
+        except Exception:
+            connection.close()
+            self._connections.connection = None
+            # Never retry an uncertain mutation automatically.
+            raise
+
+    def download(self, world: dict, target: Path):
+        if target.exists():
+            with target.open("rb") as stream:
+                if hashlib.file_digest(stream, "sha256").hexdigest() == world["sha256"]:
+                    return
+        if world.get("storage", {}).get("kind") == "worker":
+            raise RuntimeError("The admitted Worker-local scene is missing or changed")
+        connection = http.client.HTTPConnection("127.0.0.1", self.port, timeout=30)
+        temporary = target.with_suffix(".part")
+        try:
+            connection.request("GET", PREFIX + "/worlds/" + world["world_id"] + "/source.ply")
+            response = connection.getresponse()
+            if (
+                response.status != 200
+                or int(response.getheader("Content-Length") or 0) != world["byte_length"]
+            ):
+                raise RuntimeError("Scene transfer contract changed")
+            digest, size = hashlib.sha256(), 0
+            with temporary.open("wb") as stream:
+                while block := response.read(4 * 1024**2):
+                    size += len(block)
+                    if size > world["byte_length"]:
+                        raise RuntimeError("Scene transfer exceeds admitted size")
+                    digest.update(block)
+                    stream.write(block)
+            if size != world["byte_length"] or digest.hexdigest() != world["sha256"]:
+                raise RuntimeError("Scene transfer identity changed")
+            temporary.replace(target)
+        finally:
+            connection.close()
diff --git a/simulation/ai-polygon/ddrnet_server.py b/simulation/ai-polygon/ddrnet_server.py
new file mode 100644
index 0000000..46eb2d2
--- /dev/null
+++ b/simulation/ai-polygon/ddrnet_server.py
@@ -0,0 +1,83 @@
+"""Resident reference DDRNet adapter; run only inside the pinned module image.
+
+Raw pinhole RGB is the only input. No device mask, scene labels or actor truth.
+The container publishes this port on Windows loopback only and owns no weights.
+"""
+
+import hashlib
+import importlib.util
+import json
+from http.server import BaseHTTPRequestHandler, HTTPServer
+from pathlib import Path
+
+import numpy as np
+from PIL import Image
+
+
+def checked(path, expected):
+    value = Path(path)
+    if hashlib.sha256(value.read_bytes()).hexdigest() != expected:
+        raise RuntimeError("Pinned DDRNet asset changed: " + value.name)
+    return value
+
+
+def main():
+    checkpoint = checked(
+        "/assets/ddrnet-checkpoint.pth",
+        "b99c2838051bcd7b092fd3970aa62a77d5c0bbb809c9b9afb2ff4b0ebdaa4ee6",
+    )
+    runner = checked(
+        "/assets/ddrnet-goose-runner.py",
+        "b18ad60f277eea69a240a28f290611b94627fb9707faf1bb3e6e22102dad67c1",
+    )
+    spec = importlib.util.spec_from_file_location("polygon_pinned_goose", runner)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    model, _, _ = module.load_model("ddrnet", checkpoint)
+    tensor, _ = module.preprocess(Image.fromarray(np.zeros((600, 800, 3), np.uint8)))
+    module.infer(model, tensor)
+
+    class Handler(BaseHTTPRequestHandler):
+        def setup(self):
+            super().setup()
+            self.connection.settimeout(10)
+
+        def reply(self, status, body, kind="application/octet-stream"):
+            self.send_response(status)
+            self.send_header("Content-Type", kind)
+            self.send_header("Content-Length", str(len(body)))
+            self.end_headers()
+            self.wfile.write(body)
+
+        def do_GET(self):
+            if self.path != "/ready":
+                self.reply(404, b"")
+                return
+            self.reply(200, json.dumps({"model": "ddrnet-goose-pytorch-reference"}).encode())
+
+        def do_POST(self):
+            if self.path != "/infer" or self.headers.get("Content-Length") != "1440000":
+                self.reply(400, b"Expected 800x600 raw RGB uint8")
+                return
+            try:
+                raw = self.rfile.read(1440000)
+                if len(raw) != 1440000:
+                    raise ValueError("Incomplete camera frame")
+                rgb = np.frombuffer(raw, np.uint8).reshape(600, 800, 3)
+                tensor, _ = module.preprocess(Image.fromarray(rgb))
+                mask, _ = module.infer(model, tensor)
+                if mask.shape != (512, 512) or np.any(mask < 0) or np.any(mask >= 64):
+                    raise ValueError("DDRNet output contract changed")
+                self.reply(200, mask.astype(np.uint8).tobytes())
+            except (TimeoutError, ValueError, RuntimeError):
+                self.reply(500, b"DDRNet inference failed")
+
+        def log_message(self, *_):
+            pass
+
+    # External exposure is restricted by the Compose loopback publication.
+    HTTPServer(("0.0.0.0", 8010), Handler).serve_forever()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/simulation/ai-polygon/fetch_lod.py b/simulation/ai-polygon/fetch_lod.py
new file mode 100644
index 0000000..3c29e9b
--- /dev/null
+++ b/simulation/ai-polygon/fetch_lod.py
@@ -0,0 +1,86 @@
+"""Archive full-resolution chunks from an authorized public streamed-SOG scene."""
+
+import argparse
+import gzip
+import hashlib
+import json
+import re
+import subprocess
+import sys
+import urllib.request
+from pathlib import Path
+from urllib.parse import urljoin
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--metadata-url", required=True)
+parser.add_argument("--output", type=Path, required=True)
+parser.add_argument("--author", required=True)
+parser.add_argument("--license", required=True)
+parser.add_argument("--source-url", required=True)
+args = parser.parse_args()
+if not args.metadata_url.startswith("https://") or not args.metadata_url.endswith("/lod-meta.json"):
+    parser.error("Use an observed public HTTPS lod-meta.json URL")
+args.output.mkdir(parents=True, exist_ok=False)
+with urllib.request.urlopen(args.metadata_url, timeout=30) as response:
+    raw = response.read(1024**2 + 1)
+if len(raw) > 1024**2:
+    raise ValueError("LOD manifest too large")
+if raw[:2] == b"\x1f\x8b":
+    raw = gzip.decompress(raw)
+if len(raw) > 1024**2:
+    raise ValueError("LOD manifest too large")
+meta = json.loads(raw)
+if meta["version"] != 1 or not 1 <= meta["counts"][0] <= 20_000_000:
+    raise ValueError("Unsupported scene size or LOD version")
+indices = set()
+
+
+def visit(node):
+    if "0" in node.get("lods", {}):
+        indices.add(node["lods"]["0"]["file"])
+    for child in node.get("children", []):
+        visit(child)
+
+
+visit(meta["tree"])
+if not 1 <= len(indices) <= 100:
+    raise ValueError("Unexpected full-resolution chunk count")
+(args.output / "lod-meta.json").write_bytes(raw)
+for index in sorted(indices):
+    name = meta["filenames"][index]
+    if not re.fullmatch(r"[a-zA-Z0-9_-]+/meta\.json", name):
+        raise ValueError("Unexpected chunk reference")
+    command = [
+        sys.executable,
+        str(Path(__file__).with_name("fetch_sog.py")),
+        "--metadata-url",
+        urljoin(args.metadata_url, name),
+        "--output",
+        str(args.output / Path(name).parent),
+        "--author",
+        args.author,
+        "--license",
+        args.license,
+        "--source-url",
+        args.source_url,
+    ]
+    subprocess.run(command, check=True)
+manifest = {
+    "source_url": args.source_url,
+    "metadata_url": args.metadata_url,
+    "author": args.author,
+    "license": args.license,
+    "lod": 0,
+    "splat_count": meta["counts"][0],
+    "files": [
+        {
+            "path": p.relative_to(args.output).as_posix(),
+            "sha256": hashlib.sha256(p.read_bytes()).hexdigest(),
+            "byte_length": p.stat().st_size,
+        }
+        for p in sorted(args.output.rglob("*"))
+        if p.is_file()
+    ],
+}
+(args.output / "source-manifest.json").write_text(json.dumps(manifest, indent=2))
+print(json.dumps({"count": meta["counts"][0], "chunks": len(indices)}))
diff --git a/simulation/ai-polygon/fetch_sog.py b/simulation/ai-polygon/fetch_sog.py
new file mode 100644
index 0000000..80da1f7
--- /dev/null
+++ b/simulation/ai-polygon/fetch_sog.py
@@ -0,0 +1,86 @@
+"""Archive a licensed, public unbundled SOG; no cookies or account tokens.
+
+Use only a metadata URL observed in an authorized scene. This downloads assets,
+not collision/voxel geometry. The source provenance is carried into the manifest.
+"""
+
+import argparse
+import gzip
+import hashlib
+import json
+import re
+import time
+import urllib.request
+from concurrent.futures import ThreadPoolExecutor
+from datetime import UTC, datetime
+from pathlib import Path
+from urllib.parse import urljoin, urlsplit
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--metadata-url", required=True)
+parser.add_argument("--output", type=Path, required=True)
+parser.add_argument("--author", required=True)
+parser.add_argument("--license", required=True)
+parser.add_argument("--source-url", required=True)
+args = parser.parse_args()
+url = urlsplit(args.metadata_url)
+if url.scheme != "https" or url.username or not url.path.endswith("/meta.json"):
+    parser.error("Use a public HTTPS SOG meta.json URL")
+args.output.mkdir(parents=True, exist_ok=False)
+started = time.monotonic_ns()
+
+
+def fetch(name, limit):
+    target = args.output / name
+    digest = hashlib.sha256()
+    length = 0
+    with (
+        urllib.request.urlopen(urljoin(args.metadata_url, name), timeout=30) as response,
+        target.with_suffix(target.suffix + ".part").open("xb") as stream,
+    ):
+        while data := response.read(1024 * 1024):
+            length += len(data)
+            if length > limit:
+                raise ValueError("SOG asset exceeds bounded download size")
+            stream.write(data)
+            digest.update(data)
+    target.with_suffix(target.suffix + ".part").replace(target)
+    return {"path": name, "byte_length": length, "sha256": digest.hexdigest()}
+
+
+meta = fetch("meta.json", 1024 * 1024)
+raw = (args.output / "meta.json").read_bytes()
+if raw[:2] == b"\x1f\x8b":
+    raw = gzip.decompress(raw)
+    if len(raw) > 1024 * 1024:
+        raise ValueError("SOG metadata exceeds bounded size")
+    (args.output / "meta.json").write_bytes(raw)
+    meta = {"path": "meta.json", "byte_length": len(raw), "sha256": hashlib.sha256(raw).hexdigest()}
+metadata = json.loads(raw)
+if metadata.get("version") != 2 or not 1 <= metadata["count"] <= 20_000_000:
+    raise ValueError("Unsupported SOG version or size")
+files = sorted(
+    {
+        name
+        for value in metadata.values()
+        if isinstance(value, dict)
+        for name in value.get("files", [])
+    }
+)
+if not 1 <= len(files) <= 10 or any(not re.fullmatch(r"[A-Za-z0-9_-]+\.webp", n) for n in files):
+    raise ValueError("SOG contains unexpected asset references")
+with ThreadPoolExecutor(max_workers=3) as pool:
+    rows = list(pool.map(lambda name: fetch(name, 100 * 1024**2), files))
+report = {
+    "schema_version": "missioncore.ai-polygon-source-archive/v1",
+    "source_url": args.source_url,
+    "metadata_url": args.metadata_url,
+    "author": args.author,
+    "license": args.license,
+    "captured_at": datetime.now(UTC).isoformat(),
+    "elapsed_ns": time.monotonic_ns() - started,
+    "splat_count": metadata["count"],
+    "files": [meta, *rows],
+}
+(args.output / "source-manifest.json").write_text(json.dumps(report, indent=2))
+print(json.dumps({"count": report["splat_count"], "bytes": sum(r["byte_length"] for r in rows)}))
diff --git a/simulation/ai-polygon/local_state.py b/simulation/ai-polygon/local_state.py
new file mode 100644
index 0000000..de3d3d8
--- /dev/null
+++ b/simulation/ai-polygon/local_state.py
@@ -0,0 +1,102 @@
+"""Atomic local IPC snapshots; no network access in the simulation process."""
+
+import json
+import os
+import sqlite3
+import time
+from pathlib import Path
+
+
+class StateChannel:
+    """Latest-value local IPC using SQLite WAL, available in both Python hosts.
+
+    Readers never hold Windows rename/delete locks on a telemetry snapshot.
+    The durable camera/motion journals remain separate evidence artifacts.
+    Each process owns its connection; only small control/telemetry JSON enters it.
+    """
+
+    def __init__(self, directory: Path):
+        self.connection = sqlite3.connect(directory / "channel.sqlite3", timeout=0.1)
+        self.connection.execute("PRAGMA journal_mode=WAL")
+        self.connection.execute("PRAGMA synchronous=NORMAL")
+        self.connection.execute(
+            "CREATE TABLE IF NOT EXISTS state (key TEXT PRIMARY KEY, value TEXT NOT NULL)"
+        )
+        self.connection.commit()
+
+    def write(self, key, value):
+        encoded = json.dumps(value, allow_nan=False)
+        with self.connection:
+            self.connection.execute(
+                "INSERT INTO state(key,value) VALUES(?,?) "
+                "ON CONFLICT(key) DO UPDATE SET value=excluded.value",
+                (key, encoded),
+            )
+
+    def read(self, key, default=None):
+        row = self.connection.execute("SELECT value FROM state WHERE key=?", (key,)).fetchone()
+        return default if row is None else json.loads(row[0])
+
+    def close(self):
+        self.connection.close()
+
+
+def sharing_retry(operation):
+    # A concurrent Windows reader briefly denies delete/replace sharing. Retry
+    # only this local file operation, bounded to 14ms; permanent errors surface.
+    for attempt in range(8):
+        try:
+            return operation()
+        except PermissionError:
+            if attempt == 7:
+                raise
+            time.sleep(0.002)
+
+
+def write_json(path: Path, value):
+    temporary = path.with_name(path.name + f".{os.getpid()}.tmp")
+    temporary.write_text(json.dumps(value, allow_nan=False), encoding="utf-8")
+    sharing_retry(lambda: os.replace(temporary, path))
+
+
+def read_json(path: Path, default=None):
+    try:
+        return json.loads(sharing_retry(lambda: read_shared_text(path)))
+    except FileNotFoundError:
+        return default
+
+
+def read_shared_text(path: Path):
+    if os.name != "nt":
+        return path.read_text(encoding="utf-8")
+    # Python's normal Windows open does not grant FILE_SHARE_DELETE. An atomic
+    # writer must remain able to replace a snapshot while a reader holds it.
+    import ctypes
+    import msvcrt
+    from ctypes import wintypes
+
+    kernel = ctypes.WinDLL("kernel32", use_last_error=True)
+    kernel.CreateFileW.argtypes = [
+        wintypes.LPCWSTR,
+        wintypes.DWORD,
+        wintypes.DWORD,
+        ctypes.c_void_p,
+        wintypes.DWORD,
+        wintypes.DWORD,
+        wintypes.HANDLE,
+    ]
+    kernel.CreateFileW.restype = wintypes.HANDLE
+    handle = kernel.CreateFileW(str(path), 0x80000000, 7, None, 3, 0x80, None)
+    if handle == wintypes.HANDLE(-1).value:
+        code = ctypes.get_last_error()
+        if code in (2, 3):
+            raise FileNotFoundError(str(path))
+        raise ctypes.WinError(code)
+    try:
+        descriptor = msvcrt.open_osfhandle(handle, os.O_RDONLY)
+    except BaseException:
+        kernel.CloseHandle.argtypes = [wintypes.HANDLE]
+        kernel.CloseHandle(handle)
+        raise
+    with os.fdopen(descriptor, "r", encoding="utf-8") as stream:
+        return stream.read()
diff --git a/simulation/ai-polygon/model_stack.py b/simulation/ai-polygon/model_stack.py
new file mode 100644
index 0000000..ba01e84
--- /dev/null
+++ b/simulation/ai-polygon/model_stack.py
@@ -0,0 +1,155 @@
+"""Run-owned additive Docker stack. Caller must hold Core's GPU reservation."""
+
+import hashlib
+import http.client
+import json
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent
+sys.path.insert(0, str(ROOT.parents[1] / "src"))
+NAVIGATION_NAME = "ndc-mission-core-ai-module-simulation-cmu-navigation"
+NAMES = (
+    "ndc-ai-polygon-ddrnet",
+    "ndc-ai-polygon-segformer",
+    "ndc-ai-polygon-rfdetr",
+    NAVIGATION_NAME,
+)
+
+
+def sha256(path):
+    with Path(path).open("rb") as stream:
+        return hashlib.file_digest(stream, "sha256").hexdigest()
+
+
+def docker(*args, check=True):
+    return subprocess.run(
+        ["docker", *args], check=check, capture_output=True, text=True, timeout=120
+    )
+
+
+class ModelStack:
+    def __init__(self):
+        self.profile_path = ROOT / "models.worker-006.json"
+        self.profile = json.loads(self.profile_path.read_text())
+        self.ids = []
+
+    def preflight(self):
+        from k1link.simulation.ai_polygon.composition import compose
+
+        compose(ROOT)  # Portable graph admission happens before any Isaac boot.
+        for model in [*self.profile["models"], self.profile["navigation"]]:
+            for key in ("checkpoint", "runner"):
+                if key in model and sha256(model[key]) != model[key + "_sha256"]:
+                    raise RuntimeError("Simulation model identity changed: " + model["id"])
+            actual = docker(
+                "image", "inspect", model["image"], "--format", "{{.Id}}"
+            ).stdout.strip()
+            if actual != model["image"]:
+                raise RuntimeError("Simulation model image is unavailable")
+        if sha256(self.profile["labels"]) != self.profile["labels_sha256"]:
+            raise RuntimeError("GOOSE label table changed")
+
+    def start(self, cancelled=None, on_acquired=None, selection=None):
+        from k1link.simulation.ai_polygon.composition import compose
+
+        graph = compose(ROOT, selection)
+        services = {
+            "simulation-ddrnet-goose": "ddrnet",
+            "simulation-segformer-ade": "segformer",
+            "simulation-rf-detr": "detector",
+        }
+        module_ids = [node.module.module_id for node in graph.nodes]
+        installed = {*services, "simulation-cmu-navigation", "simulation-waypoint-mission"}
+        if set(module_ids) - installed:
+            raise ValueError("Uninstalled simulation provider")
+        self.preflight()
+        for name in NAMES:
+            if docker("container", "inspect", name, check=False).returncode == 0:
+                raise RuntimeError(
+                    "An existing simulation container requires reconciliation: " + name
+                )
+        if cancelled and cancelled():
+            raise InterruptedError("Model startup cancelled")
+        try:
+            selected = [services[module] for module in module_ids if module in services]
+            docker("compose", "-f", str(ROOT / "compose.models.yaml"), "up", "-d", *selected)
+            if "simulation-cmu-navigation" in module_ids:
+                docker(
+                    "run",
+                    "-d",
+                    "--name",
+                    NAVIGATION_NAME,
+                    "--restart",
+                    "no",
+                    "--label",
+                    "com.nodedc.product=mission-core",
+                    "--label",
+                    "com.nodedc.stack=ai-polygon",
+                    "--label",
+                    "com.nodedc.role=ai-module",
+                    "--label",
+                    "com.nodedc.managed-by=ai-polygon-worker",
+                    "--label",
+                    "com.nodedc.composition-sha256=" + graph.sha256,
+                    "--cpus",
+                    "3",
+                    "--memory",
+                    "2g",
+                    "-p",
+                    "127.0.0.1:18093:8010",
+                    "--mount",
+                    f"type=bind,source={ROOT},target=/adapter,readonly",
+                    self.profile["navigation"]["image"],
+                )
+        finally:
+            # Capture immutable IDs even after a partial Compose start.
+            for name in NAMES:
+                result = docker("container", "inspect", name, check=False)
+                if result.returncode == 0:
+                    row = json.loads(result.stdout)[0]
+                    if row["Config"]["Labels"].get("com.nodedc.stack") == "ai-polygon":
+                        self.ids.append(row["Id"])
+            if on_acquired:
+                on_acquired(list(self.ids))
+        deadline = time.monotonic() + 120
+        while time.monotonic() < deadline:
+            if cancelled and cancelled():
+                raise InterruptedError("Model startup cancelled")
+            ready = True
+            for port, path in (
+                (18091, "/ready"),
+                (18092, "/v2/models/rf_detr_large/versions/1/ready"),
+                (18093, "/ready"),
+            ):
+                connection = http.client.HTTPConnection("127.0.0.1", port, timeout=2)
+                try:
+                    connection.request("GET", path)
+                    response = connection.getresponse()
+                    response.read(65536)
+                    ready &= response.status == 200
+                except OSError:
+                    ready = False
+                finally:
+                    connection.close()
+            if ready:
+                return
+            time.sleep(0.5)
+        raise TimeoutError("Simulation models did not become ready")
+
+    def stop(self):
+        failures = []
+        for identity in self.ids:
+            result = docker("container", "rm", "-f", identity, check=False)
+            if (
+                result.returncode
+                and docker("container", "inspect", identity, check=False).returncode == 0
+            ):
+                failures.append(identity)
+        if failures:
+            raise RuntimeError("Simulation GPU resources were not released")
+        self.ids.clear()
+        # Only our now-unused network; Docker refuses removal if another endpoint uses it.
+        docker("network", "rm", "ndc-ai-polygon-models", check=False)
diff --git a/simulation/ai-polygon/models.worker-006.json b/simulation/ai-polygon/models.worker-006.json
new file mode 100644
index 0000000..16ff244
--- /dev/null
+++ b/simulation/ai-polygon/models.worker-006.json
@@ -0,0 +1,52 @@
+{
+  "schema_version": "missioncore.ai-polygon-model-profile/v1",
+  "profile_id": "simulation-pinhole-composable-surfaces-cmu-v2",
+  "camera": {
+    "width": 800,
+    "height": 600,
+    "color": "RGB",
+    "projection": "pinhole"
+  },
+  "authority": "virtual-only",
+  "models": [
+    {
+      "id": "ddrnet-goose-pytorch-reference",
+      "image": "sha256:a3b7d22f5d3bfdf2d84444b936c8b01abf8243be652387d7e2024ba7bda587f5",
+      "checkpoint": "D:/NDC_MISSIONCORE/runtime/assets/observatory-portable/lab-v1-static-files-v1/ddrnet-checkpoint.pth",
+      "checkpoint_sha256": "b99c2838051bcd7b092fd3970aa62a77d5c0bbb809c9b9afb2ff4b0ebdaa4ee6",
+      "preprocess": "pinned GOOSE 600-square crop, nearest 512, RGB /255",
+      "runner": "D:/NDC_MISSIONCORE/runtime/assets/observatory-portable/lab-v1-static-files-v1/ddrnet-goose-runner.py",
+      "runner_sha256": "b18ad60f277eea69a240a28f290611b94627fb9707faf1bb3e6e22102dad67c1"
+    },
+    {
+      "id": "rf_detr_large",
+      "image": "sha256:58df7489c3f2276f9591d500a012dee03e23d35543ce3c390b4c001e6bf90794",
+      "checkpoint": "D:/NDC_MISSIONCORE/runtime/experiments/m48t-fixed-detector-20260825T095425Z/triton-models/rf_detr_large/1/model.plan",
+      "checkpoint_sha256": "986399ce706b7380472cf5e473232249fed6e628971d8007f6609e83128d46b8",
+      "preprocess": "full pinhole frame, resize 704x704, ImageNet normalization; no KB4 mask"
+    },
+    {
+      "id": "segformer-b2-ade150",
+      "image": "sha256:f37681beb9feecee991e71a60a68259f15b56b168c939db3c3cb259c322e9536",
+      "checkpoint": "D:/NDC_MISSIONCORE/runtime/simulation/assets/segformer-b2-ade/de01bae28967510f9ddd496c60a969357195400c/pytorch_model.bin",
+      "checkpoint_sha256": "187ca07bea003a5717c63d04ea90b07f33cd033c0ebf44b4b89fce5070d6c8f3",
+      "revision": "de01bae28967510f9ddd496c60a969357195400c",
+      "preprocess": "RGB centre 600-square crop; pinned processor bilinear 512, ImageNet normalization; logits bilinear 512; candidate confidence >=0.55",
+      "config_sha256": "ee7400840fdb1e5045f0b2eba78bf053df8e33a309c4acec31705a48c8cc5c00",
+      "processor_sha256": "8039d1d210abaa7117ad78e58cdfd6141a2ec72c03dae891b3cd76737e422c6c"
+    }
+  ],
+  "labels": "D:/NDC_MISSIONCORE/runtime/assets/observatory-portable/lab-v1-static-files-v1/ddrnet-goose-mapping.csv",
+  "labels_sha256": "88ae319ba5a3877dd3ae0773f693a6a5fdc283934140de9dfaff029108aefd7f",
+  "navigation": {
+    "id": "cmu-terrain-navigation",
+    "image": "sha256:698228d0d378b166ceaf299aa86ab95b5e74179837532f2378c036c7f7621dcc",
+    "upstream_commit": "158e67b31b644ed1e8b06eb1d7f70e183cc62591"
+  },
+  "default_modules": [
+    "simulation-segformer-ade",
+    "simulation-rf-detr",
+    "simulation-cmu-navigation",
+    "simulation-waypoint-mission"
+  ]
+}
diff --git a/simulation/ai-polygon/motion_control.py b/simulation/ai-polygon/motion_control.py
new file mode 100644
index 0000000..6214702
--- /dev/null
+++ b/simulation/ai-polygon/motion_control.py
@@ -0,0 +1,36 @@
+"""Physics-clock command envelope; perception remains asynchronous.
+
+Limit wheel-surface acceleration without delaying a safety reduction. Scaling
+both sides together preserves the planner's requested curvature. Stops bypass
+the ramp, as required by the already qualified collision/braking envelope.
+This is a simulation actuator, not a VESC controller or a motor calibration.
+"""
+
+CONTROL_PROFILE = {
+    "clock": "physics-pre-step",
+    "rate_hz": 60,
+    "wheel_surface_acceleration_mps2": 0.2,
+    "safety_reductions": "immediate",
+    "authority": "simulation-only",
+}
+
+
+class DriveEnvelope:
+    def __init__(self, track_width=0.9, acceleration=0.2):
+        self.half_track = track_width / 2
+        self.acceleration = acceleration
+        self.wheels = (0.0, 0.0)
+
+    def step(self, velocity, yaw_rate, dt, *, stop=False):
+        requested = (velocity - yaw_rate * self.half_track, velocity + yaw_rate * self.half_track)
+        if stop or max(map(abs, requested)) < 1e-8:
+            self.wheels = (0.0, 0.0)
+            return 0.0, 0.0
+        scale = 1.0
+        for old, new in zip(self.wheels, requested, strict=True):
+            # Reversal starts from zero; never retain motion in the old direction.
+            prior = abs(old) if old * new >= 0 else 0.0
+            if abs(new) > prior:
+                scale = min(scale, (prior + self.acceleration * max(0.0, dt)) / abs(new))
+        self.wheels = tuple(value * scale for value in requested)
+        return velocity * scale, yaw_rate * scale
diff --git a/simulation/ai-polygon/navigation/Dockerfile b/simulation/ai-polygon/navigation/Dockerfile
new file mode 100644
index 0000000..009969f
--- /dev/null
+++ b/simulation/ai-polygon/navigation/Dockerfile
@@ -0,0 +1,34 @@
+FROM ros:humble-ros-base-jammy
+
+LABEL com.nodedc.product="mission-core" \
+      com.nodedc.stack="ai-polygon" \
+      com.nodedc.role="terrain-navigation" \
+      com.nodedc.managed-by="ai-polygon-worker" \
+      org.opencontainers.image.source="https://github.com/HongbiaoZ/autonomous_exploration_development_environment" \
+      org.opencontainers.image.revision="158e67b31b644ed1e8b06eb1d7f70e183cc62591"
+
+SHELL ["/bin/bash", "-o", "pipefail", "-c"]
+RUN apt-get update && apt-get install -y --no-install-recommends \
+      curl ca-certificates build-essential python3-numpy python3-yaml \
+      python3-colcon-common-extensions ros-humble-pcl-ros \
+      ros-humble-pcl-conversions ros-humble-tf2-geometry-msgs \
+      ros-humble-sensor-msgs-py ros-humble-rclpy \
+    && rm -rf /var/lib/apt/lists/*
+
+# Upstream's terrain analysis and collision-free motion primitives are unchanged.
+# Keep the exact sources and their declared BSD package licenses in the image.
+WORKDIR /opt/cmu
+RUN curl -fL --retry 3 \
+      https://codeload.github.com/HongbiaoZ/autonomous_exploration_development_environment/tar.gz/158e67b31b644ed1e8b06eb1d7f70e183cc62591 \
+      -o /tmp/upstream.tgz \
+    && mkdir upstream src \
+    && tar -xzf /tmp/upstream.tgz -C upstream --strip-components=1 \
+    && cp -a upstream/src/local_planner upstream/src/terrain_analysis src/ \
+    && source /opt/ros/humble/setup.bash \
+    && MAKEFLAGS=-j2 colcon build --base-paths src --parallel-workers 1 \
+      --cmake-args -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=OFF \
+    && rm /tmp/upstream.tgz
+
+ENV ROS_LOCALHOST_ONLY=1 ROS_DOMAIN_ID=83
+WORKDIR /adapter
+ENTRYPOINT ["/bin/bash", "-c", "source /opt/ros/humble/setup.bash && source /opt/cmu/install/setup.bash && exec python3 /adapter/navigation/server.py"]
diff --git a/simulation/ai-polygon/navigation/Dockerfile.connectivity b/simulation/ai-polygon/navigation/Dockerfile.connectivity
new file mode 100644
index 0000000..156b18a
--- /dev/null
+++ b/simulation/ai-polygon/navigation/Dockerfile.connectivity
@@ -0,0 +1,8 @@
+FROM ndc-ai-polygon-cmu:rectangle-v3
+COPY terrain_connectivity.cpp /opt/missioncore/terrain_connectivity.cpp
+RUN g++ -O3 -std=c++17 -ffp-contract=off -fno-fast-math -fPIC -shared \
+    /opt/missioncore/terrain_connectivity.cpp -o /opt/missioncore/libterrain_connectivity.so
+LABEL com.nodedc.product="mission-core" \
+      com.nodedc.stack="ai-polygon" \
+      com.nodedc.role="terrain-navigation" \
+      com.nodedc.managed-by="ai-polygon-worker"
diff --git a/simulation/ai-polygon/navigation/Dockerfile.footprint b/simulation/ai-polygon/navigation/Dockerfile.footprint
new file mode 100644
index 0000000..7a640fa
--- /dev/null
+++ b/simulation/ai-polygon/navigation/Dockerfile.footprint
@@ -0,0 +1,16 @@
+# Reuse the already-built pinned upstream image. The caller supplies its exact
+# sha256 through a local immutable build tag; no model images are replaced.
+ARG CMU_BASE
+FROM ${CMU_BASE}
+LABEL com.nodedc.product="mission-core" \
+      com.nodedc.stack="ai-polygon" \
+      com.nodedc.role="terrain-navigation" \
+      com.nodedc.managed-by="ai-polygon-worker" \
+      com.nodedc.navigation-adapter="rectangular-primitive-filter-v3"
+COPY missioncore_footprint.hpp /opt/cmu/src/local_planner/src/missioncore_footprint.hpp
+COPY patch_footprint.py /opt/cmu/patch_footprint.py
+WORKDIR /opt/cmu
+RUN python3 patch_footprint.py && source /opt/ros/humble/setup.bash \
+    && MAKEFLAGS=-j2 colcon build --base-paths src --packages-select local_planner \
+        --parallel-workers 1 --cmake-args -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=OFF
+WORKDIR /adapter
diff --git a/simulation/ai-polygon/navigation/Prepare-Connectivity.ps1 b/simulation/ai-polygon/navigation/Prepare-Connectivity.ps1
new file mode 100644
index 0000000..d031203
--- /dev/null
+++ b/simulation/ai-polygon/navigation/Prepare-Connectivity.ps1
@@ -0,0 +1,17 @@
+param([string]$Root = 'D:\NDC_MISSIONCORE\runtime\simulation')
+$ErrorActionPreference = 'Continue'
+$base = 'sha256:d9b9c842bf8f0bf5208d8568d35f41b69d6e4a9ed74f31dedbcd455f1f7f263f'
+$actual = docker image inspect 'ndc-ai-polygon-cmu:rectangle-v3' --format '{{.Id}}'
+if ($LASTEXITCODE -ne 0 -or $actual.Trim() -ne $base) { throw 'Prepare the pinned rectangular CMU image first' }
+$endpoint = docker context inspect --format '{{.Endpoints.docker.Host}}'
+if ($LASTEXITCODE -ne 0) { throw 'Docker endpoint unavailable' }
+$config = Join-Path $Root 'state\connectivity-public-build'
+New-Item -ItemType Directory -Force $config | Out-Null
+[IO.File]::WriteAllText((Join-Path $config 'config.json'), '{"auths":{"https://index.docker.io/v1/":{}}}')
+$env:DOCKER_CONFIG = $config
+$env:BUILDX_CONFIG = Join-Path $config 'buildx'
+docker --config $config --host $endpoint.Trim() build --progress plain -t ndc-ai-polygon-cmu:connectivity-v1 -f (Join-Path $PSScriptRoot 'Dockerfile.connectivity') $PSScriptRoot
+if ($LASTEXITCODE -ne 0) { throw 'Terrain connectivity build failed' }
+$image = docker --config $config --host $endpoint.Trim() image inspect ndc-ai-polygon-cmu:connectivity-v1 --format '{{.Id}}'
+if ($LASTEXITCODE -ne 0) { throw 'Built image unavailable' }
+@{base_image=$base; image=$image.Trim(); source_sha256=(Get-FileHash (Join-Path $PSScriptRoot 'terrain_connectivity.cpp') -Algorithm SHA256).Hash.ToLower(); installed_at=[DateTime]::UtcNow.ToString('o')} | ConvertTo-Json
diff --git a/simulation/ai-polygon/navigation/Prepare-Navigation.ps1 b/simulation/ai-polygon/navigation/Prepare-Navigation.ps1
new file mode 100644
index 0000000..bf55ccd
--- /dev/null
+++ b/simulation/ai-polygon/navigation/Prepare-Navigation.ps1
@@ -0,0 +1,34 @@
+param([string]$Root = 'D:\NDC_MISSIONCORE\runtime\simulation')
+# Windows PowerShell 5 treats Docker's normal stderr progress as ErrorRecords.
+# Native failures are checked by exit code after each command below.
+$ErrorActionPreference = 'Continue'
+$ProgressPreference = 'SilentlyContinue'
+$tag = 'ndc/mission-core-ai-module-cmu-navigation:158e67b3-v1'
+$receipt = Join-Path $Root 'assets\cmu-navigation-v1.json'
+$existing = docker image inspect $tag --format '{{.Id}}' 2>$null
+if ($LASTEXITCODE -ne 0) {
+    # Public upstream images need no credentials. Windows SSH cannot access the
+    # interactive user's credential helper; isolate only this build's config.
+    $endpoint = docker context inspect --format '{{.Endpoints.docker.Host}}'
+    if ($LASTEXITCODE -ne 0) { throw 'Docker endpoint unavailable' }
+    $buildConfig = Join-Path $PSScriptRoot 'public-build-config'
+    New-Item -ItemType Directory -Force $buildConfig | Out-Null
+    # An explicit anonymous registry prevents CLI auto-discovery of wincred.
+    [IO.File]::WriteAllText((Join-Path $buildConfig 'config.json'), '{"auths":{"https://index.docker.io/v1/":{}}}')
+    $env:DOCKER_CONFIG = $buildConfig
+    $env:BUILDX_CONFIG = Join-Path $buildConfig 'buildx'
+    docker --config $buildConfig --host $endpoint.Trim() build --progress plain -t $tag -f (Join-Path $PSScriptRoot 'Dockerfile') $PSScriptRoot
+    if ($LASTEXITCODE -ne 0) { throw 'CMU navigation image build failed' }
+    $existing = docker image inspect $tag --format '{{.Id}}'
+    if ($LASTEXITCODE -ne 0) { throw 'Built navigation image is unavailable' }
+}
+$value = @{
+    schema_version = 'missioncore.ai-polygon-navigation-install/v1'
+    upstream = 'HongbiaoZ/autonomous_exploration_development_environment'
+    upstream_commit = '158e67b31b644ed1e8b06eb1d7f70e183cc62591'
+    image = $existing.Trim()
+    tag = $tag
+    installed_at = [DateTime]::UtcNow.ToString('o')
+}
+[IO.File]::WriteAllText($receipt, ($value | ConvertTo-Json -Depth 5), [Text.UTF8Encoding]::new($false))
+$value | ConvertTo-Json -Compress
diff --git a/simulation/ai-polygon/navigation/Prepare-RectangleNavigation.ps1 b/simulation/ai-polygon/navigation/Prepare-RectangleNavigation.ps1
new file mode 100644
index 0000000..a2078e2
--- /dev/null
+++ b/simulation/ai-polygon/navigation/Prepare-RectangleNavigation.ps1
@@ -0,0 +1,24 @@
+param([string]$Root = 'D:\NDC_MISSIONCORE\runtime\simulation')
+$ErrorActionPreference = 'Continue'
+$base = 'sha256:7693c28fac83df9e9e6ac6d89ffec81470f660cb8b30fc78fa88be228329dc0b'
+$baseTag = 'ndc-ai-polygon-cmu:upstream-158e67b'
+$tag = 'ndc-ai-polygon-cmu:rectangle-v3'
+$actual = docker image inspect $base --format '{{.Id}}'
+if ($LASTEXITCODE -ne 0 -or $actual.Trim() -ne $base) { throw 'Prepare the pinned upstream CMU image first' }
+docker tag $base $baseTag
+if ($LASTEXITCODE -ne 0) { throw 'Unable to identify the local CMU build base' }
+docker build --build-arg "CMU_BASE=$baseTag" -f (Join-Path $PSScriptRoot 'Dockerfile.footprint') -t $tag $PSScriptRoot
+if ($LASTEXITCODE -ne 0) { throw 'Rectangular candidate filter build failed' }
+$image = docker image inspect $tag --format '{{.Id}}'
+if ($LASTEXITCODE -ne 0) { throw 'Built navigation image unavailable' }
+$value = @{
+    schema_version = 'missioncore.ai-polygon-navigation-install/v1'
+    upstream_commit = '158e67b31b644ed1e8b06eb1d7f70e183cc62591'
+    base_image = $base
+    image = $image.Trim()
+    adapter = 'rectangular-primitive-filter-v3'
+    adapter_sha256 = (Get-FileHash (Join-Path $PSScriptRoot 'missioncore_footprint.hpp') -Algorithm SHA256).Hash.ToLower()
+    installed_at = [DateTime]::UtcNow.ToString('o')
+}
+[IO.File]::WriteAllText((Join-Path $Root 'assets\cmu-navigation-rectangle-v3.json'), ($value | ConvertTo-Json), [Text.UTF8Encoding]::new($false))
+$value | ConvertTo-Json -Compress
diff --git a/simulation/ai-polygon/navigation/Prepare-Terrain.ps1 b/simulation/ai-polygon/navigation/Prepare-Terrain.ps1
new file mode 100644
index 0000000..752cb52
--- /dev/null
+++ b/simulation/ai-polygon/navigation/Prepare-Terrain.ps1
@@ -0,0 +1,74 @@
+param(
+    [Parameter(Mandatory=$true)][string]$WorldFile,
+    [string]$Root = 'D:\NDC_MISSIONCORE\runtime\simulation',
+    [string]$Python = 'D:\NDC_MISSIONCORE\runtime\simulation\isaac-sim-6.1.0\kit\python\python.exe'
+)
+$ErrorActionPreference = 'Stop'
+[Threading.Thread]::CurrentThread.CurrentCulture = [Globalization.CultureInfo]::InvariantCulture
+$world = Get-Content -Raw $WorldFile | ConvertFrom-Json
+if ($world.sha256 -notmatch '^[a-f0-9]{64}$') { throw 'Invalid source hash' }
+$source = Join-Path $Root ('state\worlds\' + $world.sha256 + '.ply')
+if ((Get-FileHash $source -Algorithm SHA256).Hash.ToLower() -ne $world.sha256) { throw 'World source changed' }
+$tool = Join-Path $Root 'tools\splat-transform-3.4.0\node_modules\@playcanvas\splat-transform'
+$package = Get-Content -Raw (Join-Path $tool 'package.json') | ConvertFrom-Json
+if ($package.version -ne '3.4.0') { throw 'SplatTransform version changed' }
+$settings = $world.settings
+$json = ($settings | ConvertTo-Json -Compress) + (Get-FileHash $PSCommandPath -Algorithm SHA256).Hash
+$sha = [Security.Cryptography.SHA256]::Create()
+$digest = [BitConverter]::ToString($sha.ComputeHash([Text.Encoding]::UTF8.GetBytes($json))).Replace('-','').ToLower()
+$out = Join-Path $Root ('assets\terrain-v1\' + $world.sha256 + '-' + $digest.Substring(0,16))
+New-Item -ItemType Directory -Force $out | Out-Null
+$mesh = Join-Path $out 'terrain.collision.glb'
+$migrator = Join-Path $PSScriptRoot 'migrate_terrain_coordinates.py'
+if ((Get-FileHash $migrator -Algorithm SHA256).Hash.ToLower() -ne 'ff08de8b97b9f2b0040b379840c402f5c7c327e3bc1915548d99b6372ff5ca37') { throw 'Coordinate migrator identity changed' }
+if (-not (Test-Path $mesh)) {
+    # Exact rigid correction of the admitted historical cache avoids rerunning
+    # a faulty third-party BVH build. Never mutates or reclassifies that mesh.
+    $ErrorActionPreference = 'Continue'
+    & $Python $migrator --root $Root --world $WorldFile --output $mesh
+    if ($LASTEXITCODE -notin @(0,2)) { throw 'Collision coordinate migration failed' }
+    $ErrorActionPreference = 'Stop'
+}
+if (-not (Test-Path $mesh)) {
+    # SplatTransform 3.4 assigns PLY an implicit Rz(180) before CLI actions.
+    # Cancel it FIRST, then apply the same XYZ rotation as the USD visual.
+    # World Z-up -> glTF Y-up is Rx(-90): (x,y,z) -> (x,z,-y).
+    # Runtime's (x,y,z) -> (x,-z,y) is exactly its inverse. Do not use Rx(+90).
+    $x = [double]$settings.spawn_xy[0]; $y = [double]$settings.spawn_xy[1]; $z = [double]$settings.ground_z
+    $box = @(($x-15),($z-4),(-$y-15),($x+15),($z+8),(-$y+15)) -join ','
+    $args = @('--max-old-space-size=12288', (Join-Path $tool 'bin\cli.mjs'), $source,
+        '--filter-nan', '--rotate', '0,0,180',
+        '--rotate', ($settings.rotation_degrees -join ','),
+        '--scale', [string]$settings.meters_per_unit,
+        '--rotate', '-90,0,0', '--filter-box', $box,
+        '--voxel-size', '0.06', '--voxel-opacity', '0.25',
+        '--voxel-floor-fill', '0.3', '--collision-mesh', 'smooth',
+        (Join-Path $out 'terrain.voxel.json'))
+    $ErrorActionPreference = 'Continue'
+    & node @args
+    if ($LASTEXITCODE -ne 0) { throw 'SplatTransform collision generation failed' }
+    $ErrorActionPreference = 'Stop'
+}
+if (-not (Test-Path $mesh)) { throw 'Collision mesh was not produced' }
+$correctionPath = Join-Path $out 'coordinate-correction.json'
+$correction = if (Test-Path $correctionPath) { Get-Content -Raw $correctionPath | ConvertFrom-Json } else { $null }
+if ($correction) {
+    if ((Get-FileHash $mesh -Algorithm SHA256).Hash.ToLower() -ne $correction.collider_sha256) { throw 'Corrected collision changed' }
+    $settings = $correction.settings
+}
+$manifest = @{
+    schema_version = 'missioncore.ai-polygon-terrain/v1'
+    source_sha256 = $world.sha256
+    settings = $settings
+    generator = '@playcanvas/splat-transform@3.4.0'
+    generator_sha256 = (Get-FileHash $PSCommandPath -Algorithm SHA256).Hash.ToLower()
+    voxel_size_m = 0.06
+    collider = $mesh
+    collider_sha256 = (Get-FileHash $mesh -Algorithm SHA256).Hash.ToLower()
+    coordinate_system = 'gltf-y-up-from-metric-world'
+    coordinate_revision = 'ply-rz180-cancelled-usd-xyz-gltf-rx-minus90-v2'
+    coordinate_correction = $correction
+    qualification = 'reconstructed-proxy-requires-contact-validation'
+}
+[IO.File]::WriteAllText((Join-Path $out 'terrain.json'), ($manifest | ConvertTo-Json -Depth 6), [Text.UTF8Encoding]::new($false))
+$manifest | ConvertTo-Json -Depth 6 -Compress
diff --git a/simulation/ai-polygon/navigation/fastdds.xml b/simulation/ai-polygon/navigation/fastdds.xml
new file mode 100644
index 0000000..eac068e
--- /dev/null
+++ b/simulation/ai-polygon/navigation/fastdds.xml
@@ -0,0 +1,18 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<profiles xmlns="http://www.eprosima.com/XMLSchemas/fastRTPS_Profiles">
+  <transport_descriptors>
+    <transport_descriptor>
+      <transport_id>worker_loopback</transport_id>
+      <type>UDPv4</type>
+      <sendBufferSize>1048576</sendBufferSize>
+      <receiveBufferSize>1048576</receiveBufferSize>
+      <interfaceWhiteList><address>127.0.0.1</address></interfaceWhiteList>
+    </transport_descriptor>
+  </transport_descriptors>
+  <participant profile_name="worker_navigation" is_default_profile="true">
+    <rtps>
+      <userTransports><transport_id>worker_loopback</transport_id></userTransports>
+      <useBuiltinTransports>false</useBuiltinTransports>
+    </rtps>
+  </participant>
+</profiles>
diff --git a/simulation/ai-polygon/navigation/footprint.py b/simulation/ai-polygon/navigation/footprint.py
new file mode 100644
index 0000000..6a11f42
--- /dev/null
+++ b/simulation/ai-polygon/navigation/footprint.py
@@ -0,0 +1,107 @@
+"""Metre-square swept-body check on CMU's observed terrain and chosen path.
+
+CMU's circular path table proposes paths. This final adapter check preserves
+the actual square chassis, including the initial turn, without widening it to
+its circumscribed circle for straight travel. No scene geometry enters here.
+"""
+
+import math
+
+import numpy as np
+
+MAX_STEP_M = 0.10
+FRAME_DEADLINE_SECONDS = 0.8
+
+
+def _obstacles(terrain, pose):
+    terrain = np.asarray(terrain, dtype=float)
+    obstacles = terrain[terrain[:, 3] > MAX_STEP_M + 1e-4, :2] - np.asarray(pose[:2])
+    x, y, z, w = pose[3:]
+    yaw = math.atan2(2 * (w * z + x * y), 1 - 2 * (y * y + z * z))
+    return obstacles @ np.array([[math.cos(yaw), -math.sin(yaw)], [math.sin(yaw), math.cos(yaw)]])
+
+
+def _clearances(obstacles):
+    return np.maximum(np.abs(obstacles[:, 0]) - 0.5, np.abs(obstacles[:, 1]) - 0.5)
+
+
+def _clear(obstacles, position, angle):
+    delta = obstacles - position
+    c, s = math.cos(angle), math.sin(angle)
+    along = delta[:, 0] * c + delta[:, 1] * s
+    across = -delta[:, 0] * s + delta[:, 1] * c
+    initial = _clearances(obstacles)
+    if np.any(initial <= 0):
+        return False  # Never excuse an overlap of the actual metre-square body.
+    clearance = np.maximum(np.abs(along) - 0.5, np.abs(across) - 0.5)
+    # An observed point may already be inside the 5 cm reserve behind the body.
+    # Permit only motion that never decreases that initial clearance. This
+    # cannot authorize moving toward it, reversing into it or corner penetration.
+    return bool(np.all(clearance + 1e-6 >= np.minimum(initial, 0.05)))
+
+
+def command_footprint_clear(speed, yaw_rate, terrain, pose):
+    """Collision monitor over deadman latency plus a conservative braking arc.
+
+    CMU replans the route continuously. A later blocked corner must not prevent
+    safe progress on its prefix; this checks the command that can actually be
+    applied before the source-frame deadline, plus braking and 0.5 s reserve.
+    """
+    obstacles = _obstacles(terrain, pose)
+    horizon = FRAME_DEADLINE_SECONDS + 0.5 + abs(speed) / 0.4 + abs(yaw_rate) / 1.6
+    for t in np.arange(0, horizon + 0.025, 0.025):
+        angle = yaw_rate * t
+        position = (
+            np.array([speed * math.sin(angle) / yaw_rate, speed * (1 - math.cos(angle)) / yaw_rate])
+            if abs(yaw_rate) > 1e-6
+            else np.array([speed * t, 0])
+        )
+        if not _clear(obstacles, position, angle):
+            return False
+    return True
+
+
+def regulate_command(speed, yaw_rate, terrain, pose):
+    """Reduce speed along CMU's same arc when its full-speed stop is unsafe.
+
+    Scaling both components preserves curvature. The shortened stopping envelope
+    is a prefix of the original arc, so search for its largest admitted scale.
+    Never choose another turn/direction, ignore a hazard or creep arbitrarily.
+    """
+    if command_footprint_clear(speed, yaw_rate, terrain, pose):
+        return speed, yaw_rate, 1.0
+    low, high = 0.2, 1.0
+    if not command_footprint_clear(speed * low, yaw_rate * low, terrain, pose):
+        return 0.0, 0.0, 0.0
+    for _ in range(7):
+        middle = (low + high) / 2
+        if command_footprint_clear(speed * middle, yaw_rate * middle, terrain, pose):
+            low = middle
+        else:
+            high = middle
+    return speed * low, yaw_rate * low, low
+
+
+def swept_footprint_clear(path, terrain, pose):
+    path = np.asarray(path, dtype=float)[:, :2]
+    terrain = np.asarray(terrain, dtype=float)
+    if len(path) < 2 or terrain.ndim != 2 or terrain.shape[1] != 4:
+        return False
+    obstacles = _obstacles(terrain, pose)
+
+    previous_angle = 0.0
+    for start, end in zip(path[:-1], path[1:], strict=True):
+        delta = end - start
+        length = np.linalg.norm(delta)
+        if length < 1e-6:
+            continue
+        angle = math.atan2(delta[1], delta[0])
+        turn = math.atan2(math.sin(angle - previous_angle), math.cos(angle - previous_angle))
+        for fraction in np.linspace(0, 1, max(2, math.ceil(abs(turn) / 0.035) + 1)):
+            if not _clear(obstacles, start, previous_angle + fraction * turn):
+                return False
+        for fraction in np.linspace(0, 1, max(2, math.ceil(length / 0.025) + 1)):
+            if not _clear(obstacles, start + fraction * delta, angle):
+                return False
+        previous_angle = angle
+    return True
diff --git a/simulation/ai-polygon/navigation/migrate_terrain_coordinates.py b/simulation/ai-polygon/navigation/migrate_terrain_coordinates.py
new file mode 100644
index 0000000..59cf350
--- /dev/null
+++ b/simulation/ai-polygon/navigation/migrate_terrain_coordinates.py
@@ -0,0 +1,120 @@
+"""Correct only the known v1 cache basis; original collision assets are immutable.
+
+For the admitted [-90,0,180] scene rotation, the old and correct pipelines
+have identical vertical axes and differ by a horizontal 180-degree rotation.
+Rotating the finished mesh is exact: no filtering, smoothing or new geometry.
+"""
+
+import argparse
+import hashlib
+import json
+import struct
+from pathlib import Path
+
+LEGACY_GENERATOR = "d7e142184c1de43cde71ebc7cd311e667c18c7c7a17725c14e123127740697a2"
+
+
+def rotate_glb(data):
+    magic, version, size = struct.unpack_from("<III", data)
+    if (magic, version, size) != (0x46546C67, 2, len(data)):
+        raise ValueError("Invalid collision GLB")
+    chunks = {}
+    offset = 12
+    while offset < size:
+        length, kind = struct.unpack_from("<II", data, offset)
+        chunks[kind] = data[offset + 8 : offset + 8 + length]
+        offset += 8 + length
+    document = json.loads(chunks[0x4E4F534A])
+    binary = bytearray(chunks[0x004E4942])
+    if any(
+        any(k in node for k in ("matrix", "rotation", "translation", "scale"))
+        for node in document.get("nodes", [])
+    ):
+        raise ValueError("Collision transforms must be baked")
+    transformed = set()
+    for mesh in document["meshes"]:
+        for primitive in mesh["primitives"]:
+            if set(primitive["attributes"]) != {"POSITION"}:
+                raise ValueError("Only the known position-only collision export is admitted")
+            index = primitive["attributes"]["POSITION"]
+            if index in transformed:
+                continue
+            transformed.add(index)
+            accessor = document["accessors"][index]
+            view = document["bufferViews"][accessor["bufferView"]]
+            if (
+                accessor["componentType"] != 5126
+                or accessor["type"] != "VEC3"
+                or "sparse" in accessor
+                or view.get("buffer", 0) != 0
+            ):
+                raise ValueError("Unsupported collision position buffer")
+            start = view.get("byteOffset", 0) + accessor.get("byteOffset", 0)
+            stride = view.get("byteStride", 12)
+            for i in range(accessor["count"]):
+                at = start + i * stride
+                x, y, z = struct.unpack_from("<fff", binary, at)
+                struct.pack_into("<fff", binary, at, -x, y, -z)
+            low, high = accessor["min"], accessor["max"]
+            accessor["min"] = [-high[0], low[1], -high[2]]
+            accessor["max"] = [-low[0], high[1], -low[2]]
+    encoded = json.dumps(document, separators=(",", ":")).encode()
+    encoded += b" " * ((-len(encoded)) % 4)
+    result = struct.pack("<III", 0x46546C67, 2, 12 + 8 + len(encoded) + 8 + len(binary))
+    return (
+        result
+        + struct.pack("<II", len(encoded), 0x4E4F534A)
+        + encoded
+        + struct.pack("<II", len(binary), 0x004E4942)
+        + binary
+    )
+
+
+def migrate(root, world, output):
+    settings = world["settings"]
+    if settings["rotation_degrees"] != [-90, 0, 180]:
+        return False
+    for path in sorted((root / "assets/terrain-v1").glob(world["sha256"] + "-*/terrain.json")):
+        manifest = json.loads(path.read_text(encoding="utf-8-sig"))
+        old = manifest["settings"]
+        if (
+            manifest.get("generator_sha256") != LEGACY_GENERATOR
+            or old["rotation_degrees"] != [-90, 0, 180]
+            or old["meters_per_unit"] != settings["meters_per_unit"]
+        ):
+            continue
+        centre = [-v for v in old["spawn_xy"]]
+        if (
+            any(abs(centre[i] - settings["spawn_xy"][i]) > 12 for i in (0, 1))
+            or abs(old["ground_z"] - settings["ground_z"]) > 2
+        ):
+            continue
+        original = Path(manifest["collider"]).read_bytes()
+        if hashlib.sha256(original).hexdigest() != manifest["collider_sha256"]:
+            raise ValueError("Legacy collision identity changed")
+        corrected = rotate_glb(original)
+        temporary = output.with_suffix(".part.glb")
+        temporary.write_bytes(corrected)
+        temporary.replace(output)
+        report = {
+            "source_manifest": str(path),
+            "source_collider_sha256": manifest["collider_sha256"],
+            "collider_sha256": hashlib.sha256(corrected).hexdigest(),
+            "transform": "gltf-rotate-y-180",
+            "settings": {**old, "spawn_xy": centre},
+        }
+        output.with_name("coordinate-correction.json").write_text(
+            json.dumps(report, indent=2), encoding="utf-8"
+        )
+        return True
+    return False
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--root", type=Path, required=True)
+    parser.add_argument("--world", type=Path, required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    args = parser.parse_args()
+    if not migrate(args.root, json.loads(args.world.read_text(encoding="utf-8-sig")), args.output):
+        raise SystemExit(2)
diff --git a/simulation/ai-polygon/navigation/missioncore_footprint.hpp b/simulation/ai-polygon/navigation/missioncore_footprint.hpp
new file mode 100644
index 0000000..66d687a
--- /dev/null
+++ b/simulation/ai-polygon/navigation/missioncore_footprint.hpp
@@ -0,0 +1,74 @@
+// Mission Core adapter for CMU's existing motion-primitive selector.
+// Reject a candidate before selection when its initial turn or first metre
+// sweeps the real rectangular body through observed terrain. Upstream's
+// angular wedge alone misses obstacles beside a rear corner.
+#pragma once
+#include <algorithm>
+#include <cmath>
+#include <vector>
+#include <utility>
+
+template <class TerrainCloud, class PathCloud>
+bool missioncoreFootprintClear(const TerrainCloud& terrain, const PathCloud& path,
+                              double rotation, double scale, double range,
+                              double length, double width, double threshold,
+                              bool twoWayDrive) {
+  const double margin = 0.05;
+  const double halfLength = length / 2.0 - margin, halfWidth = width / 2.0 - margin;
+  const double cap = std::min(range, 1.2);
+  std::vector<std::pair<double, double>> obstacles;
+  for (const auto& p : terrain.points) {
+    if (p.intensity > threshold && std::hypot(p.x, p.y) < cap + std::hypot(halfLength + margin, halfWidth + margin))
+      obstacles.emplace_back(p.x, p.y);
+  }
+  auto clear = [&](double x, double y, double yaw) {
+    const double c = std::cos(yaw), s = std::sin(yaw);
+    for (const auto& p : obstacles) {
+      const double dx = p.first - x, dy = p.second - y;
+      const double initial = std::max(std::abs(p.first) - halfLength,
+                                      std::abs(p.second) - halfWidth);
+      if (initial <= 0) return false;
+      const double clearance = std::max(std::abs(dx * c + dy * s) - halfLength,
+                                        std::abs(-dx * s + dy * c) - halfWidth);
+      if (clearance + 1e-6 < std::min(margin, initial)) return false;
+    }
+    return true;
+  };
+  auto turnClear = [&](double x, double y, double from, double to) {
+    const double delta = std::atan2(std::sin(to - from), std::cos(to - from));
+    const int steps = std::max(1, int(std::ceil(std::abs(delta) / 0.025)));
+    for (int i = 0; i <= steps; ++i)
+      if (!clear(x, y, from + delta * i / steps)) return false;
+    return true;
+  };
+  if (path.points.size() < 2 || !clear(0, 0, 0)) return false;
+  // Match pathFollower's 0.7 m look-ahead. Its initial command points here,
+  // which need not equal the selected primitive's initial tangent.
+  double lookYaw = rotation;
+  for (const auto& p : path.points) {
+    lookYaw = rotation + std::atan2(p.y, p.x);
+    if (std::hypot(p.x, p.y) * scale >= std::min(0.7, cap)) break;
+  }
+  // A reversing body follows the rear-facing primitive without first turning
+  // 180 degrees. Match pathFollower's twoWayDrive direction selection.
+  const bool reverse = twoWayDrive && std::cos(lookYaw) < 0;
+  if (reverse) lookYaw += M_PI;
+  if (!turnClear(0, 0, 0, lookYaw)) return false;
+  const double c = std::cos(rotation), s = std::sin(rotation);
+  double lastX = 0, lastY = 0, lastYaw = lookYaw;
+  for (const auto& p : path.points) {
+    if (std::hypot(p.x, p.y) * scale > cap) break;
+    const double x = scale * (c * p.x - s * p.y);
+    const double y = scale * (s * p.x + c * p.y);
+    const double distance = std::hypot(x - lastX, y - lastY);
+    if (distance < 1e-6) continue;
+    const double yaw = std::atan2(y - lastY, x - lastX) + (reverse ? M_PI : 0);
+    if (!turnClear(lastX, lastY, lastYaw, yaw)) return false;
+    const int steps = std::max(1, int(std::ceil(distance / 0.025)));
+    for (int i = 1; i <= steps; ++i)
+      if (!clear(lastX + (x - lastX) * i / steps,
+                 lastY + (y - lastY) * i / steps, yaw)) return false;
+    lastX = x; lastY = y; lastYaw = yaw;
+  }
+  return true;
+}
diff --git a/simulation/ai-polygon/navigation/patch_footprint.py b/simulation/ai-polygon/navigation/patch_footprint.py
new file mode 100644
index 0000000..f142258
--- /dev/null
+++ b/simulation/ai-polygon/navigation/patch_footprint.py
@@ -0,0 +1,14 @@
+"""Apply one explicit adapter patch to the pinned CMU selector; fail on drift."""
+from pathlib import Path
+
+path = Path("/opt/cmu/src/local_planner/src/localPlanner.cpp")
+source = path.read_text()
+needle = "            maxScore = clearPathPerGroupScore[i];\n            selectedGroupID = i;"
+replacement = """\
+            if (!missioncoreFootprintClear(*plannerCloudCrop, *startPaths[i % groupNum],
+                rotAng, pathScale, std::min(pathRange, relativeGoalDis),
+                vehicleLength, vehicleWidth, obstacleHeightThre, twoWayDrive)) continue;
+            maxScore = clearPathPerGroupScore[i];
+            selectedGroupID = i;"""
+assert source.count(needle) == 1, "CMU selector source changed"
+path.write_text('#include "missioncore_footprint.hpp"\n' + source.replace(needle, replacement))
diff --git a/simulation/ai-polygon/navigation/qualify.py b/simulation/ai-polygon/navigation/qualify.py
new file mode 100644
index 0000000..4b6acb4
--- /dev/null
+++ b/simulation/ai-polygon/navigation/qualify.py
@@ -0,0 +1,294 @@
+"""Bounded synthetic acceptance of the installed CMU image on Worker only.
+
+This tests navigation independently of perception and physics. It does not
+claim Gaussian-world or physical-rover acceptance. Owns one temporary container.
+"""
+
+import argparse
+import http.client
+import json
+import math
+import subprocess
+import time
+from datetime import UTC, datetime
+from pathlib import Path
+from uuid import uuid4
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--adapter", type=Path, required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    args = parser.parse_args()
+    profile = json.loads((args.adapter / "models.worker-006.json").read_text())
+    name = "ndc-mission-core-ai-module-navigation-check-" + uuid4().hex[:8]
+    identity = subprocess.check_output(
+        [
+            "docker",
+            "run",
+            "-d",
+            "--name",
+            name,
+            "--cpus",
+            "3",
+            "--memory",
+            "2g",
+            "--label",
+            "com.nodedc.stack=ai-polygon-qualification",
+            "--label",
+            "com.nodedc.product=mission-core",
+            "--label",
+            "com.nodedc.role=qualification",
+            "--label",
+            "com.nodedc.managed-by=ai-polygon-qualification",
+            "-p",
+            "127.0.0.1:18193:8010",
+            "--mount",
+            f"type=bind,source={args.adapter},target=/adapter,readonly",
+            profile["navigation"]["image"],
+        ],
+        text=True,
+    ).strip()
+    connection = http.client.HTTPConnection("127.0.0.1", 18193, timeout=3)
+
+    def request(path, body=None):
+        connection.request(
+            "GET" if body is None else "POST", path, body=None if body is None else json.dumps(body)
+        )
+        response = connection.getresponse()
+        value = json.loads(response.read())
+        if response.status != 200:
+            raise RuntimeError(str(value))
+        return value
+
+    def ready():
+        deadline = time.monotonic() + 15
+        while time.monotonic() < deadline:
+            try:
+                request("/ready")
+                return
+            except (OSError, RuntimeError, http.client.HTTPException):
+                time.sleep(0.2)
+        raise TimeoutError("CMU qualification container did not become ready")
+
+    report = {
+        "image": profile["navigation"]["image"],
+        "cases": [],
+        "utc": datetime.now(UTC).isoformat(),
+        "monotonic": time.monotonic(),
+    }
+    try:
+        ready()
+        floor = [[x / 10, y / 10, 0.0] for x in range(-20, 41) for y in range(-20, 21)]
+        wall = [[1.5, y / 10, z / 10] for y in range(-20, 21) for z in range(1, 16)]
+        enclosure = [
+            [side * 0.9, y / 10, z / 10]
+            for side in (-1, 1)
+            for y in range(-9, 10)
+            for z in range(1, 16)
+        ]
+        enclosure += [
+            [x / 10, side * 0.9, z / 10]
+            for side in (-1, 1)
+            for x in range(-9, 10)
+            for z in range(1, 16)
+        ]
+
+        def corridor(half_width):
+            return floor + [
+                [x / 10, side * half_width, z / 10]
+                for x in range(-15, 51)
+                for side in (-1, 1)
+                for z in range(1, 16)
+            ]
+
+        for name, points in (
+            ("clear-flat-ground", floor),
+            ("rear-reserve-forward-escape", floor + [[-0.53, 0, z / 10] for z in range(1, 10)]),
+            ("front-reserve-stop", floor + [[0.53, 0, z / 10] for z in range(1, 10)]),
+            ("body-overlap-stop", floor + [[-0.49, 0, z / 10] for z in range(1, 10)]),
+            ("reverse-clear-ground", floor),
+            (
+                "reverse-blocked-by-rear-wall",
+                floor + [[-0.7, y / 10, z / 10] for y in range(-10, 11) for z in range(1, 12)],
+            ),
+            ("turn-away-from-wall", floor + wall),
+            (
+                "close-wall-stop",
+                floor + [[0.65, y / 10, z / 10] for y in range(-20, 21) for z in range(1, 16)],
+            ),
+            ("closed-enclosure-stop", floor + enclosure),
+            ("metre-rover-in-1.3m-corridor", corridor(0.65)),
+            ("metre-rover-rejects-0.9m-corridor", corridor(0.45)),
+            (
+                "rear-corner-selects-clear-primitive",
+                floor + [[-0.62, -0.39, z / 10] for z in range(1, 15)],
+            ),
+        ):
+            request("/reset", {})
+            ready()
+            samples = []
+            for _ in range(20):
+                samples.append(
+                    request(
+                        "/plan",
+                        {
+                            "points": points,
+                            "pose": [0, 0, 0.27, 0, 0, 0, 1],
+                            "goal": [-0.65, 0, 0]
+                            if name.startswith("reverse-")
+                            else ([2, -2, 0] if name.startswith("rear-corner") else [3, 0, 0]),
+                            "allow_reverse": name.startswith("reverse-"),
+                            "max_speed_mps": 0.15,
+                        },
+                    )
+                )
+                time.sleep(0.2)
+            if name in (
+                "clear-flat-ground",
+                "metre-rover-in-1.3m-corridor",
+                "rear-reserve-forward-escape",
+            ):
+                passed = any(v["speed_mps"] > 0.05 and v["status"] == "path" for v in samples)
+            elif name == "reverse-clear-ground":
+                passed = any(v["speed_mps"] < -0.05 and v["status"] == "path" for v in samples[-5:])
+            elif name == "turn-away-from-wall":
+                valid = [v for v in samples[-5:] if v["status"] == "path"]
+                passed = len(valid) >= 3 and all(
+                    (v["speed_mps"] > 0.02 or abs(v["yaw_rate_rps"]) > 0.1)
+                    and max(p[0] for p in v["path"]) < 0.75
+                    for v in valid
+                )
+            elif name.startswith("rear-corner"):
+                passed = any(v["speed_mps"] > 0.05 for v in samples[-5:]) and all(
+                    v.get("diagnostic", {}).get("failure") != "footprint" for v in samples[-5:]
+                )
+            else:
+                passed = all(v["speed_mps"] == 0 and v["status"] == "blocked" for v in samples[-5:])
+            report["cases"].append({"name": name, "passed": passed, "samples": samples})
+        # A vanished return is occlusion, not proof of free space. Observe a
+        # near wall, then only distant ground for longer than the decay timer.
+        request("/reset", {})
+        ready()
+        close_wall = [[0.7, y / 10, z / 10] for y in range(-10, 11) for z in range(1, 10)]
+        samples = []
+        for i in range(30):
+            samples.append(
+                request(
+                    "/plan",
+                    {
+                        "points": floor + close_wall if i < 5 else [p for p in floor if p[0] > 1.2],
+                        "pose": [0, 0, 0.37, 0, 0, 0, 1],
+                        "goal": [3, 0, 0],
+                        "max_speed_mps": 0.15,
+                    },
+                )
+            )
+            time.sleep(0.2)
+        report["cases"].append(
+            {
+                "name": "occluded-near-obstacle-retained",
+                "samples": samples,
+                "passed": all(
+                    x["speed_mps"] == 0 and x["status"] == "blocked" for x in samples[-5:]
+                ),
+            }
+        )
+        # Navigation must admit the same continuous grades as the measured
+        # physical profile, while retaining a discontinuous 15 cm ledge.
+        for angle, quantized in [(a, False) for a in (5, 10, 15, 20, 25)] + [
+            (10, True),
+            (20, True),
+        ]:
+            request("/reset", {})
+            ready()
+            radians = math.radians(angle)
+            slope = math.tan(radians)
+            samples = []
+            for _ in range(12):
+                samples.append(
+                    request(
+                        "/plan",
+                        {
+                            "points": [
+                                [x, y, round(x * slope / 0.06) * 0.06 if quantized else x * slope]
+                                for x, y, _ in floor
+                            ],
+                            "pose": [
+                                0,
+                                0,
+                                0.37,
+                                0,
+                                -math.sin(radians / 2),
+                                0,
+                                math.cos(radians / 2),
+                            ],
+                            "goal": [3, 0, 3 * slope],
+                            "max_speed_mps": 0.15,
+                        },
+                    )
+                )
+                time.sleep(0.2)
+            report["cases"].append(
+                {
+                    "name": f"supported-{'voxel-' if quantized else ''}ramp-{angle}-degrees",
+                    "samples": samples,
+                    "passed": any(s["speed_mps"] > 0.05 for s in samples[-5:]),
+                }
+            )
+        for height in (0.12, 0.15, -0.4):
+            request("/reset", {})
+            ready()
+            points = [[x, y, height if x >= 0.7 else 0] for x, y, _ in floor]
+            points += [
+                [x / 10, side * 0.65, z / 10]
+                for x in range(-15, 41)
+                for side in (-1, 1)
+                for z in range(1, 16)
+            ]
+            samples = []
+            for _ in range(12):
+                samples.append(
+                    request(
+                        "/plan",
+                        {
+                            "points": points,
+                            "pose": [0, 0, 0.37, 0, 0, 0, 1],
+                            "goal": [3, 0, 0],
+                            "max_speed_mps": 0.15,
+                        },
+                    )
+                )
+                time.sleep(0.2)
+            report["cases"].append(
+                {
+                    "name": f"discontinuous-height-{height:+.2f}m-stop",
+                    "samples": samples,
+                    "passed": all(
+                        s["speed_mps"] == 0 and s["status"] == "blocked" for s in samples[-5:]
+                    ),
+                }
+            )
+        report["passed"] = all(case["passed"] for case in report["cases"])
+    finally:
+        report["container_log_tail"] = subprocess.run(
+            ["docker", "logs", "--tail", "30", identity], capture_output=True, text=True
+        ).stderr
+        args.output.parent.mkdir(parents=True, exist_ok=True)
+        args.output.write_text(json.dumps(report, indent=2), encoding="utf-8")
+        connection.close()
+        subprocess.run(["docker", "rm", "-f", identity], check=True, capture_output=True)
+    print(
+        json.dumps(
+            {
+                "passed": report["passed"],
+                "cases": [
+                    {"name": row["name"], "passed": row["passed"]} for row in report["cases"]
+                ],
+            }
+        )
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/simulation/ai-polygon/navigation/qualify_connectivity.py b/simulation/ai-polygon/navigation/qualify_connectivity.py
new file mode 100644
index 0000000..e6179e3
--- /dev/null
+++ b/simulation/ai-polygon/navigation/qualify_connectivity.py
@@ -0,0 +1,31 @@
+"""Worker-only numeric equivalence at hazard boundaries and disconnected patches."""
+
+import argparse
+import json
+from pathlib import Path
+
+import numpy as np
+import terrain_costs as costs
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--output", type=Path, required=True)
+args = parser.parse_args()
+assert costs._NATIVE is not None, "Compiled implementation must be installed"
+engine = costs._NATIVE
+rng = np.random.default_rng(230923)
+cases = []
+for dtype in (np.float32, np.float64):
+    for _ in range(200):
+        points = rng.uniform(-0.4, 0.4, (rng.integers(8, 180), 3)).astype(dtype)
+        points[:, 2] *= 0.3
+        cases.append(points)
+    for distance in (0.119999, 0.12, 0.120001):
+        for jump in (0.100099, 0.1001, 0.100101):
+            cases.append(np.array([[0, 0, 0], [distance, 0, jump]], dtype=dtype))
+for points in cases:
+    costs._NATIVE = None
+    expected = costs.connected_grade(points)
+    costs._NATIVE = engine
+    assert costs.connected_grade(points) == expected
+args.output.write_text(json.dumps({"passed": True, "cases": len(cases)}))
+print(json.dumps({"passed": True, "cases": len(cases)}))
diff --git a/simulation/ai-polygon/navigation/qualify_ipc.py b/simulation/ai-polygon/navigation/qualify_ipc.py
new file mode 100644
index 0000000..c968f41
--- /dev/null
+++ b/simulation/ai-polygon/navigation/qualify_ipc.py
@@ -0,0 +1,50 @@
+"""Bounded Windows snapshot read/replace acceptance; no simulator required."""
+
+import json
+import sys
+import tempfile
+import threading
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+from local_state import StateChannel
+
+
+def main():
+    errors = []
+    done = threading.Event()
+    with tempfile.TemporaryDirectory(prefix="ndc-polygon-ipc-") as folder:
+        directory = Path(folder)
+        writer = StateChannel(directory)
+        writer.write("snapshot", {"sequence": -1})
+
+        def reader():
+            channel = StateChannel(directory)
+            try:
+                while not done.is_set():
+                    row = channel.read("snapshot")
+                    if row is None or not -1 <= row["sequence"] < 1000:
+                        raise RuntimeError("Incomplete snapshot read")
+            except Exception as exc:
+                errors.append(type(exc).__name__ + ": " + str(exc))
+            finally:
+                channel.close()
+
+        thread = threading.Thread(target=reader)
+        thread.start()
+        try:
+            for sequence in range(1000):
+                writer.write("snapshot", {"sequence": sequence, "telemetry": [sequence] * 100})
+        finally:
+            done.set()
+            thread.join()
+        if errors:
+            raise RuntimeError(str(errors))
+        if writer.read("snapshot")["sequence"] != 999:
+            raise RuntimeError("Final snapshot was not retained")
+        writer.close()
+    print(json.dumps({"passed": True, "concurrent_replacements": 1000}))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/simulation/ai-polygon/navigation/qualify_perception.py b/simulation/ai-polygon/navigation/qualify_perception.py
new file mode 100644
index 0000000..d5c996e
--- /dev/null
+++ b/simulation/ai-polygon/navigation/qualify_perception.py
@@ -0,0 +1,57 @@
+"""Offline numeric check of the pinned GOOSE decoder on retained Worker RGB.
+
+Runs in the existing pinned image, without changing its runner or checkpoint.
+This diagnoses the adapter; it is not semantic accuracy or navigation acceptance.
+"""
+
+import argparse
+import csv
+import hashlib
+import importlib.util
+import json
+import time
+from pathlib import Path
+
+import numpy as np
+import torch
+from PIL import Image
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--image", type=Path, required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    args = parser.parse_args()
+    spec = importlib.util.spec_from_file_location("reference", "/assets/ddrnet-goose-runner.py")
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    model, _, _ = module.load_model("ddrnet", Path("/assets/ddrnet-checkpoint.pth"))
+    tensor, _ = module.preprocess(Image.open(args.image))
+    tensor = tensor.cuda()
+    with torch.inference_mode():
+        logits = module.logits_from_output(model(tensor)).float()
+        legacy = torch.sigmoid(logits).argmax(1)
+        direct = logits.argmax(1)
+        saturated = (torch.sigmoid(logits) == 1).sum(1)
+    names = {}
+    with open("/assets/ddrnet-goose-mapping.csv") as stream:
+        names = {int(r["label_key"]): r["class_name"] for r in csv.DictReader(stream)}
+    args.output.mkdir(exist_ok=True, parents=True)
+    report = {
+        "source_sha256": hashlib.sha256(args.image.read_bytes()).hexdigest(),
+        "monotonic_ns": time.monotonic_ns(),
+        "logit_range": [float(logits.min()), float(logits.max())],
+        "changed_pixels": int((legacy != direct).sum()),
+        "saturated_tie_pixels": int((saturated > 1).sum()),
+    }
+    for name, mask in (("reference", legacy), ("direct", direct)):
+        mask = mask[0].cpu().numpy().astype(np.uint8)
+        Image.fromarray(mask).save(args.output / (name + ".png"))
+        ids, counts = np.unique(mask, return_counts=True)
+        report[name] = {names[int(i)]: int(counts[index]) for index, i in enumerate(ids)}
+    (args.output / "report.json").write_text(json.dumps(report, indent=2), encoding="utf-8")
+    print(json.dumps(report))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/simulation/ai-polygon/navigation/qualify_placement.py b/simulation/ai-polygon/navigation/qualify_placement.py
new file mode 100644
index 0000000..a0a6fea
--- /dev/null
+++ b/simulation/ai-polygon/navigation/qualify_placement.py
@@ -0,0 +1,105 @@
+"""Offline authoring check for a stable, metre-wide start on a scan proxy.
+
+Uses world geometry only to prepare a scene. No candidate map enters navigation.
+An operator/engineer still verifies the chosen start against the visual trail.
+"""
+
+import argparse
+import json
+import math
+import sys
+from pathlib import Path
+
+import numpy as np
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
+from spawn_clearance import obstructing_triangles
+from terrain import load_glb
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--terrain", type=Path, required=True)
+    args = parser.parse_args()
+    manifest = json.loads((args.terrain / "terrain.json").read_text(encoding="utf-8-sig"))
+    settings = manifest["settings"]
+    points, indices = load_glb(args.terrain / "terrain.collision.glb")
+    triangles = points[indices]
+    low, high = triangles.min(axis=1), triangles.max(axis=1)
+    center = np.array(settings["spawn_xy"])
+    selected = (
+        (low[:, :2] <= center + 2.5).all(axis=1)
+        & (high[:, :2] >= center - 2.5).all(axis=1)
+        & (low[:, 2] < settings["ground_z"] + 1.5)
+        & (high[:, 2] > settings["ground_z"] - 0.8)
+    )
+    triangles, low, high = triangles[selected], low[selected], high[selected]
+
+    def heights(x, y):
+        hits = triangles[
+            (low[:, 0] <= x) & (high[:, 0] >= x) & (low[:, 1] <= y) & (high[:, 1] >= y)
+        ]
+        if not len(hits):
+            return np.empty(0)
+        a, b, c = hits[:, 0], hits[:, 1], hits[:, 2]
+        den = (b[:, 1] - c[:, 1]) * (a[:, 0] - c[:, 0]) + (c[:, 0] - b[:, 0]) * (a[:, 1] - c[:, 1])
+        valid = np.abs(den) > 1e-8
+        a, b, c, den = a[valid], b[valid], c[valid], den[valid]
+        u = ((b[:, 1] - c[:, 1]) * (x - c[:, 0]) + (c[:, 0] - b[:, 0]) * (y - c[:, 1])) / den
+        v = ((c[:, 1] - a[:, 1]) * (x - c[:, 0]) + (a[:, 0] - c[:, 0]) * (y - c[:, 1])) / den
+        inside = (u >= -1e-6) & (v >= -1e-6) & (u + v <= 1 + 1e-6)
+        return (u * a[:, 2] + v * b[:, 2] + (1 - u - v) * c[:, 2])[inside]
+
+    angle = math.radians(settings["heading_degrees"])
+    rotation = np.array([[math.cos(angle), -math.sin(angle)], [math.sin(angle), math.cos(angle)]])
+    footprint = np.array([[x, y] for x in (-0.5, 0, 0.5) for y in (-0.5, 0, 0.5)]) @ rotation.T
+    candidates = []
+    for dx in np.arange(-2, 2.01, 0.2):
+        for dy in np.arange(-2, 2.01, 0.2):
+            position = center + [dx, dy]
+            support = []
+            for x, y in footprint + position:
+                z = heights(x, y)
+                near = z[np.abs(z - settings["ground_z"]) < 0.8]
+                if not len(near):
+                    break
+                ground = near.max()
+                if np.any((z > ground + 0.12) & (z < ground + 1)):
+                    break
+                support.append(float(ground))
+            if len(support) != 9:
+                continue
+            design = np.column_stack((footprint, np.ones(9)))
+            plane = np.linalg.lstsq(design, np.asarray(support), rcond=None)[0]
+            residual = float(np.max(np.abs(design @ plane - support)))
+            slope = math.degrees(math.atan(np.linalg.norm(plane[:2])))
+            if residual > 0.08 or slope > 20:
+                continue
+            if obstructing_triangles(
+                points, indices, position, settings["heading_degrees"], plane
+            ):
+                continue
+            candidates.append(
+                {
+                    "xy": position.tolist(),
+                    "ground_z": float(np.median(support)),
+                    "height_span": max(support) - min(support),
+                    "offset_m": math.hypot(dx, dy),
+                    "residual_m": residual,
+                    "slope_degrees": slope,
+                }
+            )
+    candidates.sort(key=lambda row: row["offset_m"] + 2 * row["height_span"])
+    print(
+        json.dumps(
+            {
+                "world_sha256": manifest["source_sha256"],
+                "candidate_count": len(candidates),
+                "candidates": candidates[:12],
+            }
+        )
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/simulation/ai-polygon/navigation/replay.py b/simulation/ai-polygon/navigation/replay.py
new file mode 100644
index 0000000..0f123cf
--- /dev/null
+++ b/simulation/ai-polygon/navigation/replay.py
@@ -0,0 +1,121 @@
+"""Replay retained sensor evidence against the shipped CMU adapter on Worker.
+
+Owns exactly one bounded CPU container. Never starts rendering or model GPU jobs.
+"""
+
+import argparse
+import hashlib
+import http.client
+import json
+import subprocess
+import time
+from collections import Counter
+from datetime import UTC, datetime
+from pathlib import Path
+from uuid import uuid4
+
+import numpy as np
+
+
+def main():
+    p = argparse.ArgumentParser()
+    p.add_argument("--adapter", type=Path, required=True)
+    p.add_argument("--run", type=Path, required=True)
+    p.add_argument("--output", type=Path, required=True)
+    p.add_argument("--limit", type=int, default=180)
+    p.add_argument("--terrain", action="store_true")
+    args = p.parse_args()
+    profile = json.loads((args.adapter / "models.worker-006.json").read_text())
+    identity = subprocess.check_output(
+        [
+            "docker",
+            "run",
+            "-d",
+            "--name",
+            "ndc-ai-polygon-navigation-replay-" + uuid4().hex[:8],
+            "--cpus",
+            "3",
+            "--memory",
+            "2g",
+            "--label",
+            "com.nodedc.product=mission-core",
+            "--label",
+            "com.nodedc.stack=ai-polygon-qualification",
+            "--label",
+            "com.nodedc.role=qualification",
+            "--label",
+            "com.nodedc.managed-by=ai-polygon-qualification",
+            "-p",
+            "127.0.0.1:18193:8010",
+            "--mount",
+            f"type=bind,source={args.adapter},target=/adapter,readonly",
+            profile["navigation"]["image"],
+        ],
+        text=True,
+    ).strip()
+    connection = http.client.HTTPConnection("127.0.0.1", 18193, timeout=3)
+    source = args.run / "camera/decisions.jsonl"
+    rows = [json.loads(line) for line in source.read_text().splitlines()]
+    report = dict(
+        utc=datetime.now(UTC).isoformat(),
+        monotonic=time.monotonic(),
+        input_sha256=hashlib.sha256(source.read_bytes()).hexdigest(),
+        samples=[],
+    )
+    try:
+        deadline = time.monotonic() + 20
+        while True:
+            try:
+                connection.request("GET", "/ready")
+                response = connection.getresponse()
+                response.read()
+                if response.status == 200:
+                    break
+            except (OSError, http.client.HTTPException):
+                connection.close()
+            if time.monotonic() > deadline:
+                raise TimeoutError("Replay navigation unavailable")
+            time.sleep(0.2)
+        for row in rows[: args.limit]:
+            if row["goal"] is None:
+                continue
+            observation = np.load(args.run / "camera" / f"{row['frame_id']:08d}.range.npz")
+            payload = dict(
+                points=observation["points"].tolist(),
+                pose=observation["pose"].tolist(),
+                goal=row["goal"],
+                max_speed_mps=0.15,
+                allow_reverse=row.get("mission", {}).get("state") == "reversing",
+                include_terrain=args.terrain,
+            )
+            connection.request("POST", "/plan", body=json.dumps(payload))
+            response = connection.getresponse()
+            result = json.loads(response.read())
+            if response.status != 200:
+                raise RuntimeError(result)
+            report["samples"].append(
+                dict(
+                    frame=row["frame_id"],
+                    pose=payload["pose"],
+                    original=row["decision"],
+                    result=result,
+                )
+            )
+            time.sleep(0.18)
+    finally:
+        connection.close()
+        subprocess.run(["docker", "rm", "-f", identity], check=True, capture_output=True)
+        args.output.parent.mkdir(parents=True, exist_ok=True)
+        args.output.write_text(json.dumps(report, indent=2), encoding="utf-8")
+    failures = Counter(
+        v["result"].get("diagnostic", {}).get("failure", v["result"]["status"])
+        for v in report["samples"]
+    )
+    blocked = [v for v in report["samples"] if v["result"]["status"] == "blocked"]
+    print(
+        json.dumps(dict(failures=failures, first_blocked=blocked[:1], last=report["samples"][-1:]))
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/simulation/ai-polygon/navigation/server.py b/simulation/ai-polygon/navigation/server.py
new file mode 100644
index 0000000..059934a
--- /dev/null
+++ b/simulation/ai-polygon/navigation/server.py
@@ -0,0 +1,449 @@
+"""Bounded local HTTP adapter for the unchanged CMU ROS 2 navigation nodes.
+
+Only simulated sensor observations enter ROS; no scene mesh or oracle route.
+The container is owned by one episode. A reset restarts all causal ROS state.
+"""
+
+import json
+import math
+import os
+import signal
+import subprocess
+import sys
+import threading
+import time
+from collections import OrderedDict
+from http.server import BaseHTTPRequestHandler, HTTPServer
+from pathlib import Path
+
+import numpy as np
+import rclpy
+from footprint import MAX_STEP_M, regulate_command
+from geometry_msgs.msg import PointStamped, TwistStamped
+from nav_msgs.msg import Odometry
+from nav_msgs.msg import Path as RosPath
+from rclpy.node import Node
+from sensor_msgs.msg import PointCloud2, PointField
+from sensor_msgs_py import point_cloud2
+from std_msgs.msg import Float32, Header
+from terrain_costs import TerrainCostNormalizer, underbody_support_costs
+
+
+def stamp_ns(stamp):
+    return stamp.sec * 1_000_000_000 + stamp.nanosec
+
+
+class Navigation(Node):
+    def __init__(self):
+        super().__init__("missioncore_navigation_adapter")
+        self.condition = threading.Condition()
+        self.processes = []
+        self.path = self.command = self.terrain = None
+        self.odom = self.create_publisher(Odometry, "/state_estimation", 5)
+        self.scan = self.create_publisher(PointCloud2, "/registered_scan", 5)
+        self.goal = self.create_publisher(PointStamped, "/way_point", 5)
+        self.speed = self.create_publisher(Float32, "/speed", 5)
+        self.obstacles = self.create_publisher(PointCloud2, "/added_obstacles", 5)
+        self.surface = self.create_publisher(PointCloud2, "/terrain_map", 5)
+        self.create_subscription(RosPath, "/path", self.on_path, 5)
+        self.create_subscription(TwistStamped, "/cmd_vel", self.on_command, 5)
+        self.create_subscription(PointCloud2, "/terrain_map_raw", self.on_terrain, 5)
+        self.slope_corrected = 0
+        self.terrain_processing_ms = 0.0
+        self.normalize_costs = TerrainCostNormalizer()
+        self.support_poses = OrderedDict()
+        self.underbody_corrected = 0
+        self.start_nodes()
+
+    def start_nodes(self):
+        common = dict(
+            autonomyMode=True,
+            autonomySpeed=0.3,
+            maxSpeed=1.0,
+            twoWayDrive=True,
+            joyToSpeedDelay=0.0,
+        )
+        configs = [
+            (
+                "terrain_analysis",
+                "terrainAnalysis",
+                dict(
+                    scanVoxelSize=0.06,
+                    # Keep the upstream near-field memory: an obstacle hidden
+                    # by our own chassis must not disappear after one second.
+                    decayTime=2.0,
+                    noDecayDis=4.0,
+                    useSorting=True,
+                    # Keep CMU's upstream ground quantile. Lower values make
+                    # shallow scan depressions the reference for the entire
+                    # 0.6 m neighbourhood; the median admits too much wall.
+                    quantileZ=0.25,
+                    considerDrop=True,
+                    clearDyObs=False,
+                    noDataObstacle=False,
+                    vehicleHeight=0.9,
+                    minRelZ=-2.0,
+                    maxRelZ=1.0,
+                    voxelPointUpdateThre=1,
+                    voxelTimeUpdateThre=0.0,
+                ),
+            ),
+            (
+                "local_planner",
+                "localPlanner",
+                dict(
+                    **common,
+                    pathFolder="/opt/cmu/install/local_planner/share/local_planner/paths",
+                    # Match the final monitor's 5 cm margin on every side;
+                    # otherwise CMU repeatedly proposes a forbidden corner turn.
+                    vehicleLength=1.1,
+                    vehicleWidth=1.1,
+                    useTerrainAnalysis=True,
+                    checkObstacle=True,
+                    # The pinned rectangular-filter image checks the complete
+                    # initial turn and primitive before selection. The upstream
+                    # angular wedge can wrongly exclude a clear straight escape
+                    # from an obstacle beside the rear corner.
+                    checkRotObstacle=False,
+                    adjacentRange=5.0,
+                    obstacleHeightThre=MAX_STEP_M,
+                    groundHeightThre=0.08,
+                    costHeightThre=0.08,
+                    useCost=True,
+                    pointPerPathThre=1,
+                    terrainVoxelSize=0.08,
+                    minRelZ=-0.5,
+                    maxRelZ=0.9,
+                    # Propose with a 56 cm half-width. The final swept square
+                    # check below covers front/rear corners and turning.
+                    pathScale=1.25,
+                    minPathScale=1.25,
+                    pathScaleBySpeed=False,
+                    pathRangeBySpeed=False,
+                    # Permit a safe short prefix when a full metre is obstructed.
+                    # The swept-body monitor still covers command latency and
+                    # braking; a prefix is not permission to cross its endpoint.
+                    # Upstream decrements range by 0.5 m by default, so merely
+                    # lowering the minimum skips every shorter candidate.
+                    minPathRange=0.2,
+                    pathRangeStep=0.1,
+                    dirThre=80.0,
+                    goalClearRange=0.0,
+                ),
+            ),
+            (
+                "local_planner",
+                "pathFollower",
+                dict(
+                    **common,
+                    lookAheadDis=0.7,
+                    yawRateGain=2.0,
+                    stopYawRateGain=2.0,
+                    maxYawRate=20.0,
+                    maxAccel=0.4,
+                    dirDiffThre=0.3,
+                    # The follower sees the cropped local prefix, not the
+                    # mission endpoint. Do not stop before its 0.2 m minimum;
+                    # waypoint arrival and the braking monitor remain separate.
+                    stopDisThre=0.08,
+                    slowDwnDisThre=0.7,
+                    useInclToStop=True,
+                    inclThre=30.0,
+                    stopTime=0.5,
+                    noRotAtGoal=True,
+                    pubSkipNum=0,
+                ),
+            ),
+        ]
+        for package, executable, parameters in configs:
+            args = ["ros2", "run", package, executable, "--ros-args"]
+            if executable == "terrainAnalysis":
+                args += ["-r", "/terrain_map:=/terrain_map_raw"]
+            for key, value in parameters.items():
+                args += ["-p", f"{key}:={str(value).lower() if isinstance(value, bool) else value}"]
+            self.processes.append(subprocess.Popen(args, start_new_session=True))
+
+    def stop_nodes(self):
+        for process in self.processes:
+            if process.poll() is None:
+                os.killpg(process.pid, signal.SIGTERM)
+        for process in self.processes:
+            try:
+                process.wait(timeout=3)
+            except subprocess.TimeoutExpired:
+                os.killpg(process.pid, signal.SIGKILL)
+                process.wait()
+        self.processes.clear()
+
+    def ready(self):
+        return (
+            len(self.processes) == 3
+            and all(p.poll() is None for p in self.processes)
+            and self.odom.get_subscription_count() >= 3
+            and self.scan.get_subscription_count() >= 2
+        )
+
+    def on_path(self, message):
+        with self.condition:
+            self.path = (
+                stamp_ns(message.header.stamp),
+                time.monotonic(),
+                [[p.pose.position.x, p.pose.position.y, p.pose.position.z] for p in message.poses],
+            )
+            self.condition.notify_all()
+
+    def on_command(self, message):
+        with self.condition:
+            self.command = (
+                stamp_ns(message.header.stamp),
+                time.monotonic(),
+                message.twist.linear.x,
+                message.twist.angular.z,
+            )
+            self.condition.notify_all()
+
+    def on_terrain(self, message):
+        started = time.monotonic()
+        points = point_cloud2.read_points_numpy(
+            message, field_names=["x", "y", "z", "intensity"], skip_nans=True
+        ).copy()
+        points, corrected = self.normalize_costs(points)
+        with self.condition:
+            # CMU round-trips the stamp through double seconds. Match the same
+            # sub-microsecond tolerance as the observation transaction below;
+            # never substitute an unrelated latest pose for a delayed map.
+            stamp = stamp_ns(message.header.stamp)
+            support = next(
+                (
+                    value
+                    for key, value in reversed(self.support_poses.items())
+                    if abs(key - stamp) <= 1000
+                ),
+                None,
+            )
+        underbody = 0
+        if support is not None:
+            points, underbody = underbody_support_costs(points, *support)
+        fields = [
+            PointField(name=name, offset=i * 4, datatype=PointField.FLOAT32, count=1)
+            for i, name in enumerate(("x", "y", "z", "intensity"))
+        ]
+        self.surface.publish(point_cloud2.create_cloud(message.header, fields, points))
+        with self.condition:
+            self.terrain = (stamp_ns(message.header.stamp), len(points), points)
+            self.slope_corrected = corrected
+            self.underbody_corrected = underbody
+            self.terrain_processing_ms = (time.monotonic() - started) * 1000
+            self.condition.notify_all()
+
+    def plan(self, value):
+        if not self.ready():
+            raise RuntimeError("navigation nodes are not ready")
+        points = np.asarray(value["points"], dtype=np.float32)
+        pose = np.asarray(value["pose"], dtype=np.float64)
+        goal = np.asarray(value["goal"], dtype=np.float64)
+        speed = float(value["max_speed_mps"])
+        reverse = value.get("allow_reverse", False)
+        contact_height = float(value.get("body_contact_height_m", 0.37))
+        if (
+            points.ndim != 2
+            or points.shape[1] != 3
+            or not 50 <= len(points) <= 30000
+            or pose.shape != (7,)
+            or goal.shape != (3,)
+            or not 0 <= speed <= 1
+            or not isinstance(reverse, bool)
+            or not math.isfinite(contact_height)
+            or not 0.1 <= contact_height <= 1.0
+            or not all(np.isfinite(v).all() for v in (points, pose, goal))
+            or abs(float(np.linalg.norm(pose[3:])) - 1) > 0.01
+        ):
+            raise ValueError("invalid range/odometry contract")
+        header = Header(stamp=self.get_clock().now().to_msg(), frame_id="map")
+        identity = stamp_ns(header.stamp)
+        with self.condition:
+            self.support_poses[identity] = (pose.copy(), contact_height)
+            while len(self.support_poses) > 8:
+                self.support_poses.popitem(last=False)
+        odom = Odometry(header=header, child_frame_id="vehicle")
+        odom.pose.pose.position.x, odom.pose.pose.position.y, odom.pose.pose.position.z = map(
+            float, pose[:3]
+        )
+        q = odom.pose.pose.orientation
+        q.x, q.y, q.z, q.w = map(float, pose[3:])
+        target = PointStamped(header=header)
+        target.point.x, target.point.y, target.point.z = map(float, goal)
+        fields = [
+            PointField(name=n, offset=i * 4, datatype=PointField.FLOAT32, count=1)
+            for i, n in enumerate(("x", "y", "z", "intensity"))
+        ]
+        cloud = point_cloud2.create_cloud(
+            header, fields, np.column_stack((points, np.zeros(len(points), np.float32)))
+        )
+        # The single HTTP writer establishes one observation transaction.
+        self.goal.publish(target)
+        self.speed.publish(Float32(data=speed))
+        self.odom.publish(odom)
+        self.scan.publish(cloud)
+        with self.condition:
+            fresh = self.condition.wait_for(
+                lambda: (
+                    self.path is not None
+                    and abs(self.path[0] - identity) <= 1000
+                    and self.command is not None
+                    and abs(self.command[0] - identity) <= 1000
+                    and self.command[1] >= self.path[1]
+                    and self.terrain is not None
+                    and abs(self.terrain[0] - identity) <= 1000
+                ),
+                # R26's accumulated 26k-point map needs ~0.33 s. Returning at
+                # 0.3 s perpetually abandons each matching observation just
+                # before its terrain/path arrives. Wait for that transaction,
+                # bounded below the independent 0.8 s camera deadman. A late
+                # result is still rejected by LatestInference, never reused.
+                timeout=0.6,
+            )
+            if not fresh:
+                return {
+                    "speed_mps": 0.0,
+                    "yaw_rate_rps": 0.0,
+                    "status": "waiting-for-plan",
+                    "path": [],
+                    "pending": {
+                        "path_stamp_delta_ns": None
+                        if self.path is None
+                        else self.path[0] - identity,
+                        "command_stamp_delta_ns": None
+                        if self.command is None
+                        else self.command[0] - identity,
+                        "terrain_stamp_delta_ns": None
+                        if self.terrain is None
+                        else self.terrain[0] - identity,
+                        "path_points": None if self.path is None else len(self.path[2]),
+                        "terrain_processing_ms": self.terrain_processing_ms,
+                        "terrain_points": None if self.terrain is None else self.terrain[1],
+                    },
+                    **(
+                        {"observed_terrain": self.terrain[2].tolist()}
+                        if value.get("include_terrain") is True and self.terrain is not None
+                        else {}
+                    ),
+                }
+            path, command = self.path, self.command
+            valid = len(path[2]) > 1 and all(math.isfinite(v) for v in command[2:])
+            velocity = max(-speed, min(speed, command[2]))
+            # Reverse is admitted only by the composed recovery policy after
+            # observing full-width support. Bound heading changes to that strip.
+            direction_clear = (
+                velocity <= 0 and abs(command[3]) <= 0.15 if reverse else velocity >= 0
+            )
+            velocity, yaw_rate, command_scale = (
+                regulate_command(velocity, command[3], self.terrain[2], pose)
+                if valid and direction_clear
+                else (0.0, 0.0, 0.0)
+            )
+            footprint_clear = command_scale > 0
+            valid = valid and footprint_clear and direction_clear
+            qx, qy, qz, qw = pose[3:]
+            tilt = math.degrees(math.acos(max(-1, min(1, 1 - 2 * (qx * qx + qy * qy)))))
+            failure = (
+                "inclination"
+                if tilt >= 30
+                else "no-path"
+                if len(path[2]) <= 1
+                else "footprint"
+                if not footprint_clear
+                else "direction"
+                if not direction_clear
+                else "controller-hold"
+                if abs(command[2]) + abs(command[3]) < 1e-5
+                else "none"
+            )
+            obstacles = self.terrain[2][self.terrain[2][:, 3] > MAX_STEP_M]
+            distances = np.linalg.norm(obstacles[:, :2] - pose[:2], axis=1)
+            near = obstacles[np.argsort(distances)[:12]]
+            return {
+                "speed_mps": velocity if valid else 0.0,
+                "yaw_rate_rps": max(-0.8, min(0.8, yaw_rate)) if valid else 0.0,
+                "status": "path" if valid else "blocked",
+                "path": path[2][::3],
+                "terrain_points": self.terrain[1],
+                "footprint_clear": bool(footprint_clear),
+                "path_frame": "vehicle-yaw",
+                "diagnostic": {
+                    "failure": failure,
+                    "tilt_degrees": tilt,
+                    "controller_command": list(command[2:]),
+                    "near_obstacles": near.tolist(),
+                    "slope_corrected_points": self.slope_corrected,
+                    "underbody_support_points": self.underbody_corrected,
+                    "terrain_processing_ms": self.terrain_processing_ms,
+                    "command_scale": command_scale,
+                },
+                # Engineering replay only; this local endpoint never forwards
+                # dense geometry to the operator or changes the control input.
+                **(
+                    {"observed_terrain": self.terrain[2].tolist()}
+                    if value.get("include_terrain") is True
+                    else {}
+                ),
+            }
+
+
+def main():
+    # All ROS traffic stays inside this container. Avoid persistent Fast DDS
+    # shared-memory segments across causal node resets on Docker/WSL.
+    os.environ["FASTRTPS_DEFAULT_PROFILES_FILE"] = str(Path(__file__).with_name("fastdds.xml"))
+    rclpy.init()
+    node = Navigation()
+    thread = threading.Thread(target=rclpy.spin, args=(node,), daemon=True)
+    thread.start()
+
+    class Handler(BaseHTTPRequestHandler):
+        def reply(self, status, value):
+            body = json.dumps(value, allow_nan=False).encode()
+            self.send_response(status)
+            self.send_header("Content-Type", "application/json")
+            self.send_header("Content-Length", str(len(body)))
+            self.end_headers()
+            self.wfile.write(body)
+
+        def do_GET(self):
+            self.reply(
+                200 if self.path == "/ready" and node.ready() else 503, {"ready": node.ready()}
+            )
+
+        def do_POST(self):
+            try:
+                size = int(self.headers.get("Content-Length", "0"))
+                if not 0 < size <= 3_000_000:
+                    raise ValueError("bounded JSON body required")
+                value = json.loads(self.rfile.read(size))
+                if self.path == "/reset":
+                    node.stop_nodes()
+                    self.reply(200, {"reset": True})
+                    # Reset DDS publishers/subscribers as well as child nodes.
+                    # Replacing PID 1 preserves container ownership and clears
+                    # all cached graph/history state before the next observation.
+                    os.execv(sys.executable, [sys.executable, str(Path(__file__).resolve())])
+                elif self.path == "/plan":
+                    self.reply(200, node.plan(value))
+                else:
+                    self.reply(404, {"error": "unknown endpoint"})
+            except (ValueError, KeyError, TypeError) as exc:
+                self.reply(400, {"error": str(exc)})
+            except Exception as exc:
+                self.reply(503, {"error": str(exc)})
+
+        def log_message(self, *_):
+            pass
+
+    try:
+        HTTPServer(("0.0.0.0", 8010), Handler).serve_forever()
+    finally:
+        node.stop_nodes()
+        rclpy.shutdown()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/simulation/ai-polygon/navigation/terrain_connectivity.cpp b/simulation/ai-polygon/navigation/terrain_connectivity.cpp
new file mode 100644
index 0000000..fabf9ea
--- /dev/null
+++ b/simulation/ai-polygon/navigation/terrain_connectivity.cpp
@@ -0,0 +1,23 @@
+// Exact close-pair hazard and connectedness test; same predicates as NumPy.
+// C ABI keeps this numerical hot loop independent of ROS and Python versions.
+#include <cmath>
+#include <vector>
+template<class T> int connected_grade(const T* p, int n) {
+    if (n <= 0) return 0;
+    std::vector<int> parent(n);
+    for(int i=0;i<n;i++) parent[i]=i;
+    auto root=[&](int i) { while(parent[i]!=i) {parent[i]=parent[parent[i]];i=parent[i];} return i; };
+    for(int i=0;i<n;i++) for(int j=i+1;j<n;j++) {
+        const T dx=p[3*i]-p[3*j], dy=p[3*i+1]-p[3*j+1];
+        const T distance=std::sqrt(dx*dx+dy*dy);
+        if(distance<=T(0.12)) {
+            if(std::abs(p[3*i+2]-p[3*j+2])>T(0.1001)) return 0;
+            parent[root(i)]=root(j);
+        }
+    }
+    const int first=root(0);
+    for(int i=1;i<n;i++) if(root(i)!=first) return 0;
+    return 1;
+}
+extern "C" int terrain_connected_f32(const float* p,int n) {return connected_grade(p,n);}
+extern "C" int terrain_connected_f64(const double* p,int n) {return connected_grade(p,n);}
diff --git a/simulation/ai-polygon/navigation/terrain_costs.py b/simulation/ai-polygon/navigation/terrain_costs.py
new file mode 100644
index 0000000..abc000e
--- /dev/null
+++ b/simulation/ai-polygon/navigation/terrain_costs.py
@@ -0,0 +1,170 @@
+"""Keep CMU height hazards, except an observed supported grade.
+
+Height above a cell's low quantile is not step height. A smooth 15-degree ramp
+can exceed 10 cm across that cell. A 6 cm voxel mesh also quantizes a continuous
+grade. Admit that surface only with broad support, a bounded plane residual and
+no observed short-range height jump exceeding the qualified 10 cm step. Vertical
+surfaces, excessive roughness and sparse/unknown support retain the CMU cost.
+"""
+
+import ctypes
+import hashlib
+import math
+from collections import OrderedDict
+from pathlib import Path
+
+import numpy as np
+
+_NATIVE_PATH = Path("/opt/missioncore/libterrain_connectivity.so")
+_NATIVE = ctypes.CDLL(str(_NATIVE_PATH)) if _NATIVE_PATH.is_file() else None
+if _NATIVE is not None:
+    for suffix, dtype in (("f32", ctypes.c_float), ("f64", ctypes.c_double)):
+        function = getattr(_NATIVE, "terrain_connected_" + suffix)
+        function.argtypes = [ctypes.POINTER(dtype), ctypes.c_int]
+        function.restype = ctypes.c_int
+
+
+def connected_grade(nearby):
+    if _NATIVE is not None and nearby.dtype in (np.dtype("float32"), np.dtype("float64")):
+        points = np.ascontiguousarray(nearby[:, :3])
+        dtype, suffix = (
+            (ctypes.c_float, "f32") if points.dtype.itemsize == 4 else (ctypes.c_double, "f64")
+        )
+        return bool(
+            getattr(_NATIVE, "terrain_connected_" + suffix)(
+                points.ctypes.data_as(ctypes.POINTER(dtype)), len(points)
+            )
+        )
+    # Reference implementation retained for portable CPU tests and comparison.
+    dx = nearby[:, None, 0] - nearby[None, :, 0]
+    dy = nearby[:, None, 1] - nearby[None, :, 1]
+    separation = np.sqrt(dx * dx + dy * dy)
+    jump = np.abs(nearby[:, None, 2] - nearby[None, :, 2])
+    if np.any((separation <= 0.12) & (jump > 0.1001)):
+        return False
+    connected = separation <= 0.12
+    reached = connected[np.argmin(np.linalg.norm(nearby[:, :2], axis=1))].copy()
+    while True:
+        expanded = np.any(connected[reached], axis=0)
+        if np.array_equal(expanded, reached):
+            return bool(reached.all())
+        reached = expanded
+
+
+def underbody_support_costs(terrain, pose, contact_height_m=0.37):
+    """Reconcile low returns already inside the current chassis footprint.
+
+    CMU's neighbourhood ground reference can label the supported terrain under
+    the chassis as a body collision. Use measured pose and the declared contact
+    height only inside the body (with a 5 cm inset), never for terrain ahead.
+    Retain drops, high returns and excessive tilt. The 8 cm band has 2 cm reserve
+    below the physically qualified 10 cm step; raw terrain memory is unchanged.
+    """
+    result = terrain.copy()
+    x, y, z, w = pose[3:]
+    up = np.array([2 * (x * z + w * y), 2 * (y * z - w * x), 1 - 2 * (x * x + y * y)])
+    if up[2] < math.cos(math.radians(25)):
+        return result, 0
+    yaw = math.atan2(2 * (w * z + x * y), 1 - 2 * (y * y + z * z))
+    axes = np.array([[math.cos(yaw), -math.sin(yaw)], [math.sin(yaw), math.cos(yaw)]])
+    local = (terrain[:, :2] - pose[:2]) @ axes
+    height = (terrain[:, :3] - pose[:3]) @ up + contact_height_m
+    supported = (
+        (np.abs(local) < 0.45).all(axis=1) & (np.abs(height) <= 0.08) & (terrain[:, 3] > 0.1)
+    )
+    result[supported, 3] = np.abs(height[supported])
+    return result, int(supported.sum())
+
+
+def _neighborhoods(terrain, radius=0.4):
+    """Exact radius neighborhoods without scanning the whole accumulated map.
+
+    Returns in the nine adjacent cells include every possible neighbor. Keep
+    source order so fitting and thresholds remain identical to the full scan.
+    """
+    cells = np.floor(terrain[:, :2] / radius).astype(np.int64)
+    buckets = {}
+    for index, (x, y) in enumerate(cells):
+        buckets.setdefault((x, y), []).append(index)
+    cached = {}
+
+    def around(index):
+        key = tuple(cells[index])
+        if key not in cached:
+            x, y = key
+            cached[key] = np.array(
+                sorted(
+                    i
+                    for dx in (-1, 0, 1)
+                    for dy in (-1, 0, 1)
+                    for i in buckets.get((x + dx, y + dy), ())
+                ),
+                dtype=np.int64,
+            )
+        delta = terrain[cached[key], :3] - terrain[index, :3]
+        return delta[np.linalg.norm(delta[:, :2], axis=1) <= radius]
+
+    return around
+
+
+def supported_slope_costs(terrain, cache=None):
+    result = terrain.copy()
+    if len(terrain) < 8:
+        return result, 0
+    around = _neighborhoods(terrain)
+    corrected = 0
+    for index in np.flatnonzero(terrain[:, 3] > 0.1):
+        nearby = around(index)
+        if len(nearby) < 8:
+            continue
+        # Any admitted plane spans at most a 0.8 m diameter at 25 degrees,
+        # plus the two 7.5 cm residuals. Reject tall foliage/walls before fitting.
+        if np.ptp(nearby[:, 2]) > 0.8 * math.tan(math.radians(25 + 1e-4)) + 0.15:
+            continue
+        key = None
+        if cache is not None:
+            key = hashlib.blake2b(nearby.tobytes(), digest_size=24).digest()
+            if key in cache:
+                if cache[key]:
+                    result[index, 3] = 0.0
+                    corrected += 1
+                cache.move_to_end(key)
+                continue
+            cache[key] = False
+            if len(cache) > 16384:
+                cache.popitem(last=False)
+        # No collinear strip, hidden region, multiple height layers or vertical
+        # surface is admitted as a plane. All observed points must agree.
+        covariance = np.cov(nearby[:, :2], rowvar=False)
+        if np.linalg.eigvalsh(covariance)[0] < 0.0036:
+            continue
+        matrix = np.column_stack((nearby[:, :2], np.ones(len(nearby))))
+        plane = np.linalg.lstsq(matrix, nearby[:, 2], rcond=None)[0]
+        slope = math.degrees(math.atan(np.linalg.norm(plane[:2])))
+        if not 2 <= slope <= 25 + 1e-4 or abs(plane[2]) > 0.05:
+            continue
+        if np.max(np.abs(matrix @ plane - nearby[:, 2])) > 0.075:
+            continue
+        # A permissive fit alone could erase a real ledge. Test close measured
+        # returns explicitly: even a narrow step/drop must retain its hazard.
+        if not connected_grade(nearby):
+            continue  # Do not fit a road across a gap with no returns.
+        result[index, 3] = 0.0
+        corrected += 1
+        if cache is not None:
+            cache[key] = True
+    return result, corrected
+
+
+class TerrainCostNormalizer:
+    """Reuse fits only for byte-identical observed neighborhoods, bounded in RAM.
+
+    New or changed returns always trigger a new fit. This stores no occupancy
+    belief and clears with the owning ROS node on every episode/reset.
+    """
+
+    def __init__(self):
+        self.cache = OrderedDict()
+
+    def __call__(self, terrain):
+        return supported_slope_costs(terrain, self.cache)
diff --git a/simulation/ai-polygon/navigation_client.py b/simulation/ai-polygon/navigation_client.py
new file mode 100644
index 0000000..f0884c0
--- /dev/null
+++ b/simulation/ai-polygon/navigation_client.py
@@ -0,0 +1,433 @@
+"""Installed provider adapters executed in shared composition dependency order."""
+
+import http.client
+import json
+import math
+import time
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+
+import numpy as np
+
+from k1link.simulation.ai_polygon.composition import compose
+from k1link.simulation.ai_polygon.contracts import Decision
+from k1link.simulation.ai_polygon.inference import ModelInference
+from k1link.simulation.ai_polygon.mission_policy import WaypointMission
+
+
+def project(points, calibration):
+    rotation = np.asarray(calibration["rotation"], dtype=np.float64).reshape(3, 3)
+    local = (points - np.asarray(calibration["origin"])) @ rotation
+    front = local[:, 0] > 0.1
+    depth = np.maximum(local[:, 0], 0.1)
+    fx, fy, cx, cy = calibration["intrinsics"]
+    u = cx - fx * local[:, 1] / depth
+    v = cy - fy * local[:, 2] / depth
+    visible = front & (u >= 100) & (u < 700) & (v >= 0) & (v < 600)
+    return local, u, v, visible
+
+
+def visual_goal(surface, points, pose, calibration, prior=None, target=None, excluded=()):
+    """Project semantic surface candidates into range; CMU owns path search."""
+    local, u, v, visible = project(points, calibration)
+    semantic = np.zeros(len(points), dtype=bool)
+    semantic[visible] = surface[
+        ((v[visible] * 512 / 600).astype(int)), (((u[visible] - 100) * 512 / 600).astype(int))
+    ]
+    ground_z = pose[2] - calibration.get("body_contact_height_m", 0.27)
+    distance = np.linalg.norm(points[:, :2] - np.asarray(pose[:2]), axis=1)
+    slope = np.abs(points[:, 2] - ground_z) / np.maximum(distance, 0.1)
+    good = (
+        visible
+        & semantic
+        & (distance >= (0.4 if target is not None else 1.2))
+        & (distance <= 3.5)
+        & (slope <= np.tan(np.radians(25)))
+        & (local[:, 2] < 0)
+    )
+    candidates = points[good]
+    # A pixel at the centre is insufficient for a metre-wide chassis. Require
+    # visual surface support on both sides, at the same candidate ground level.
+    if len(candidates):
+        delta = candidates[:, :2] - np.asarray(pose[:2])
+        perpendicular = np.column_stack((-delta[:, 1], delta[:, 0]))
+        perpendicular /= np.maximum(np.linalg.norm(perpendicular, axis=1)[:, None], 0.1)
+        supported = np.ones(len(candidates), dtype=bool)
+        for side in (-0.5, 0.5):
+            edge = candidates.copy()
+            edge[:, :2] += perpendicular * side
+            _, eu, ev, in_view = project(edge, calibration)
+            ok = np.zeros(len(edge), dtype=bool)
+            ok[in_view] = surface[
+                (ev[in_view] * 512 / 600).astype(int), ((eu[in_view] - 100) * 512 / 600).astype(int)
+            ]
+            supported &= ok
+        for failed in excluded:
+            supported &= np.linalg.norm(candidates[:, :2] - np.asarray(failed[:2]), axis=1) > 0.6
+        good[np.flatnonzero(good)[~supported]] = False
+        candidates = points[good]
+    if len(candidates) < 8:
+        return None
+    if target is not None:
+        target = np.asarray(target)
+        if prior is not None and np.linalg.norm(np.asarray(prior[:2]) - target) < 1e-4:
+            direction = target - pose[:2]
+            remaining = np.linalg.norm(direction)
+            lateral = np.array([-direction[1], direction[0]]) / max(remaining, 0.1)
+            footprint = np.tile(prior, (3, 1))
+            footprint[:, :2] += np.array([-0.5, 0, 0.5])[:, None] * lateral
+            _, pu, pv, in_view = project(footprint, calibration)
+            # The measured camera projection defines its blind strip; a fixed
+            # 0.8 m cutoff forgot goals that disappeared around 0.95 m. Retain
+            # only the exact, previously observed task goal. Any still-visible
+            # non-drivable part invalidates it; live CMU range/collision checks
+            # remain mandatory before every motion command.
+            if 0.35 < remaining <= 3.5 and not in_view.all():
+                if surface[
+                    (pv[in_view] * 512 / 600).astype(int),
+                    ((pu[in_view] - 100) * 512 / 600).astype(int),
+                ].all():
+                    return prior
+                return None
+        near_target = np.linalg.norm(candidates[:, :2] - target, axis=1) <= 0.25
+        if near_target.sum() >= 3:
+            # Once the requested waypoint has observed semantic/range support,
+            # keep its XY identity. A moving pixel median can slide laterally
+            # past the arrival radius, leaving CMU chasing successive forward
+            # goals after the actual task waypoint is already behind the rover.
+            requested = np.array([*target.tolist(), float(np.median(candidates[near_target, 2]))])
+            direction = target - pose[:2]
+            lateral = np.array([-direction[1], direction[0]]) / max(
+                np.linalg.norm(direction), 0.1
+            )
+            footprint = np.tile(requested, (3, 1))
+            footprint[:, :2] += np.array([-0.5, 0, 0.5])[:, None] * lateral
+            _, tu, tv, in_view = project(footprint, calibration)
+            if in_view.all() and surface[
+                (tv * 512 / 600).astype(int), ((tu - 100) * 512 / 600).astype(int)
+            ].all():
+                return requested.tolist()
+    if prior is not None and target is not None:
+        remaining = np.linalg.norm(np.asarray(prior[:2]) - pose[:2])
+        # Finish an already observed close waypoint after it enters the camera's
+        # under-body blind strip. Current visual support is still required above;
+        # CMU's causal terrain memory and live collision monitor retain authority.
+        if 0.35 < remaining <= 0.8:
+            return prior
+    # Retain a still-observed waypoint until reached; never keep a stale visual
+    # goal after the supporting surface disappears or changes classification.
+    if (
+        prior is not None
+        and np.linalg.norm(np.asarray(prior[:2]) - pose[:2]) > 0.8
+        and np.linalg.norm(candidates[:, :2] - np.asarray(prior[:2]), axis=1).min() < 0.35
+    ):
+        return prior
+    good_local = local[good]
+    score = np.abs(distance[good] - 2.3) + 2.0 * np.abs(
+        np.arctan2(good_local[:, 1], good_local[:, 0])
+    )
+    if target is not None:
+        remaining = np.linalg.norm(target - pose[:2])
+        target_distance = np.linalg.norm(candidates[:, :2] - target, axis=1)
+        progress = target_distance < remaining - 0.1
+        if not progress.any():
+            return None
+        score = target_distance + 0.2 * score
+        score[~progress] = np.inf
+    center = candidates[int(np.argmin(score))]
+    neighbors = candidates[np.linalg.norm(candidates[:, :2] - center[:2], axis=1) < 0.3]
+    return np.median(neighbors, axis=0).tolist()
+
+
+def recovery_goal(points, pose, calibration, prior=None):
+    """A short straight retreat requires fresh support across its whole width.
+
+    This is admission of a task goal, not a replacement for CMU. Its obstacle
+    map, path selector and swept-body monitor still reject the actual command.
+    Missing returns, ledges, steep or discontinuous ground forbid retreat.
+    """
+    points = np.asarray(points)
+    qx, qy, qz, qw = pose[3:]
+    yaw = math.atan2(2 * (qw * qz + qx * qy), 1 - 2 * (qy * qy + qz * qz))
+    axes = np.array([[math.cos(yaw), -math.sin(yaw)], [math.sin(yaw), math.cos(yaw)]])
+    local = (points[:, :2] - np.asarray(pose[:2])) @ axes
+    if prior is None:
+        distance = 0.65
+    else:
+        offset = (np.asarray(prior[:2]) - np.asarray(pose[:2])) @ axes
+        distance = -float(offset[0])
+        if not 0 < distance <= 0.8 or abs(offset[1]) > 0.1:
+            return None
+    ground = pose[2] - calibration["body_contact_height_m"]
+    heights = []
+    # 15 cm lateral reserve covers the small heading correction allowed during
+    # reverse. Never extrapolate through an unobserved cell behind either wheel.
+    for x in np.arange(-0.6 - distance, -0.5, 0.15):
+        for y in np.arange(-0.6, 0.61, 0.15):
+            near = np.linalg.norm(local - [x, y], axis=1) <= 0.18
+            if near.sum() < 3:
+                return None
+            z = points[near, 2]
+            if np.ptp(z) > 0.1 or abs(float(np.median(z)) - ground) > 0.1 + abs(x) * math.tan(
+                math.radians(25)
+            ):
+                return None
+            heights.append([x, y, float(np.median(z))])
+    samples = np.asarray(heights)
+    design = np.column_stack((samples[:, :2], np.ones(len(samples))))
+    plane = np.linalg.lstsq(design, samples[:, 2], rcond=None)[0]
+    if np.linalg.norm(plane[:2]) > math.tan(math.radians(25)):
+        return None
+    if np.max(np.abs(samples[:, 2] - design @ plane)) > 0.05:
+        return None
+    if prior is not None:
+        return prior
+    xy = np.asarray(pose[:2]) + axes @ [-distance, 0]
+    return [*xy.tolist(), float(plane[2] - distance * plane[0])]
+
+
+class NavigationClient:
+    def __init__(self):
+        self.connection = http.client.HTTPConnection("127.0.0.1", 18093, timeout=2)
+
+    def request(self, path, body=None):
+        self.connection.request(
+            "GET" if body is None else "POST",
+            path,
+            body=None if body is None else json.dumps(body, allow_nan=False),
+            headers={"Content-Type": "application/json"},
+        )
+        response = self.connection.getresponse()
+        raw = response.read(3_000_001)
+        if response.status != 200 or len(raw) > 3_000_000:
+            raise RuntimeError("Local CMU navigation is unavailable")
+        return json.loads(raw)
+
+    def close(self):
+        self.connection.close()
+
+
+class ComposedInference:
+    def __init__(self, root: Path, run, directory: Path, mission=None):
+        profile = json.loads((root / "models.worker-006.json").read_text())
+        self.graph = compose(root, run["request"].get("composition"))
+        self.segmenter_id = next(
+            n.module.module_id for n in self.graph.nodes if n.module.group == "segmentation"
+        )
+        self.models = ModelInference(
+            "http://127.0.0.1:18092",
+            Path(profile["labels"]),
+            "http://127.0.0.1:18091",
+            segmenter_id=self.segmenter_id,
+        )
+        self.navigation = NavigationClient()
+        self.max_speed = run["world"]["settings"]["max_speed_mps"]
+        self.goal = None
+        self.mission = mission or WaypointMission(run["world"]["settings"].get("route_xy", []))
+        self.navigation_evidence = {}
+        self.pool = ThreadPoolExecutor(max_workers=2, thread_name_prefix="simulation-module")
+        self.directory = directory
+        self.providers = {
+            "simulation-ddrnet-goose": lambda inputs: self.models.surface(
+                inputs["source.camera.rgb"]
+            ),
+            "simulation-segformer-ade": lambda inputs: self.models.surface(
+                inputs["source.camera.rgb"]
+            ),
+            "simulation-rf-detr": lambda inputs: {
+                "detection.boxes": self.models.detect(inputs["source.camera.rgb"])
+            },
+            "simulation-cmu-navigation": self.navigate,
+            "simulation-waypoint-mission": self.plan_mission,
+        }
+        if {node.module.module_id for node in self.graph.nodes} - set(self.providers):
+            raise ValueError("Composition contains an uninstalled executor")
+        directory.mkdir(exist_ok=True, parents=True)
+        (directory.parent / "composition.json").write_text(
+            json.dumps(
+                {
+                    "sha256": self.graph.sha256,
+                    "graph": self.graph.as_dict(),
+                    "modules": [node.module.identity_document() for node in self.graph.nodes],
+                },
+                indent=2,
+            ),
+            encoding="utf-8",
+        )
+
+    def ready(self):
+        self.models.ready()
+        self.navigation.request("/ready")
+
+    def reset(self):
+        self.goal = None
+        self.mission.resume()
+        self.navigation.request("/reset", {})
+        deadline = time.monotonic() + 8
+        while time.monotonic() < deadline:
+            try:
+                self.navigation.request("/ready")
+                return
+            except (OSError, RuntimeError):
+                time.sleep(0.1)
+        raise RuntimeError("Navigation reset timed out")
+
+    def plan_mission(self, inputs):
+        self.goal, intent = self.mission.update(
+            inputs["source.pose"],
+            inputs["source.simulation-time"],
+            lambda target, prior, excluded: visual_goal(
+                inputs["segmentation.surface"],
+                inputs["source.lidar"],
+                inputs["source.pose"],
+                inputs["source.camera.calibration"],
+                prior,
+                target,
+                excluded,
+            ),
+            lambda prior: recovery_goal(
+                inputs["source.lidar"],
+                inputs["source.pose"],
+                inputs["source.camera.calibration"],
+                prior,
+            ),
+        )
+        return {"navigation.goal": self.goal, "navigation.intent": intent}
+
+    def navigate(self, inputs):
+        surface, boxes = inputs["segmentation.surface"], inputs["detection.boxes"]
+        points, pose = inputs["source.lidar"], inputs["source.pose"]
+        calibration = inputs["source.camera.calibration"]
+        intent = inputs["navigation.intent"]
+        fraction = float(surface[320:500, 100:412].mean())
+        reason, result = "uncertain", {"speed_mps": 0.0, "yaw_rate_rps": 0.0, "path": []}
+        if len(points) >= 50:
+            local, u, v, visible = project(points, calibration)
+            distance = np.linalg.norm(points[:, :2] - np.asarray(pose[:2]), axis=1)
+            danger = False
+            for x1, y1, x2, y2 in boxes:
+                covered = (
+                    visible & (u >= 800 * x1) & (u <= 800 * x2) & (v >= 600 * y1) & (v <= 600 * y2)
+                )
+                if np.any(
+                    covered & (distance < 1.5) & (local[:, 0] > 0) & (np.abs(local[:, 1]) < 0.8)
+                ):
+                    danger = True
+            if intent["state"] in {"stuck", "goal-reached", "unstable"}:
+                reason = intent["state"]
+            elif danger:
+                reason = "obstacle"
+            elif self.goal is None:
+                reason = "no-road"
+            else:
+                result = self.navigation.request(
+                    "/plan",
+                    {
+                        "points": points.tolist(),
+                        "pose": pose,
+                        "body_contact_height_m": calibration["body_contact_height_m"],
+                        "goal": self.goal,
+                        "max_speed_mps": min(self.max_speed, 0.1)
+                        if intent["state"] == "reversing"
+                        else self.max_speed,
+                        "allow_reverse": intent["state"] == "reversing",
+                    },
+                )
+                reason = {
+                    "path": "road",
+                    "blocked": "obstacle",
+                    "waiting-for-plan": "waiting",
+                }.get(result["status"], "uncertain")
+                if (
+                    reason == "road"
+                    and abs(result["speed_mps"]) + abs(result["yaw_rate_rps"]) < 1e-5
+                ):
+                    reason = "waiting"
+                elif reason == "road" and intent["state"] in {"replanning", "reversing"}:
+                    reason = "replanning"
+        self.navigation_evidence = {
+            "mission": intent,
+            "navigation": {k: v for k, v in result.items() if k != "path"},
+        }
+        decision = Decision(
+            speed_mps=result["speed_mps"],
+            yaw_rate_rps=result["yaw_rate_rps"],
+            reason=intent["state"]
+            if intent["state"] in {"stuck", "goal-reached", "unstable"}
+            else reason,
+            road_fraction=fraction,
+            obstacle_count=len(boxes),
+        )
+        return {"motion.command": decision.model_dump(), "motion.path": result["path"]}
+
+    def infer_observation(self, rgb, observation, frame_id):
+        values = {
+            "source.camera.rgb": rgb,
+            "source.lidar": observation["points"],
+            "source.pose": observation["pose"],
+            "source.camera.calibration": observation["calibration"],
+            "source.simulation-time": observation["simulation_time_ns"] / 1e9,
+        }
+        timings = {}
+        pending = list(self.graph.nodes)
+
+        def execute(node, inputs):
+            started = time.monotonic()
+            output = self.providers[node.module.module_id](inputs)
+            if set(output) != set(node.module.provides):
+                raise ValueError("Provider output does not match its composition contract")
+            return output, (time.monotonic() - started) * 1000
+
+        while pending:
+            ready = [node for node in pending if all(port in values for port, _ in node.inputs)]
+            if not ready:
+                raise ValueError("Unresolved composition input")
+            tasks = [
+                (
+                    node,
+                    self.pool.submit(
+                        execute, node, {port: values[port] for port, _ in node.inputs}
+                    ),
+                )
+                for node in ready
+            ]
+            for node, task in tasks:
+                output, elapsed = task.result()
+                values.update(output)
+                timings[node.module.module_id] = elapsed
+                pending.remove(node)
+        from PIL import Image
+
+        Image.fromarray(values["segmentation.labels"]).save(
+            self.directory / f"{frame_id:08d}.labels.png"
+        )
+        Image.fromarray(values["segmentation.surface"].astype(np.uint8) * 255).save(
+            self.directory / f"{frame_id:08d}.surface.png"
+        )
+        np.savez_compressed(
+            self.directory / f"{frame_id:08d}.range.npz",
+            points=observation["points"],
+            pose=observation["pose"],
+            origin=observation["calibration"]["origin"],
+            rotation=observation["calibration"]["rotation"],
+            intrinsics=observation["calibration"]["intrinsics"],
+            body_contact_height_m=observation["calibration"]["body_contact_height_m"],
+            range_origins=observation["calibration"].get("range_origins", []),
+        )
+        return (
+            values["motion.command"],
+            values["detection.boxes"],
+            {
+                "composition_sha256": self.graph.sha256,
+                "module_ms": timings,
+                "goal": self.goal,
+                "path": values["motion.path"],
+                "range_points": len(observation["points"]),
+                **self.navigation_evidence,
+            },
+        )
+
+    def close(self):
+        self.pool.shutdown(wait=True, cancel_futures=True)
+        self.models.close()
+        self.navigation.close()
diff --git a/simulation/ai-polygon/prepare_terrain.py b/simulation/ai-polygon/prepare_terrain.py
new file mode 100644
index 0000000..e2cf3b2
--- /dev/null
+++ b/simulation/ai-polygon/prepare_terrain.py
@@ -0,0 +1,78 @@
+"""Versioned Worker-only collision preparation; never uploads generated assets."""
+
+import hashlib
+import json
+import subprocess
+import time
+from pathlib import Path
+
+from local_state import write_json
+
+from k1link.simulation.ai_polygon.terrain_contract import terrain_matches
+
+
+def prepare_terrain(root: Path, episode: Path, world: dict, job=None, pulse=None) -> Path:
+    if world.get("storage", {}).get("kind") == "worker":
+        path = root / "assets/prepared-worlds" / world["sha256"] / "terrain.json"
+        manifest = json.loads(path.read_text(encoding="utf-8-sig"))
+        if (
+            manifest.get("generator") != "paired-source"
+            or manifest.get("collider_sha256") != world.get("collider_sha256")
+            or not terrain_matches(manifest, world)
+        ):
+            raise ValueError("Paired collision asset does not match the admitted world")
+        return path  # install_terrain hashes the actual collider before physics.
+    generator = Path(__file__).parent / "navigation/Prepare-Terrain.ps1"
+    generator_sha256 = hashlib.sha256(generator.read_bytes()).hexdigest()
+
+    def matching():
+        for path in (root / "assets/terrain-v1").glob(world["sha256"] + "-*/terrain.json"):
+            value = json.loads(path.read_text(encoding="utf-8-sig"))
+            if terrain_matches(value, world, generator_sha256):
+                return path
+        return None
+
+    found = matching()
+    if found:
+        return found
+    world_path = episode / "terrain-world.json"
+    write_json(world_path, world)
+    with (episode / "terrain-preparation.log").open("wb") as log:
+        process = subprocess.Popen(
+            [
+                "powershell.exe",
+                "-NoProfile",
+                "-NonInteractive",
+                "-ExecutionPolicy",
+                "Bypass",
+                "-File",
+                str(generator),
+                "-Root",
+                str(root),
+                "-WorldFile",
+                str(world_path),
+            ],
+            stdout=log,
+            stderr=subprocess.STDOUT,
+        )
+        if job is not None:
+            job.assign(process)
+        deadline = time.monotonic() + 180
+        try:
+            while process.poll() is None:
+                if time.monotonic() > deadline:
+                    raise TimeoutError("Terrain preparation timed out")
+                if pulse:
+                    pulse()
+                time.sleep(1)
+            if process.returncode != 0:
+                raise RuntimeError("Terrain preparation failed; inspect the Worker episode log")
+        finally:
+            if process.poll() is None:
+                from worker import terminate_episode
+
+                terminate_episode(process)
+    found = matching()
+    if found is None:
+        raise RuntimeError("Terrain preparation did not produce a matching collider")
+    return found
diff --git a/simulation/ai-polygon/qualify_recovery_sensor.py b/simulation/ai-polygon/qualify_recovery_sensor.py
new file mode 100644
index 0000000..3d62f06
--- /dev/null
+++ b/simulation/ai-polygon/qualify_recovery_sensor.py
@@ -0,0 +1,84 @@
+"""Worker-only regression: observed retreat with the shipped body and sensors."""
+
+import argparse
+import hashlib
+import json
+import sys
+import time
+from datetime import UTC, datetime
+from pathlib import Path
+
+from isaacsim import SimulationApp
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--output", type=Path, required=True)
+args = parser.parse_args()
+sys.path.insert(0, str(Path(__file__).resolve().parents[2] / "src"))
+app = SimulationApp({"headless": True, "hide_ui": True})
+try:
+    import numpy as np
+    import omni.timeline
+    import omni.usd
+    from isaacsim.core.experimental.prims import Articulation
+    from isaacsim.core.simulation_manager import SimulationManager
+    from navigation_client import recovery_goal
+    from pxr import Gf, UsdGeom, UsdPhysics
+    from rover_profile import PROFILE, DifferentialDrive, create_rover
+    from terrain import RangeSensor
+
+    stage = omni.usd.get_context().get_stage()
+    UsdGeom.SetStageUpAxis(stage, UsdGeom.Tokens.z)
+    UsdGeom.SetStageMetersPerUnit(stage, 1)
+    floor = UsdGeom.Cube.Define(stage, "/World/Ground")
+    floor.CreateSizeAttr(1)
+    floor.AddTranslateOp().Set(Gf.Vec3d(0, 0, -0.1))
+    floor.AddScaleOp().Set(Gf.Vec3f(8, 8, 0.2))
+    UsdPhysics.CollisionAPI.Apply(floor.GetPrim())
+    create_rover(stage, [0, 0, 0.29], 0)
+    SimulationManager.setup_simulation(dt=1 / 60, device="cpu")
+    omni.timeline.get_timeline_interface().play()
+    for _ in range(30):
+        app.update()
+    robot = Articulation("/World/Rover")
+    sensor = RangeSensor()
+    drive = DifferentialDrive(robot)
+    rows = []
+    for tick in range(25):
+        p, q = (v.numpy()[0] for v in robot.get_world_poses())
+        transform = Gf.Matrix4d(1).SetRotate(Gf.Quatd(float(q[0]), Gf.Vec3d(*map(float, q[1:]))))
+        transform.SetTranslateOnly(Gf.Vec3d(*map(float, p)))
+        origins = [
+            transform.Transform(Gf.Vec3d(x, 0, 0.8 - PROFILE["body_contact_height_m"]))
+            for x in (PROFILE["camera_forward_m"], PROFILE["rear_range_forward_m"])
+        ]
+        points = sensor.capture(origins[0], transform, origins[1])
+        pose = [*p.tolist(), *q[1:].tolist(), float(q[0])]
+        goal = recovery_goal(
+            points, pose, {"body_contact_height_m": PROFILE["body_contact_height_m"]}
+        )
+        rows.append({"seconds": tick / 5, "pose": pose, "returns": len(points), "goal": goal})
+        drive.command(-0.1 if goal is not None else 0)
+        SimulationManager.step(steps=12)
+    drive.command(0)
+    before = robot.get_world_poses()[0].numpy()[0].copy()
+    SimulationManager.step(steps=120)
+    end = robot.get_world_poses()[0].numpy()[0]
+    drift = float(np.linalg.norm(end - before))
+    passed = all(row["goal"] is not None for row in rows) and end[0] < -0.4 and drift < 0.03
+    report = {
+        "utc": datetime.now(UTC).isoformat(),
+        "monotonic": time.monotonic(),
+        "profile": PROFILE,
+        "passed": bool(passed),
+        "end": end.tolist(),
+        "braking_drift_m": drift,
+        "samples": rows,
+    }
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(json.dumps(report, indent=2), encoding="utf-8")
+    args.output.with_suffix(".sha256").write_text(
+        hashlib.sha256(args.output.read_bytes()).hexdigest(), encoding="ascii"
+    )
+    print(json.dumps({k: v for k, v in report.items() if k not in {"samples", "profile"}}))
+finally:
+    app.close()
diff --git a/simulation/ai-polygon/qualify_rover.py b/simulation/ai-polygon/qualify_rover.py
new file mode 100644
index 0000000..43068b4
--- /dev/null
+++ b/simulation/ai-polygon/qualify_rover.py
@@ -0,0 +1,103 @@
+"""Bounded Worker-only drive/braking test, independent of AI and Gaussian mesh."""
+
+import argparse
+import hashlib
+import json
+import math
+import time
+from datetime import UTC, datetime
+from pathlib import Path
+
+from isaacsim import SimulationApp
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--output", type=Path, required=True)
+args = parser.parse_args()
+
+app = SimulationApp({"headless": True, "hide_ui": True})
+try:
+    import omni.timeline
+    import omni.usd
+    from isaacsim.core.experimental.prims import Articulation
+    from isaacsim.core.simulation_manager import SimulationManager
+    from pxr import Gf, UsdGeom, UsdPhysics, UsdShade
+    from rover_profile import PROFILE, DifferentialDrive, create_rover
+
+    stage = omni.usd.get_context().get_stage()
+    UsdGeom.SetStageUpAxis(stage, UsdGeom.Tokens.z)
+    UsdGeom.SetStageMetersPerUnit(stage, 1)
+    ground = UsdGeom.Cube.Define(stage, "/World/Ground")
+    ground.CreateSizeAttr(1)
+    ground.AddTranslateOp().Set(Gf.Vec3d(0, 0, -0.1))
+    ground.AddScaleOp().Set(Gf.Vec3f(20, 20, 0.2))
+    UsdPhysics.CollisionAPI.Apply(ground.GetPrim())
+    material = UsdShade.Material.Define(stage, "/World/Material")
+    physics = UsdPhysics.MaterialAPI.Apply(material.GetPrim())
+    physics.CreateStaticFrictionAttr(0.9)
+    physics.CreateDynamicFrictionAttr(0.8)
+    UsdShade.MaterialBindingAPI.Apply(ground.GetPrim()).Bind(
+        material, UsdShade.Tokens.weakerThanDescendants, "physics"
+    )
+    create_rover(stage, [0, 0, PROFILE["wheel_radius_m"] + 0.05], 0)
+    SimulationManager.setup_simulation(dt=1 / 60, device="cpu")
+    omni.timeline.get_timeline_interface().play()
+    for _ in range(20):
+        app.update()
+    robot = Articulation("/World/Rover")
+    indices = robot.get_dof_indices(PROFILE["wheel_names"])
+    drive = DifferentialDrive(robot)
+
+    def position():
+        # Tensor pose is authoritative, independent of USD/Fabric writeback.
+        return robot.get_world_poses()[0].numpy().tolist()[0]
+
+    start = position()
+    drive.command(0.3)
+    SimulationManager.step(steps=300)
+    moved = position()
+    drive.command(0)
+    SimulationManager.step(steps=120)
+    stopped = position()
+    turns = []
+
+    def heading():
+        w, x, y, z = robot.get_world_poses()[1].numpy()[0].tolist()
+        return math.atan2(2 * (w * z + x * y), 1 - 2 * (y * y + z * z))
+
+    for direction in (1, -1):
+        before, angle = position(), heading()
+        drive.command(0, direction * 0.35)
+        SimulationManager.step(steps=180)
+        turn = math.atan2(math.sin(heading() - angle), math.cos(heading() - angle))
+        after = position()
+        drive.command(0)
+        SimulationManager.step(steps=120)
+        turns.append(
+            dict(
+                command_yaw_rps=direction * 0.35,
+                measured_yaw_radians=turn,
+                displacement_m=math.dist(before[:2], after[:2]),
+                brake_drift_m=math.dist(after, position()),
+                passed=direction * turn > 0.5 and math.dist(before[:2], after[:2]) < 0.15,
+            )
+        )
+    report = dict(
+        utc=datetime.now(UTC).isoformat(),
+        monotonic=time.monotonic(),
+        profile_sha256=hashlib.sha256(
+            Path(__file__).with_name("rover_profile.py").read_bytes()
+        ).hexdigest(),
+        profile=PROFILE,
+        start=start,
+        after_5s=moved,
+        after_stop_2s=stopped,
+        turns=turns,
+        passed=moved[0] - start[0] > 1.0
+        and abs(stopped[0] - moved[0]) < 0.15
+        and all(turn["passed"] for turn in turns),
+        wheel_velocity_rps=robot.get_dof_velocities(dof_indices=indices).numpy().tolist(),
+    )
+    args.output.write_text(json.dumps(report, indent=2), encoding="utf-8")
+    print(json.dumps(report))
+finally:
+    app.close()
diff --git a/simulation/ai-polygon/qualify_runtime.py b/simulation/ai-polygon/qualify_runtime.py
new file mode 100644
index 0000000..00376a4
--- /dev/null
+++ b/simulation/ai-polygon/qualify_runtime.py
@@ -0,0 +1,116 @@
+"""Bounded synthetic camera/model probe. This is NOT navigation acceptance.
+
+Run with Isaac's python.bat after reserving Worker through the Core job queue.
+Outputs stay private in the chosen evidence directory.
+"""
+
+import argparse
+import hashlib
+import json
+import sys
+import time
+from datetime import UTC, datetime
+from pathlib import Path
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--output", type=Path, required=True)
+parser.add_argument("--models", action="store_true")
+args = parser.parse_args()
+args.output.mkdir(parents=True, exist_ok=False)
+sys.path.insert(0, str(Path(__file__).resolve().parents[2] / "src"))
+report = {
+    "schema_version": "missioncore.ai-polygon-runtime-probe/v1",
+    "started_at": datetime.now(UTC).isoformat(),
+    "started_monotonic_ns": time.monotonic_ns(),
+    "source": "synthetic-cube-only",
+    "navigation_accepted": False,
+    "passed": False,
+}
+app = stack = inference = None
+try:
+    from isaacsim import SimulationApp
+
+    app = SimulationApp({"headless": True, "multi_gpu": False, "width": 800, "height": 600})
+    import numpy as np
+    import omni.usd
+    from isaacsim.core.experimental.utils import app as app_utils
+    from isaacsim.core.simulation_manager import SimulationManager
+    from isaacsim.sensors.experimental.rtx import CameraSensor, RtxCamera
+    from PIL import Image
+    from pxr import Gf, UsdGeom, UsdLux
+
+    stage = omni.usd.get_context().get_stage()
+    UsdGeom.SetStageUpAxis(stage, UsdGeom.Tokens.z)
+    UsdGeom.SetStageMetersPerUnit(stage, 1.0)
+    cube = UsdGeom.Cube.Define(stage, "/World/Cube")
+    cube.AddTranslateOp().Set(Gf.Vec3d(0, 3, 0.5))
+    cube.CreateDisplayColorAttr([(0.8, 0.15, 0.05)])
+    UsdLux.DomeLight.Define(stage, "/World/Light").CreateIntensityAttr(500)
+    camera = RtxCamera(
+        "/World/Camera",
+        tick_rate=10,
+        translations=np.array([0.0, 0.0, 0.5]),
+        orientations=np.array([1.0, 1.0, 0.0, 0.0]) / np.sqrt(2),
+    )
+    camera.camera.set_focal_lengths(24.0)
+    sensor = CameraSensor(camera, resolution=(600, 800), annotators=["rgb"])
+    SimulationManager.setup_simulation(dt=1 / 60, device="cpu")
+    app_utils.play()
+    app_utils.update_app(steps=12)
+    app_utils.pause()
+    baseline = SimulationManager.get_num_physics_steps()
+    for _ in range(10):
+        app.update()
+    assert SimulationManager.get_num_physics_steps() == baseline, "Render advanced physics"
+    raw, _ = sensor.get_data("rgb")
+    assert raw is not None, "RTX camera did not produce a frame"
+    rgb = np.ascontiguousarray(raw.numpy()[:, :, :3])
+    assert rgb.shape == (600, 800, 3) and rgb.dtype == np.uint8
+    assert float(rgb.std()) > 1, "Camera frame is empty/uniform"
+    image_path = args.output / "synthetic-camera.png"
+    Image.fromarray(rgb).save(image_path)
+    report.update(
+        frame_sha256=hashlib.sha256(image_path.read_bytes()).hexdigest(),
+        frame_shape=list(rgb.shape),
+        paused_physics_steps=baseline,
+    )
+    SimulationManager.step(steps=6)
+    assert SimulationManager.get_num_physics_steps() == baseline + 6, "Wrong lockstep increment"
+    report["lockstep_physics_steps"] = 6
+    if args.models:
+        from model_stack import ModelStack
+
+        from k1link.simulation.ai_polygon.inference import ModelInference
+        from k1link.simulation.ai_polygon.policy import RoadPolicy
+
+        stack = ModelStack()
+        stack.start()
+        inference = ModelInference(
+            "http://127.0.0.1:18092", Path(stack.profile["labels"]), "http://127.0.0.1:18091"
+        )
+        inference.ready()
+        rows = []
+        policy = RoadPolicy(0.3)
+        for _ in range(3):
+            started = time.monotonic_ns()
+            road, boxes = inference.infer(rgb)
+            rows.append(
+                {
+                    "inference_ms": (time.monotonic_ns() - started) / 1e6,
+                    "decision": policy.decide(road, boxes).model_dump(),
+                }
+            )
+        report["model_probe"] = rows
+    report["passed"] = True
+except Exception as exc:
+    report["error"] = type(exc).__name__ + ": " + str(exc)
+    raise
+finally:
+    if inference is not None:
+        inference.close()
+    if stack is not None:
+        stack.stop()
+    report["finished_at"] = datetime.now(UTC).isoformat()
+    (args.output / "report.json").write_text(json.dumps(report, indent=2))
+    if app is not None:
+        app.close()
diff --git a/simulation/ai-polygon/qualify_scene_contact.py b/simulation/ai-polygon/qualify_scene_contact.py
new file mode 100644
index 0000000..5c1d8f2
--- /dev/null
+++ b/simulation/ai-polygon/qualify_scene_contact.py
@@ -0,0 +1,101 @@
+"""Worker-only contact replay on a retained episode's exact collision asset.
+
+This isolates physical support from perception/planning. It never supplies a
+route or collision truth to inference and never writes to the original run.
+"""
+
+import argparse
+import hashlib
+import json
+import math
+import sys
+import time
+from datetime import UTC, datetime
+from pathlib import Path
+
+from isaacsim import SimulationApp
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--run", type=Path, required=True)
+parser.add_argument("--output", type=Path, required=True)
+parser.add_argument("--mode", choices=("straight", "replay"), default="straight")
+parser.add_argument("--seconds", type=float, default=40)
+args = parser.parse_args()
+sys.path.insert(0, str(Path(__file__).resolve().parents[2] / "src"))
+
+app = SimulationApp({"headless": True, "hide_ui": True})
+try:
+    import omni.timeline
+    import omni.usd
+    from isaacsim.core.experimental.prims import Articulation
+    from isaacsim.core.simulation_manager import SimulationManager
+    from pxr import UsdGeom
+    from rover_profile import PROFILE, DifferentialDrive, create_rover
+    from terrain import install_terrain
+
+    from k1link.simulation.ai_polygon.mission_policy import inclination
+
+    run = json.loads(args.run.read_text())
+    settings = run["world"]["settings"]
+    stage = omni.usd.get_context().get_stage()
+    UsdGeom.SetStageUpAxis(stage, UsdGeom.Tokens.z)
+    UsdGeom.SetStageMetersPerUnit(stage, 1)
+    height, terrain = install_terrain(stage, run["terrain_manifest"], run["world"])
+    create_rover(
+        stage,
+        [*settings["spawn_xy"], height],
+        settings["heading_degrees"],
+        terrain["initial_ground_normal"],
+    )
+    SimulationManager.setup_simulation(dt=1 / 60, device="cpu")
+    omni.timeline.get_timeline_interface().play()
+    for _ in range(30):
+        app.update()
+    robot = Articulation("/World/Rover")
+    drive = DifferentialDrive(robot)
+    playback = []
+    if args.mode == "replay":
+        playback = [
+            json.loads(x) for x in (args.run.parent / "motion.jsonl").read_text().splitlines()
+        ]
+    trace, cursor, stopped = [], 0, False
+    started = time.monotonic()
+    for step in range(round(args.seconds * 60)):
+        p, q = robot.get_world_poses()
+        p, q = p.numpy()[0].tolist(), q.numpy()[0].tolist()
+        pose = p + q[1:] + q[:1]
+        tilt = inclination(pose)
+        stopped |= tilt >= PROFILE["stop_tilt_degrees"]
+        v, w = settings["max_speed_mps"], 0.0
+        if playback:
+            while (
+                cursor + 1 < len(playback)
+                and playback[cursor + 1]["simulation_time_ns"] / 1e9 <= step / 60
+            ):
+                cursor += 1
+            v, w = playback[cursor]["applied_speed_mps"], playback[cursor]["applied_yaw_rate_rps"]
+        drive.command(0 if stopped else v, 0 if stopped else w)
+        if step % 10 == 0:
+            trace.append(dict(time_s=step / 60, pose=pose, tilt_degrees=tilt, stopped=stopped))
+        SimulationManager.step(steps=1)
+    report = dict(
+        schema_version="missioncore.scene-contact-qualification/v1",
+        utc=datetime.now(UTC).isoformat(),
+        monotonic=time.monotonic(),
+        elapsed_wall_s=time.monotonic() - started,
+        run_sha256=hashlib.sha256(args.run.read_bytes()).hexdigest(),
+        mode=args.mode,
+        wheel_collision=PROFILE["wheel_collision"],
+        seconds=args.seconds,
+        profile=PROFILE,
+        terrain=terrain,
+        trace=trace,
+        displacement_m=math.dist(trace[0]["pose"][:2], trace[-1]["pose"][:2]),
+        max_tilt_degrees=max(x["tilt_degrees"] for x in trace),
+        stopped=stopped,
+    )
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(json.dumps(report, indent=2), encoding="utf-8")
+    print(json.dumps({k: v for k, v in report.items() if k not in ("trace", "terrain", "profile")}))
+finally:
+    app.close()
diff --git a/simulation/ai-polygon/qualify_terrain_coordinates.py b/simulation/ai-polygon/qualify_terrain_coordinates.py
new file mode 100644
index 0000000..c0d2c0a
--- /dev/null
+++ b/simulation/ai-polygon/qualify_terrain_coordinates.py
@@ -0,0 +1,105 @@
+"""Worker-only landmark regression through the actual collision preparer.
+
+The analytic landmark is transformed like the USD scan. This detects the
+SplatTransform PLY convention and glTF basis mistakes before a scene is used.
+"""
+
+import argparse
+import hashlib
+import json
+import struct
+import subprocess
+import time
+from datetime import UTC, datetime
+from pathlib import Path
+
+import numpy as np
+from terrain import load_glb
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--root", type=Path, required=True)
+parser.add_argument("--output", type=Path, required=True)
+args = parser.parse_args()
+args.output.mkdir(parents=True, exist_ok=True)
+names = [
+    "x",
+    "y",
+    "z",
+    "f_dc_0",
+    "f_dc_1",
+    "f_dc_2",
+    "opacity",
+    "scale_0",
+    "scale_1",
+    "scale_2",
+    "rot_0",
+    "rot_1",
+    "rot_2",
+    "rot_3",
+]
+header = (
+    "ply\nformat binary_little_endian 1.0\nelement vertex 1\n"
+    + "".join("property float " + n + "\n" for n in names)
+    + "end_header\n"
+)
+payload = header.encode() + struct.pack("<14f", 1, 2, 3, 0, 0, 0, 9, -2.3, -2.3, -2.3, 1, 0, 0, 0)
+sha = hashlib.sha256(payload).hexdigest()
+# A private fixture root uses the installed tool directory but never overwrites
+# a downloaded world or an existing terrain artifact.
+fixture = args.output / "fixture"
+(fixture / "state/worlds").mkdir(parents=True, exist_ok=True)
+(fixture / "state/worlds" / (sha + ".ply")).write_bytes(payload)
+tools = fixture / "tools"
+if not tools.exists():
+    subprocess.run(
+        ["cmd", "/c", "mklink", "/J", str(tools), str(args.root / "tools")],
+        check=True,
+        capture_output=True,
+    )
+world = {
+    "sha256": sha,
+    "settings": {
+        "spawn_xy": [-1, -3],
+        "ground_z": -2,
+        "rotation_degrees": [-90, 0, 180],
+        "meters_per_unit": 1,
+    },
+}
+world_file = fixture / "world.json"
+world_file.write_text(json.dumps(world), encoding="utf-8")
+command = [
+    "powershell",
+    "-NoProfile",
+    "-ExecutionPolicy",
+    "Bypass",
+    "-File",
+    str(Path(__file__).parent / "navigation/Prepare-Terrain.ps1"),
+    "-Root",
+    str(fixture),
+    "-WorldFile",
+    str(world_file),
+]
+with (args.output / "prepare.log").open("w") as log:
+    subprocess.run(command, stdout=log, stderr=subprocess.STDOUT, check=True, timeout=90)
+manifest_path = next((fixture / "assets/terrain-v1").glob("*/terrain.json"))
+manifest = json.loads(manifest_path.read_text(encoding="utf-8-sig"))
+points, _ = load_glb(manifest["collider"])
+# Floor-fill extends the fixture downward. The sphere's upper surface is
+# unaffected; its horizontal centre and uppermost Z must agree with the visual.
+centre_xy = (points[:, :2].min(axis=0) + points[:, :2].max(axis=0)) / 2
+expected = np.array([-1, -3])
+passed = bool(np.linalg.norm(centre_xy - expected) < 0.06 and -2 < points[:, 2].max() < -1.6)
+report = {
+    "utc": datetime.now(UTC).isoformat(),
+    "monotonic": time.monotonic(),
+    "source_sha256": sha,
+    "generator_sha256": manifest["generator_sha256"],
+    "expected_xy": expected.tolist(),
+    "actual_xy": centre_xy.tolist(),
+    "top_z": float(points[:, 2].max()),
+    "passed": passed,
+}
+(args.output / "report.json").write_text(json.dumps(report, indent=2), encoding="utf-8")
+print(json.dumps(report))
+if not passed:
+    raise SystemExit(1)
diff --git a/simulation/ai-polygon/qualify_terrain_drive.py b/simulation/ai-polygon/qualify_terrain_drive.py
new file mode 100644
index 0000000..4b6dfca
--- /dev/null
+++ b/simulation/ai-polygon/qualify_terrain_drive.py
@@ -0,0 +1,135 @@
+"""Worker-only physical capability measurement on isolated metric lanes.
+
+All lanes share the shipped chassis and tire materials; no AI or Gaussian
+geometry. Results qualify this virtual profile only, never the real vehicles.
+"""
+
+import argparse
+import hashlib
+import json
+import math
+import time
+from datetime import UTC, datetime
+from pathlib import Path
+
+from isaacsim import SimulationApp
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--output", type=Path, required=True)
+args = parser.parse_args()
+app = SimulationApp({"headless": True, "hide_ui": True})
+try:
+    import omni.timeline
+    import omni.usd
+    from isaacsim.core.experimental.prims import Articulation
+    from isaacsim.core.simulation_manager import SimulationManager
+    from pxr import Gf, UsdGeom, UsdPhysics, UsdShade
+    from rover_profile import PROFILE, DifferentialDrive, create_rover
+
+    stage = omni.usd.get_context().get_stage()
+    UsdGeom.SetStageUpAxis(stage, UsdGeom.Tokens.z)
+    UsdGeom.SetStageMetersPerUnit(stage, 1)
+    material = UsdShade.Material.Define(stage, "/World/Materials/Terrain")
+    physics = UsdPhysics.MaterialAPI.Apply(material.GetPrim())
+    physics.CreateStaticFrictionAttr(0.9)
+    physics.CreateDynamicFrictionAttr(0.8)
+    physics.CreateRestitutionAttr(0)
+
+    def surface(prim):
+        UsdPhysics.CollisionAPI.Apply(prim)
+        UsdShade.MaterialBindingAPI.Apply(prim).Bind(
+            material, UsdShade.Tokens.weakerThanDescendants, "physics"
+        )
+
+    def cube(path, center, scale):
+        shape = UsdGeom.Cube.Define(stage, path)
+        shape.CreateSizeAttr(1)
+        shape.AddTranslateOp().Set(Gf.Vec3d(*center))
+        shape.AddScaleOp().Set(Gf.Vec3f(*scale))
+        surface(shape.GetPrim())
+
+    cases = [("flat", 0)] + [("step", h) for h in (0.05, 0.1, 0.12, 0.14, 0.15, 0.18, 0.2, 0.25)]
+    cases += [("slope", a) for a in (10, 15, 20, 25)]
+    for i, (kind, value) in enumerate(cases):
+        y = i * 5.0
+        path = f"/World/Lane{i}"
+        cube(path + "/Ground", (3, y, -0.1), (12, 3, 0.2))
+        if kind == "step":
+            cube(path + "/Step", (4.5, y, value / 2), (6, 3, value))
+        elif kind == "slope":
+            # A continuous supported ramp; x=1.5 is the toe, no hidden step.
+            h = 6 * math.tan(math.radians(value))
+            ramp = UsdGeom.Mesh.Define(stage, path + "/Ramp")
+            ramp.CreatePointsAttr(
+                [(1.5, y - 1.5, 0), (1.5, y + 1.5, 0), (7.5, y - 1.5, h), (7.5, y + 1.5, h)]
+            )
+            ramp.CreateFaceVertexCountsAttr([3, 3])
+            ramp.CreateFaceVertexIndicesAttr([0, 2, 1, 1, 2, 3])
+            ramp.CreateSubdivisionSchemeAttr(UsdGeom.Tokens.none)
+            surface(ramp.GetPrim())
+        create_rover(stage, [0, y, PROFILE["wheel_radius_m"] + 0.04], 0, root_path=path + "/Rover")
+    SimulationManager.setup_simulation(dt=1 / 60, device="cpu")
+    omni.timeline.get_timeline_interface().play()
+    for _ in range(30):
+        app.update()
+    robots = [Articulation(f"/World/Lane{i}/Rover") for i in range(len(cases))]
+    indices = [r.get_dof_indices(PROFILE["wheel_names"]) for r in robots]
+    drives = [DifferentialDrive(r) for r in robots]
+    traces = [[] for _ in cases]
+    starts = [r.get_world_poses()[0].numpy()[0].tolist() for r in robots]
+    started = time.monotonic()
+    for drive in drives:
+        drive.command(0.3)
+    for tick in range(180):
+        SimulationManager.step(steps=10)
+        for i, robot in enumerate(robots):
+            pos, quat = robot.get_world_poses()
+            p, q = pos.numpy()[0], quat.numpy()[0]
+            tilt = math.degrees(
+                math.acos(max(-1, min(1, 1 - 2 * (float(q[1]) ** 2 + float(q[2]) ** 2))))
+            )
+            traces[i].append(dict(time_s=(tick + 1) / 6, xyz=p.tolist(), tilt_degrees=tilt))
+            if p[0] >= 5.5 or tilt > 35:
+                drives[i].command(0)
+    stops = [r.get_world_poses()[0].numpy()[0].tolist() for r in robots]
+    for drive in drives:
+        drive.command(0)
+    SimulationManager.step(steps=120)
+    rows = []
+    for i, (kind, value) in enumerate(cases):
+        end = robots[i].get_world_poses()[0].numpy()[0].tolist()
+        tilt = max(t["tilt_degrees"] for t in traces[i])
+        rows.append(
+            dict(
+                kind=kind,
+                value=value,
+                start=starts[i],
+                end=end,
+                max_tilt_degrees=tilt,
+                braking_drift_m=math.dist(stops[i], end),
+                wheel_velocity_rps=robots[i]
+                .get_dof_velocities(dof_indices=indices[i])
+                .numpy()
+                .tolist(),
+                reached=end[0] >= 5.4,
+                upright=tilt < 35,
+                passed=end[0] >= 5.4 and tilt < 35 and math.dist(stops[i], end) < 0.1,
+                trace=traces[i],
+            )
+        )
+    report = dict(
+        schema_version="missioncore.virtual-rover-capability/v1",
+        utc=datetime.now(UTC).isoformat(),
+        monotonic=time.monotonic(),
+        elapsed_wall_s=time.monotonic() - started,
+        profile=PROFILE,
+        profile_sha256=hashlib.sha256(
+            Path(__file__).with_name("rover_profile.py").read_bytes()
+        ).hexdigest(),
+        cases=rows,
+    )
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(json.dumps(report, indent=2), encoding="utf-8")
+    print(json.dumps({"cases": [{k: v for k, v in row.items() if k != "trace"} for row in rows]}))
+finally:
+    app.close()
diff --git a/simulation/ai-polygon/realtime_ai.py b/simulation/ai-polygon/realtime_ai.py
new file mode 100644
index 0000000..eb4a746
--- /dev/null
+++ b/simulation/ai-polygon/realtime_ai.py
@@ -0,0 +1,154 @@
+"""Worker-local, bounded asynchronous inference. No operator-network dependencies."""
+
+import hashlib
+import json
+import threading
+import time
+import traceback
+from pathlib import Path
+
+
+class LatestInference:
+    def __init__(self, factory, policy, directory: Path, clock=time.monotonic):
+        self.factory, self.policy, self.directory, self.clock = factory, policy, directory, clock
+        self.condition = threading.Condition()
+        self.enabled = self.closed = self.ready = False
+        self.ever_ready = False
+        self.epoch = self.count = self.dropped = 0
+        self.pending = self.result = None
+        self.error = None
+        self.thread = threading.Thread(target=self._run, name="polygon-inference", daemon=True)
+
+    def start(self):
+        self.thread.start()
+
+    def enable(self, enabled):
+        with self.condition:
+            if enabled != self.enabled:
+                self.enabled = enabled
+                self.epoch += 1
+                self.pending = self.result = None
+            self.condition.notify_all()
+
+    def submit(self, rgb, frame_id, captured_at, simulation_time_ns, observation=None):
+        with self.condition:
+            if not self.enabled or self.closed:
+                return
+            if self.pending is not None:
+                self.dropped += 1
+            self.pending = (rgb, frame_id, captured_at, simulation_time_ns, self.epoch, observation)
+            self.condition.notify_all()
+
+    def command(self, now, frame_deadline=0.5, command_deadline=0.5):
+        with self.condition:
+            result = self.result
+            if not self.enabled:
+                return 0.0, 0.0, "paused", result
+            if self.error:
+                return 0.0, 0.0, "inference-error", result
+            if result is None or now - result["captured_at"] > frame_deadline:
+                return 0.0, 0.0, "stale-camera", result
+            if now - result["completed_at"] > command_deadline:
+                return 0.0, 0.0, "stale-command", result
+            decision = result["decision"]
+            return decision["speed_mps"], decision["yaw_rate_rps"], "none", result
+
+    def _run(self):
+        model = None
+        policy_epoch = None
+        self.directory.mkdir(parents=True, exist_ok=True)
+        try:
+            with (self.directory / "decisions.jsonl").open("a", encoding="utf-8") as journal:
+                while True:
+                    with self.condition:
+                        self.condition.wait_for(
+                            lambda: self.closed or (self.enabled and self.pending is not None)
+                        )
+                        if self.closed:
+                            return
+                        rgb, frame_id, captured_at, sim_ns, epoch, observation = self.pending
+                        self.pending = None
+                    try:
+                        if model is None:
+                            model = self.factory()
+                            model.ready()
+                        started = self.clock()
+                        evidence = {}
+                        if self.policy is None:
+                            if policy_epoch is not None and policy_epoch != epoch:
+                                model.reset()
+                            policy_epoch = epoch
+                            decision, boxes, evidence = model.infer_observation(
+                                rgb, observation, frame_id
+                            )
+                        else:
+                            road, boxes = model.infer(rgb)
+                        with self.condition:
+                            if epoch != self.epoch or not self.enabled:
+                                continue
+                            if self.policy is not None and policy_epoch != epoch:
+                                self.policy.reset()
+                                policy_epoch = epoch
+                            if self.policy is not None:
+                                decision = self.policy.decide(road, boxes).model_dump()
+                        completed = self.clock()
+                        record = dict(
+                            frame_id=frame_id,
+                            captured_at=captured_at,
+                            simulation_time_ns=sim_ns,
+                            completed_at=completed,
+                            inference_ms=(completed - started) * 1000,
+                            decision=decision,
+                        )
+                        with self.condition:
+                            self.ready, self.ever_ready, self.error = True, True, None
+                            if epoch != self.epoch or not self.enabled:
+                                continue  # A pause/resume fences every earlier in-flight decision.
+                            self.result = record
+                            self.count += 1
+                        # Archive source RGB on Worker, independently of the video stream.
+                        from PIL import Image
+
+                        image_path = self.directory / f"{frame_id:08d}.jpg"
+                        Image.fromarray(rgb).save(image_path, "JPEG", quality=90)
+                        record = {
+                            **record,
+                            "image": image_path.name,
+                            "image_sha256": hashlib.sha256(image_path.read_bytes()).hexdigest(),
+                            "obstacles": boxes,
+                            **evidence,
+                        }
+                        journal.write(json.dumps(record, allow_nan=False) + "\n")
+                        journal.flush()
+                    except Exception as exc:
+                        with (self.directory / "errors.jsonl").open(
+                            "a", encoding="utf-8"
+                        ) as errors:
+                            errors.write(
+                                json.dumps(
+                                    {
+                                        "frame_id": frame_id,
+                                        "monotonic": self.clock(),
+                                        "error": str(exc),
+                                        "type": type(exc).__name__,
+                                        "traceback": traceback.format_exc(),
+                                    }
+                                )
+                                + "\n"
+                            )
+                        with self.condition:
+                            self.error, self.ready, self.result = type(exc).__name__, False, None
+                        if model is not None:
+                            model.close()
+                            model = None
+                        with self.condition:
+                            self.condition.wait(timeout=0.5)
+        finally:
+            if model is not None:
+                model.close()
+
+    def close(self):
+        with self.condition:
+            self.closed = True
+            self.condition.notify_all()
+        self.thread.join(timeout=12)
diff --git a/simulation/ai-polygon/realtime_worker.py b/simulation/ai-polygon/realtime_worker.py
new file mode 100644
index 0000000..772d5c0
--- /dev/null
+++ b/simulation/ai-polygon/realtime_worker.py
@@ -0,0 +1,377 @@
+"""Persistent host coordinator. Control/telemetry I/O never runs in the simulator.
+
+The Windows task owns the coordinator; a Job Object fences its native children.
+Closing the viewer or the private control tunnel cannot stop the physics clock.
+Runs are bounded locally; reconnect reconciles the same identity, without replay.
+"""
+
+import argparse
+import ctypes
+import hashlib
+import json
+import logging
+import os
+import subprocess
+import threading
+import time
+from pathlib import Path
+from uuid import uuid4
+
+from core_client import CoreClient
+from local_state import StateChannel, read_json, write_json
+from model_stack import ROOT, ModelStack, docker, sha256
+from worker import terminate_episode
+
+
+class NativeJob:
+    """Windows kills all native descendants if the coordinator unexpectedly exits."""
+
+    def __init__(self):
+        from ctypes import wintypes
+
+        self.kernel = ctypes.WinDLL("kernel32", use_last_error=True)
+        self.kernel.CreateJobObjectW.restype = wintypes.HANDLE
+        self.kernel.CreateJobObjectW.argtypes = [ctypes.c_void_p, wintypes.LPCWSTR]
+        self.kernel.SetInformationJobObject.argtypes = [
+            wintypes.HANDLE,
+            ctypes.c_int,
+            ctypes.c_void_p,
+            wintypes.DWORD,
+        ]
+        self.kernel.AssignProcessToJobObject.argtypes = [wintypes.HANDLE, wintypes.HANDLE]
+        self.kernel.CloseHandle.argtypes = [wintypes.HANDLE]
+        self.handle = self.kernel.CreateJobObjectW(None, None)
+        # JOBOBJECT_EXTENDED_LIMIT_INFORMATION: 144 bytes on 64-bit Windows;
+        # BasicLimitInformation.LimitFlags is DWORD at offset 16.
+        limits = ctypes.create_string_buffer(144)
+        ctypes.c_uint32.from_buffer(limits, 16).value = 0x2000  # KILL_ON_JOB_CLOSE
+        if not self.handle or not self.kernel.SetInformationJobObject(self.handle, 9, limits, 144):
+            raise ctypes.WinError(ctypes.get_last_error())
+
+    def assign(self, child):
+        if not self.kernel.AssignProcessToJobObject(self.handle, int(child._handle)):
+            terminate_episode(child)
+            raise ctypes.WinError(ctypes.get_last_error())
+
+    def close(self):
+        if self.handle:
+            self.kernel.CloseHandle(self.handle)
+            self.handle = None
+
+
+def acquire_service_lock(path):
+    import msvcrt
+
+    stream = path.open("a+b")
+    stream.seek(0)
+    if not stream.read(1):
+        stream.write(b"0")
+        stream.flush()
+    stream.seek(0)
+    msvcrt.locking(stream.fileno(), msvcrt.LK_NBLCK, 1)
+    return stream
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--core", default="http://127.0.0.1:18080")
+    parser.add_argument("--token-file", type=Path, required=True)
+    parser.add_argument("--state", type=Path, required=True)
+    parser.add_argument("--isaac", type=Path, required=True)
+    parser.add_argument("--stream-address", required=True)
+    args = parser.parse_args()
+    args.state.mkdir(parents=True, exist_ok=True)
+    lock = acquire_service_lock(args.state / "realtime-service.lock")
+    logging.basicConfig(
+        filename=args.state / "realtime-service.log",
+        level=logging.INFO,
+        format="%(asctime)s %(levelname)s %(message)s",
+    )
+    identity_path = args.state / "realtime-identity.json"
+    identity = read_json(identity_path)
+    release = sha256(ROOT / "realtime_worker.py")
+    active_path = args.state / "active.json"
+    if identity is None or (identity["release"] != release and not active_path.exists()):
+        identity = {"instance_id": uuid4().hex, "release": release}
+        write_json(identity_path, identity)
+    instance = identity["instance_id"]
+    client = CoreClient(args.core, args.token_file, instance)
+    stack = ModelStack()
+    stack.preflight()
+    sources = {
+        "worker": release,
+        "scene": hashlib.sha256(
+            "".join(
+                sha256(ROOT / n)
+                for n in (
+                    "run_realtime.py",
+                    "motion_control.py",
+                    "realtime_ai.py",
+                    "local_state.py",
+                    "terrain.py",
+                    "spawn_clearance.py",
+                    "prepare_terrain.py",
+                    "navigation_client.py",
+                    "navigation/server.py",
+                    "navigation/footprint.py",
+                    "navigation/fastdds.xml",
+                    "navigation/Prepare-Terrain.ps1",
+                )
+            ).encode()
+            + "".join(
+                sha256(ROOT.parents[1] / "src/k1link/simulation/ai_polygon" / n)
+                for n in (
+                    "inference.py",
+                    "policy.py",
+                    "contracts.py",
+                    "composition.py",
+                    "terrain_contract.py",
+                )
+            ).encode()
+            + "".join(
+                sha256(ROOT.parents[1] / "src/k1link" / n)
+                for n in (
+                    "artifacts.py",
+                    "observatory/__init__.py",
+                    "observatory/modular_composition.py",
+                    "simulation/__init__.py",
+                )
+            ).encode()
+        ).hexdigest(),
+        "models": hashlib.sha256(
+            "".join(
+                sha256(ROOT / name)
+                for name in (
+                    "models.worker-006.json",
+                    "model_stack.py",
+                    "compose.models.yaml",
+                    "ddrnet_server.py",
+                    "segformer/server.py",
+                )
+            ).encode()
+        ).hexdigest(),
+        "robot": sha256(ROOT / "rover_profile.py"),
+    }
+    hello = {
+        "worker_id": "worker-006-ai-polygon",
+        "instance_id": instance,
+        "runtime": "isaac-sim-6.1",
+        "execution_modes": ["realtime"],
+        "model_ids": [m["id"] for m in stack.profile["models"]]
+        + [stack.profile["navigation"]["id"]],
+        "profile_sha256": hashlib.sha256(json.dumps(sources, sort_keys=True).encode()).hexdigest(),
+        "runtime_sources": sources,
+        "stream": {
+            "server": args.stream_address,
+            "signaling_port": 49100,
+            "media_port": 47998,
+            "width": 1280,
+            "height": 720,
+            "fps": 30,
+        },
+    }
+    write_json(
+        args.state / "service-status.json",
+        {"pid": os.getpid(), "instance_id": instance, "hello": hello},
+    )
+
+    def connect():
+        while True:
+            try:
+                client.request("/worker/register", hello)
+                return
+            except Exception as exc:
+                logging.warning("Control unavailable: %s", type(exc).__name__)
+                time.sleep(3)
+
+    # A previous process died. Its Job Object has killed native children.
+    # Reconcile only exact persisted container IDs; never infer release from a lease.
+    orphan = read_json(active_path)
+    if orphan:
+        if orphan.get("instance_id") != instance:
+            raise RuntimeError("Reconciliation identity mismatch")
+        for cid in orphan.get("model_container_ids", []):
+            record = docker("container", "inspect", cid, check=False)
+            if record.returncode == 0:
+                data = json.loads(record.stdout)[0]
+                if data["Config"]["Labels"].get("com.nodedc.stack") != "ai-polygon":
+                    raise RuntimeError("Reconciliation resource owner changed")
+                stack.ids.append(cid)
+        stack.stop()
+        connect()
+        client.request(
+            "/worker/runs/" + orphan["run_id"] + "/finish",
+            {
+                "instance_id": instance,
+                "outcome": "failed",
+                "resources_released": True,
+                "message": "Worker перезапустился. Прогон завершён без повторного запуска.",
+            },
+        )
+        active_path.unlink()
+    connect()
+    while True:
+        try:
+            polled = client.request("/worker/poll", {"instance_id": instance})
+        except Exception:
+            connect()
+            continue
+        if polled["action"] == "idle":
+            time.sleep(0.5)
+            continue
+        run = polled["run"]
+        if run.get("clock") != "realtime":
+            raise RuntimeError("A realtime worker cannot execute lockstep work")
+        episode = args.state / run["run_id"]
+        if episode.exists():
+            raise RuntimeError("An existing episode must never be replayed")
+        episode.mkdir()
+        run_file = episode / "run.json"
+        write_json(run_file, run)
+        channel = StateChannel(episode)
+        channel.write("control", run)
+        active = {"run_id": run["run_id"], "instance_id": instance, "model_container_ids": []}
+        write_json(active_path, active)
+        child = job = starter = None
+        cancelled = threading.Event()
+        model_failure = []
+        result = {"outcome": "failed", "message": "Прогон прерван. Журнал сохранён на Worker."}
+        try:
+            cache = args.state / "worlds"
+            cache.mkdir(exist_ok=True)
+            source = cache / (run["world"]["sha256"] + ".ply")
+            client.download(run["world"], source)
+            client.request("/worker/runs/" + run["run_id"] + "/progress", {"phase": "scene"})
+            from prepare_terrain import prepare_terrain
+
+            job = NativeJob()
+
+            def preparation_pulse(run=run):
+                desired = client.request(
+                    "/worker/poll", {"instance_id": instance, "run_id": run["run_id"]}
+                )
+                if desired["run"]["control"] == "stop":
+                    raise InterruptedError("Terrain preparation cancelled")
+
+            terrain_path = prepare_terrain(
+                args.state.parent, episode, run["world"], job, preparation_pulse
+            )
+            run["terrain_manifest"] = str(terrain_path)
+            write_json(run_file, run)
+            with (episode / "isaac.log").open("wb") as output:
+                child = subprocess.Popen(
+                    [
+                        str(args.isaac / "python.bat"),
+                        str(ROOT / "run_realtime.py"),
+                        "--run",
+                        str(run_file),
+                        "--source",
+                        str(source),
+                        "--stream-address",
+                        args.stream_address,
+                    ],
+                    stdout=output,
+                    stderr=subprocess.STDOUT,
+                )
+                job.assign(child)
+
+                def acquired(ids, active=active):
+                    active["model_container_ids"] = ids
+                    write_json(active_path, active)
+
+                def start_models(
+                    cancelled=cancelled, model_failure=model_failure, acquired=acquired, run=run
+                ):
+                    try:
+                        stack.start(
+                            cancelled=cancelled.is_set,
+                            on_acquired=acquired,
+                            selection=run["request"].get("composition"),
+                        )
+                    except InterruptedError:
+                        pass
+                    except Exception as exc:
+                        model_failure.append(type(exc).__name__)
+                        logging.exception("Model startup failed")
+
+                stop_deadline = None
+                deadline = time.monotonic() + 300 + run["request"]["duration_seconds"]
+                last_snapshot = -1
+                registered = True
+                while child.poll() is None:
+                    now = time.monotonic()
+                    if now >= deadline or (stop_deadline and now >= stop_deadline):
+                        terminate_episode(child)
+                        break
+                    if model_failure:
+                        raise RuntimeError("Local model startup failed")
+                    try:
+                        if not registered:
+                            client.request("/worker/register", hello)
+                            registered = True
+                        polled = client.request(
+                            "/worker/poll", {"instance_id": instance, "run_id": run["run_id"]}
+                        )
+                        desired = polled["run"]
+                        channel.write("control", desired)
+                        if desired["control"] == "stop":
+                            cancelled.set()
+                            stop_deadline = stop_deadline or now + 15
+                        if desired["control"] == "play" and starter is None:
+                            starter = threading.Thread(
+                                target=start_models, name="polygon-model-start", daemon=True
+                            )
+                            starter.start()
+                        snapshot = channel.read("snapshot")
+                        if snapshot and snapshot["sequence"] > last_snapshot:
+                            client.request("/worker/runs/" + run["run_id"] + "/snapshot", snapshot)
+                            last_snapshot = snapshot["sequence"]
+                    except Exception as exc:
+                        registered = False
+                        logging.warning(
+                            "Episode control/telemetry disconnected: %s", type(exc).__name__
+                        )
+                    time.sleep(0.25)
+                result = read_json(episode / "result.json", result)
+                if cancelled.is_set():
+                    result = {"outcome": "stopped", "message": "Движение и inference остановлены."}
+        except InterruptedError:
+            result = {"outcome": "stopped", "message": "Подготовка симуляции остановлена."}
+        except Exception:
+            logging.exception("Episode failed")
+        finally:
+            channel.close()
+            cancelled.set()
+            if child is not None and child.poll() is None:
+                terminate_episode(child)
+            if job is not None:
+                job.close()
+            if starter is not None:
+                starter.join(timeout=150)
+                if starter.is_alive():
+                    raise RuntimeError("Model startup has not released ownership")
+            stack.stop()
+            active["resources_released"] = True
+            write_json(active_path, active)
+            write_json(episode / "worker-result.json", result)
+        while True:
+            try:
+                client.request("/worker/register", hello)
+                client.request(
+                    "/worker/runs/" + run["run_id"] + "/finish",
+                    {**result, "instance_id": instance, "resources_released": True},
+                )
+                active_path.unlink()
+                break
+            except Exception:
+                logging.warning("Finish pending reconciliation")
+                time.sleep(3)
+    lock.close()
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except Exception:
+        logging.exception("Simulation coordinator failed")
+        raise
diff --git a/simulation/ai-polygon/register_paired_scene.py b/simulation/ai-polygon/register_paired_scene.py
new file mode 100644
index 0000000..3c96327
--- /dev/null
+++ b/simulation/ai-polygon/register_paired_scene.py
@@ -0,0 +1,105 @@
+"""Admit validated Worker-local Gaussian + collision assets, sending metadata only.
+
+Run with the prepared Isaac Python environment on the Worker. The descriptor
+declares hashes, provenance, size and scene settings; registration does not
+assert that AI has completed a route. No model/profile is changed here.
+"""
+
+import argparse
+import hashlib
+import json
+import os
+import shutil
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[2] / "src"))
+
+from core_client import CoreClient
+from local_state import read_json, write_json
+from terrain import load_glb
+
+from k1link.simulation.ai_polygon.contracts import WorkerWorldCreate
+from k1link.simulation.ai_polygon.terrain_contract import terrain_matches
+from k1link.simulation.ai_polygon.worlds import inspect_gaussian_ply
+
+
+def digest(path):
+    with path.open("rb") as stream:
+        return hashlib.file_digest(stream, "sha256").hexdigest()
+
+
+def preserve(source, target, expected):
+    if target.exists():
+        if digest(target) != expected:
+            raise ValueError("Existing immutable asset identity differs")
+        return
+    temporary = target.with_suffix(target.suffix + ".part")
+    try:
+        os.link(source, temporary)
+    except OSError:
+        shutil.copyfile(source, temporary)
+    if digest(temporary) != expected:
+        raise ValueError("Asset identity changed while staging")
+    temporary.replace(target)
+
+
+def stage(root, source, collider, descriptor):
+    if (root / "state/active.json").exists():
+        raise RuntimeError("Finish the active simulation before importing a scene")
+    request = WorkerWorldCreate.model_validate(descriptor)
+    if (
+        source.stat().st_size != request.byte_length
+        or digest(source) != request.sha256
+        or inspect_gaussian_ply(source) != request.splat_count
+        or digest(collider) != request.collider_sha256
+    ):
+        raise ValueError("Paired asset does not match its declared identity")
+    load_glb(collider)  # Reject unsupported transforms, external buffers and non-finite data.
+    cache = root / "state/worlds"
+    cache.mkdir(parents=True, exist_ok=True)
+    preserve(source, cache / (request.sha256 + ".ply"), request.sha256)
+    directory = root / "assets/prepared-worlds" / request.sha256
+    directory.mkdir(parents=True, exist_ok=True)
+    destination = directory / "collision.glb"
+    preserve(collider, destination, request.collider_sha256)
+    manifest = {
+        "schema_version": "missioncore.ai-polygon-terrain/v1",
+        "source_sha256": request.sha256,
+        "settings": request.settings.model_dump(mode="json"),
+        "generator": "paired-source",
+        "collider": str(destination),
+        "collider_sha256": request.collider_sha256,
+        "source_url": str(request.source_url) if request.source_url else None,
+        "qualification": "paired-source-requires-contact-and-route-validation",
+    }
+    path = directory / "terrain.json"
+    if path.exists():
+        existing = read_json(path)
+        if existing.get("collider_sha256") != request.collider_sha256 or not terrain_matches(
+            existing, request.model_dump(mode="json")
+        ):
+            raise ValueError("Existing paired-scene calibration differs")
+    else:
+        write_json(path, manifest)
+    return request
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--root", type=Path, required=True)
+    parser.add_argument("--source", type=Path, required=True)
+    parser.add_argument("--collider", type=Path, required=True)
+    parser.add_argument("--descriptor", type=Path, required=True)
+    parser.add_argument("--core", default="http://127.0.0.1:18080")
+    args = parser.parse_args()
+    request = stage(args.root, args.source, args.collider, read_json(args.descriptor))
+    identity = read_json(args.root / "state/realtime-identity.json")
+    client = CoreClient(args.core, args.root / "private/worker.token", identity["instance_id"])
+    result = client.request("/worker/worlds", request.model_dump(mode="json"))
+    write_json(args.root / "assets/prepared-worlds" / request.sha256 / "registration.json", result)
+    print(json.dumps(result, ensure_ascii=True))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/simulation/ai-polygon/route_author/Dockerfile b/simulation/ai-polygon/route_author/Dockerfile
new file mode 100644
index 0000000..7aad2a9
--- /dev/null
+++ b/simulation/ai-polygon/route_author/Dockerfile
@@ -0,0 +1,10 @@
+FROM ros:humble-ros-base-jammy
+LABEL com.nodedc.product="mission-core" \
+      com.nodedc.stack="ai-polygon-route-author" \
+      com.nodedc.role="offline-route-author" \
+      com.nodedc.managed-by="ai-polygon-worker"
+RUN apt-get update && apt-get install -y --no-install-recommends g++ librecast-dev \
+    && rm -rf /var/lib/apt/lists/*
+COPY habitat_route.cpp /src/habitat_route.cpp
+RUN g++ -O2 -std=c++17 -I/usr/include/recastnavigation /src/habitat_route.cpp -lDetour -o /usr/local/bin/habitat-route
+ENTRYPOINT ["habitat-route"]
diff --git a/simulation/ai-polygon/route_author/README.md b/simulation/ai-polygon/route_author/README.md
new file mode 100644
index 0000000..996f38e
--- /dev/null
+++ b/simulation/ai-polygon/route_author/README.md
@@ -0,0 +1,25 @@
+# Offline Habitat route authoring
+
+This tool uses Ubuntu's packaged Recast/Detour library to extract a complete
+route from a supplied Habitat navmesh. It runs on Worker only. It outputs task
+waypoints, never runtime collision knowledge for the AI composition.
+
+Build the `Dockerfile` into `ndc-ai-polygon-route-author:habitat-v1` using the
+isolated anonymous Docker configuration described in the Worker runbook. Mount
+only the source asset directory read-only, disable container networking, and run:
+
+```
+habitat-route /data/scene.navmesh startX startY startZ goalX goalY goalZ
+```
+
+Coordinates remain Habitat Y-up in the output. Mission Core conversion is
+`(x,y,z) -> (x,-z,y)`. The source navmesh's agent radius and step capability may
+be incompatible with the 1 x 1 m rover. Extracted points require full-footprint
+support/collision checks and a real closed-loop run before acceptance. Partial
+paths, incompatible binary layouts, and disconnected endpoints are rejected.
+
+Sources: Habitat-Sim `src/esp/nav/PathFinder.cpp` and `PathFinder.h` define the
+version 1/2 binary container; Recast Navigation's Detour performs the actual
+path search. The loader supports 32-bit Detour references and the v2 56-byte
+settings layout. The image build and source/output hashes belong in the
+private qualification receipt. No asset is modified by route extraction.
diff --git a/simulation/ai-polygon/route_author/habitat_route.cpp b/simulation/ai-polygon/route_author/habitat_route.cpp
new file mode 100644
index 0000000..5458376
--- /dev/null
+++ b/simulation/ai-polygon/route_author/habitat_route.cpp
@@ -0,0 +1,49 @@
+// Offline mission authoring using the supplied Habitat/Detour navmesh.
+// Binary layout: facebookresearch/habitat-sim src/esp/nav/PathFinder.cpp.
+// No navmesh or privileged scene geometry is supplied to runtime inference.
+#include <DetourNavMesh.h>
+#include <DetourNavMeshQuery.h>
+#include <DetourAlloc.h>
+#include <fstream>
+#include <iostream>
+#include <iomanip>
+#include <stdexcept>
+#include <cmath>
+struct Header { int magic, version, tiles; dtNavMeshParams params; };
+struct Tile { dtTileRef ref; int size; };
+void require(bool ok, const char* message) { if (!ok) throw std::runtime_error(message); }
+int main(int argc, char** argv) { try {
+  require(argc == 8, "navmesh start-x start-y start-z goal-x goal-y goal-z required");
+  std::ifstream file(argv[1], std::ios::binary); Header h{};
+  require(bool(file.read(reinterpret_cast<char*>(&h), sizeof(h))), "header read failed");
+  require(h.magic == (('M'<<24)|('S'<<16)|('E'<<8)|'T') && (h.version==1 || h.version==2), "unsupported navmesh");
+  require(h.tiles > 0 && h.tiles < 100000, "invalid tile count");
+  // v2 stores thirteen float settings and four boolean flags (56 bytes).
+  if (h.version == 2) file.seekg(56, std::ios::cur);
+  dtNavMesh* mesh = dtAllocNavMesh(); require(mesh, "allocation failed");
+  require(dtStatusSucceed(mesh->init(&h.params)), "init failed");
+  for (int i=0;i<h.tiles;i++) {
+    Tile t{}; require(bool(file.read(reinterpret_cast<char*>(&t),sizeof(t))), "tile header failed");
+    require(t.ref && t.size>0 && t.size<100000000, "invalid tile");
+    auto* data=static_cast<unsigned char*>(dtAlloc(t.size,DT_ALLOC_PERM));
+    require(data && bool(file.read(reinterpret_cast<char*>(data),t.size)), "tile read failed");
+    require(dtStatusSucceed(mesh->addTile(data,t.size,DT_TILE_FREE_DATA,t.ref,nullptr)), "tile version mismatch");
+  }
+  require(file.peek()==std::char_traits<char>::eof(), "unexpected trailing bytes");
+  dtNavMeshQuery query; require(dtStatusSucceed(query.init(mesh,65535)), "query init failed");
+  float start[3],goal[3],nearStart[3],nearGoal[3],extent[3]={2,4,2};
+  for(int i=0;i<3;i++){start[i]=std::stof(argv[i+2]);goal[i]=std::stof(argv[i+5]);}
+  dtQueryFilter filter; dtPolyRef first=0,last=0;
+  require(dtStatusSucceed(query.findNearestPoly(start,extent,&filter,&first,nearStart)) && first,"start outside navmesh");
+  require(dtStatusSucceed(query.findNearestPoly(goal,extent,&filter,&last,nearGoal)) && last,"goal outside navmesh");
+  dtPolyRef path[4096];int count=0;
+  auto status=query.findPath(first,last,nearStart,nearGoal,&filter,path,&count,4096);
+  require(dtStatusSucceed(status) && !dtStatusDetail(status,DT_BUFFER_TOO_SMALL) && count && path[count-1]==last,"no complete route");
+  float points[4096*3];int n=0;
+  status=query.findStraightPath(nearStart,nearGoal,path,count,points,nullptr,nullptr,&n,4096,DT_STRAIGHTPATH_ALL_CROSSINGS);
+  require(dtStatusSucceed(status) && !dtStatusDetail(status,DT_BUFFER_TOO_SMALL) && n>=2,"route extraction failed");
+  double length=0;for(int i=1;i<n;i++){double d=0;for(int j=0;j<3;j++)d+=std::pow(points[i*3+j]-points[(i-1)*3+j],2);length+=std::sqrt(d);}
+  std::cout<<std::setprecision(10)<<"{\"length_m\":"<<length<<",\"points_habitat\":[";
+  for(int i=0;i<n;i++){if(i)std::cout<<',';std::cout<<'['<<points[i*3]<<','<<points[i*3+1]<<','<<points[i*3+2]<<']';}
+  std::cout<<"]}"<<std::endl;dtFreeNavMesh(mesh);return 0;
+} catch(const std::exception& e){std::cerr<<e.what()<<std::endl;return 1;} }
diff --git a/simulation/ai-polygon/rover_profile.py b/simulation/ai-polygon/rover_profile.py
new file mode 100644
index 0000000..f63753e
--- /dev/null
+++ b/simulation/ai-polygon/rover_profile.py
@@ -0,0 +1,145 @@
+"""Metric 1 x 1 m, four-wheel differential laboratory chassis.
+
+Geometry is explicit. Mass/friction/drive gains are provisional simulation
+parameters, NOT a calibrated braking or terrain model of the owner's vehicles.
+"""
+
+PROFILE = {
+    "schema_version": "missioncore.virtual-rover/v1",
+    "length_m": 1.0,
+    "width_m": 1.0,
+    "wheel_radius_m": 0.25,
+    "wheel_collision": "physx-convex-cylinder",
+    "wheelbase_m": 0.5,
+    "body_contact_height_m": 0.37,
+    "track_width_m": 0.9,
+    "camera_forward_m": 0.38,  # Lens ahead of the mast's front face (x=0.33).
+    "camera_pitch_degrees": -10.0,
+    "rear_range_forward_m": -0.38,
+    "range_sensor_profile": "front-360-rear-160-first-hit-v2",
+    "camera_focal_length_mm": 14.0,  # Wide pinhole view retains ground on slopes.
+    "camera_horizontal_aperture_mm": 36.0,
+    "mass_kg": 24.0,
+    "calibration": "geometry-only",
+    "max_step_m": 0.10,
+    "max_slope_degrees": 25.0,
+    "stop_tilt_degrees": 30.0,
+    "drive_damping": 1000.0,
+    "brake_stiffness": 2000.0,
+    "brake_damping": 100.0,
+    "solver_position_iterations": 32,
+    "solver_velocity_iterations": 8,
+    "contact_offset_m": 0.005,
+    "wheel_names": ["front_left", "rear_left", "front_right", "rear_right"],
+}
+
+
+class DifferentialDrive:
+    """Velocity drive plus physical wheel-position hold when braking.
+
+    Holding uses joint drive forces, never teleportation or a frozen chassis.
+    The same actuator is used by capability qualification and the live scene.
+    """
+
+    def __init__(self, articulation):
+        self.robot = articulation
+        self.indices = articulation.get_dof_indices(PROFILE["wheel_names"])
+        self.holding = None
+
+    def command(self, velocity, yaw_rate=0.0):
+        hold = abs(velocity) + abs(yaw_rate) < 1e-5
+        if hold != self.holding:
+            if hold:
+                position = self.robot.get_dof_positions(dof_indices=self.indices)
+                self.robot.set_dof_position_targets(position, dof_indices=self.indices)
+            self.robot.set_dof_gains(
+                stiffnesses=PROFILE["brake_stiffness"] if hold else 0.0,
+                dampings=PROFILE["brake_damping"] if hold else PROFILE["drive_damping"],
+                dof_indices=self.indices,
+            )
+            self.holding = hold
+        half_track, radius = PROFILE["track_width_m"] / 2, PROFILE["wheel_radius_m"]
+        left = (velocity - yaw_rate * half_track) / radius
+        right = (velocity + yaw_rate * half_track) / radius
+        self.robot.set_dof_velocity_targets([[left, left, right, right]], dof_indices=self.indices)
+
+
+def create_rover(stage, spawn, heading, ground_normal=(0, 0, 1), root_path="/World/Rover"):
+    import carb.settings
+    from omni.physx.bindings._physx import SETTING_COLLISION_APPROXIMATE_CYLINDERS
+    from pxr import Gf, PhysxSchema, UsdGeom, UsdPhysics, UsdShade
+
+    # Official PhysX cylinder approximation avoids the pathological contact
+    # cost measured on the Forest reconstruction. This changes collision
+    # representation, not wheel radius, chassis dimensions or terrain limits.
+    carb.settings.get_settings().set_bool(SETTING_COLLISION_APPROXIMATE_CYLINDERS, True)
+
+    tire = UsdShade.Material.Define(stage, "/World/Materials/RoverTire")
+    tire_physics = UsdPhysics.MaterialAPI.Apply(tire.GetPrim())
+    tire_physics.CreateStaticFrictionAttr(1.0)
+    tire_physics.CreateDynamicFrictionAttr(0.9)
+    tire_physics.CreateRestitutionAttr(0)
+
+    root = UsdGeom.Xform.Define(stage, root_path)
+    root.AddTranslateOp().Set(Gf.Vec3d(*spawn))
+    tilt = Gf.Rotation(Gf.Vec3d(0, 0, 1), Gf.Vec3d(*ground_normal)).GetQuat()
+    yaw = Gf.Rotation(Gf.Vec3d(0, 0, 1), heading).GetQuat()
+    root.AddOrientOp().Set(Gf.Quatf(tilt * yaw))
+    UsdPhysics.ArticulationRootAPI.Apply(root.GetPrim())
+    articulation = PhysxSchema.PhysxArticulationAPI.Apply(root.GetPrim())
+    articulation.CreateEnabledSelfCollisionsAttr(False)
+    articulation.CreateSolverPositionIterationCountAttr(PROFILE["solver_position_iterations"])
+    articulation.CreateSolverVelocityIterationCountAttr(PROFILE["solver_velocity_iterations"])
+
+    def contact(prim):
+        UsdPhysics.CollisionAPI.Apply(prim)
+        collision = PhysxSchema.PhysxCollisionAPI.Apply(prim)
+        collision.CreateContactOffsetAttr(PROFILE["contact_offset_m"])
+        collision.CreateRestOffsetAttr(0.0)
+
+    body = UsdGeom.Xform.Define(stage, root_path + "/chassis")
+    body.AddTranslateOp().Set(Gf.Vec3d(0, 0, 0.12))
+    UsdPhysics.RigidBodyAPI.Apply(body.GetPrim())
+    PhysxSchema.PhysxRigidBodyAPI.Apply(body.GetPrim()).CreateMaxDepenetrationVelocityAttr(1.0)
+    UsdPhysics.MassAPI.Apply(body.GetPrim()).CreateMassAttr(20)
+    hull = UsdGeom.Cube.Define(stage, root_path + "/chassis/hull")
+    hull.CreateSizeAttr(1)
+    hull.AddScaleOp().Set(Gf.Vec3f(0.8, 0.78, 0.24))
+    hull.CreateDisplayColorAttr([Gf.Vec3f(0.65, 0.67, 0.61)])
+    contact(hull.GetPrim())
+    # Mast is visual geometry under the rigid chassis (not a separate body).
+    sensor = UsdGeom.Cube.Define(stage, root_path + "/chassis/sensor")
+    sensor.CreateSizeAttr(1)
+    sensor.AddTranslateOp().Set(Gf.Vec3d(0.25, 0, 0.3))
+    sensor.AddScaleOp().Set(Gf.Vec3f(0.16, 0.16, 0.6))
+    sensor.CreateDisplayColorAttr([Gf.Vec3f(0.12, 0.13, 0.13)])
+    for name in PROFILE["wheel_names"]:
+        x = PROFILE["wheelbase_m"] / 2 * (1 if name.startswith("front") else -1)
+        y = 0.45 if name.endswith("left") else -0.45
+        path = root_path + "/" + name
+        wheel = UsdGeom.Cylinder.Define(stage, path)
+        wheel.CreateAxisAttr("Y")
+        wheel.CreateRadiusAttr(PROFILE["wheel_radius_m"])
+        wheel.CreateHeightAttr(0.1)
+        wheel.AddTranslateOp().Set(Gf.Vec3d(x, y, 0))
+        wheel.CreateDisplayColorAttr([Gf.Vec3f(0.08, 0.08, 0.08)])
+        UsdPhysics.RigidBodyAPI.Apply(wheel.GetPrim())
+        PhysxSchema.PhysxRigidBodyAPI.Apply(wheel.GetPrim()).CreateMaxDepenetrationVelocityAttr(1.0)
+        contact(wheel.GetPrim())
+        UsdShade.MaterialBindingAPI.Apply(wheel.GetPrim()).Bind(
+            tire, UsdShade.Tokens.weakerThanDescendants, "physics"
+        )
+        UsdPhysics.MassAPI.Apply(wheel.GetPrim()).CreateMassAttr(1)
+        joint = UsdPhysics.RevoluteJoint.Define(stage, root_path + "/joints/" + name)
+        joint.CreateBody0Rel().SetTargets([body.GetPath()])
+        joint.CreateBody1Rel().SetTargets([wheel.GetPath()])
+        joint.CreateLocalPos0Attr(Gf.Vec3f(x, y, -0.12))
+        joint.CreateLocalPos1Attr(Gf.Vec3f(0))
+        joint.CreateAxisAttr("Y")
+        drive = UsdPhysics.DriveAPI.Apply(joint.GetPrim(), "angular")
+        drive.CreateTypeAttr("force")
+        drive.CreateStiffnessAttr(0)
+        drive.CreateDampingAttr(PROFILE["drive_damping"])
+        drive.CreateMaxForceAttr(50)
+        drive.CreateTargetVelocityAttr(0)
+    return PROFILE
diff --git a/simulation/ai-polygon/run_realtime.py b/simulation/ai-polygon/run_realtime.py
new file mode 100644
index 0000000..47b4a6f
--- /dev/null
+++ b/simulation/ai-polygon/run_realtime.py
@@ -0,0 +1,566 @@
+"""Native Worker-local physics/render/camera loop, independent of Core and AI latency."""
+
+import argparse
+import json
+import math
+import shutil
+import sys
+import time
+import traceback
+from datetime import UTC, datetime
+from pathlib import Path
+
+from local_state import StateChannel, read_json, write_json
+from motion_control import CONTROL_PROFILE, DriveEnvelope
+from realtime_ai import LatestInference
+from rover_profile import PROFILE, DifferentialDrive, create_rover
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--run", type=Path, required=True)
+parser.add_argument("--source", type=Path, required=True)
+parser.add_argument("--stream-address", required=True)
+args, _ = parser.parse_known_args()
+root = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(root / "src"))
+run = read_json(args.run)
+settings = run["world"]["settings"]
+directory = args.run.parent
+channel = StateChannel(directory)
+result = {"outcome": "failed", "message": "Симуляция не завершена."}
+app = ai = None
+physics_callbacks = []
+
+
+def startup_phase(phase):
+    record = {"phase": phase, "utc": datetime.now(UTC).isoformat(), "monotonic": time.monotonic()}
+    write_json(directory / "startup.json", record)
+    print(json.dumps({"startup": record}), flush=True)
+
+
+try:
+    startup_phase("native-runtime")
+    from isaacsim import SimulationApp
+
+    app = SimulationApp(
+        {
+            "headless": True,
+            "hide_ui": True,
+            "multi_gpu": False,
+            "width": 1280,
+            "height": 720,
+            "window_width": 1280,
+            "window_height": 720,
+            "renderer": "RaytracedLighting",
+            "display_options": 0,
+            "extra_args": [
+                "--/app/player/useFixedTimeStepping=false",
+                "--/app/runLoops/main/manualModeEnabled=false",
+                "--/exts/isaacsim.core.throttling/enable_manualmode=false",
+            ],
+        }
+    )
+    import carb.settings
+    import numpy as np
+    import omni.kit.app
+    import omni.replicator.core as rep
+    import omni.timeline
+    import omni.usd
+    from isaacsim.core.experimental.prims import Articulation
+    from isaacsim.core.experimental.utils.app import enable_extension
+    from isaacsim.core.simulation_manager import SimulationEvent, SimulationManager
+    from navigation.footprint import FRAME_DEADLINE_SECONDS
+    from navigation_client import ComposedInference
+    from omni.kit.loop import _loop as omni_loop
+    from omni.kit.viewport.utility import get_active_viewport
+    from pxr import Gf, UsdGeom, UsdLux
+    from terrain import RangeSensor, install_terrain
+
+    from k1link.simulation.ai_polygon.mission_policy import inclination
+
+    startup_phase("stream-and-gaussian-runtime")
+    config = carb.settings.get_settings()
+    prefix = "/exts/omni.kit.livestream.app/primaryStream/"
+    for key, value in {
+        "publicIp": args.stream_address,
+        "signalPort": 49100,
+        "streamPort": 47998,
+        "targetFps": 30,
+        "enableEventTracing": False,
+    }.items():
+        config.set(prefix + key, value)
+    enable_extension("omni.kit.livestream.app")
+    enable_extension("omni.kit.converter.gsplat")
+    from usd_convert_gsplat import read_ply, write_gaussian_splat_usd
+
+    converted = args.source.with_suffix(".usd")
+    if not converted.exists():
+        temporary = converted.with_suffix(".part.usd")
+        write_gaussian_splat_usd(
+            read_ply(str(args.source)),
+            str(temporary),
+            source_file=str(args.source),
+            prim_name="Gaussians",
+            up_axis="Z",
+        )
+        temporary.replace(converted)
+    stage = omni.usd.get_context().get_stage()
+    UsdGeom.SetStageUpAxis(stage, UsdGeom.Tokens.z)
+    UsdGeom.SetStageMetersPerUnit(stage, 1)
+    visual = UsdGeom.Xform.Define(stage, "/World/Scan")
+    visual.AddRotateXYZOp().Set(Gf.Vec3f(*settings["rotation_degrees"]))
+    visual.AddScaleOp().Set(Gf.Vec3f(settings["meters_per_unit"]))
+    stage.DefinePrim("/World/Scan/Gaussians").GetReferences().AddReference(str(converted))
+    startup_phase("terrain")
+    wheel_center_z, terrain_record = install_terrain(stage, run["terrain_manifest"], run["world"])
+    write_json(directory / "terrain.json", terrain_record)
+    UsdLux.DomeLight.Define(stage, "/World/Light").CreateIntensityAttr(500)
+    create_rover(
+        stage,
+        [*settings["spawn_xy"], wheel_center_z],
+        settings["heading_degrees"],
+        terrain_record["initial_ground_normal"],
+    )
+    write_json(directory / "robot-profile.json", PROFILE)
+    write_json(directory / "motion-control.json", CONTROL_PROFILE)
+    startup_phase("cameras")
+
+    def camera(path, aspect, focal_length=24):
+        value = UsdGeom.Camera.Define(stage, path)
+        value.CreateFocalLengthAttr(focal_length)
+        value.CreateHorizontalApertureAttr(36)
+        value.CreateVerticalApertureAttr(36 / aspect)
+        value.CreateClippingRangeAttr(Gf.Vec2f(0.05, 1000))
+        return value, value.AddTranslateOp(), value.AddOrientOp()
+
+    sensor, sensor_pos, sensor_rot = camera(
+        "/World/Sensor", 4 / 3, PROFILE["camera_focal_length_mm"]
+    )
+    focal_pixels = (
+        800 * PROFILE["camera_focal_length_mm"] / PROFILE["camera_horizontal_aperture_mm"]
+    )
+    observer, observer_pos, observer_rot = camera("/World/Observer", 16 / 9)
+    product = rep.create.render_product(str(sensor.GetPath()), (800, 600))
+    annotator = rep.AnnotatorRegistry.get_annotator("rgb")
+    annotator.attach(product)
+    config.set("/omni/replicator/captureOnPlay", True)
+    viewport = get_active_viewport()
+    if viewport is None:
+        raise RuntimeError("Streaming viewport is unavailable")
+    viewport.camera_path = str(observer.GetPath())
+    viewport.set_texture_resolution((1280, 720))
+    # Initialise the final camera before the first rendered frames. Starting
+    # WebRTC with an uninitialised observer and an already paused timeline can
+    # leave the native stream without its first image/offer.
+    initial = Gf.Vec3d(*settings["spawn_xy"], wheel_center_z + 0.12)
+    heading = math.radians(settings["heading_degrees"])
+    initial_eye = initial + Gf.Vec3d(-math.cos(heading) * 2.5, -math.sin(heading) * 2.5, 1.5)
+    observer_pos.Set(initial_eye)
+    observer_rot.Set(
+        Gf.Quatf(
+            Gf.Matrix4d()
+            .SetLookAt(initial_eye, initial + Gf.Vec3d(0.1, 0, 0.2), Gf.Vec3d(0, 0, 1))
+            .GetInverse()
+            .ExtractRotationQuat()
+        )
+    )
+    startup_phase("physics-setup")
+    SimulationManager.setup_simulation(dt=1 / 60, device="cpu")
+    timeline = omni.timeline.get_timeline_interface()
+    # Render rate and physics rate are separate. Physics steps follow elapsed time.
+    config.set("/app/player/useFixedTimeStepping", False)
+    config.set("/exts/isaacsim.core.throttling/enable_manualmode", False)
+    omni_loop.acquire_loop_interface().set_manual_mode(False)
+    timeline.set_play_every_frame(False)
+    config.set("/persistent/simulation/minFrameRate", 1)
+    config.set("/app/runLoops/main/rateLimitEnabled", True)
+    config.set("/app/runLoops/main/rateLimitFrequency", 30)
+    startup_phase("initial-render-and-contact")
+    timeline.play()
+    for _ in range(30):
+        app.update()
+    articulation = Articulation("/World/Rover")
+    drive = DifferentialDrive(articulation)
+    wheel_indices = articulation.get_dof_indices(PROFILE["wheel_names"])
+
+    def body_transform():
+        positions, orientations = articulation.get_world_poses()
+        position, orientation = positions.numpy()[0], orientations.numpy()[0]
+        transform = Gf.Matrix4d(1)
+        transform.SetRotate(Gf.Quatd(float(orientation[0]), Gf.Vec3d(*map(float, orientation[1:]))))
+        transform.SetTranslateOnly(Gf.Vec3d(*map(float, position)))
+        return transform
+
+    settled = body_transform().ExtractTranslation()
+    write_json(
+        directory / "initial-contact.json",
+        {
+            "pose": list(settled),
+            "expected_xy": settings["spawn_xy"],
+            "spawn_contact_z": terrain_record["initial_contact_z"],
+        },
+    )
+    if math.hypot(settled[0] - settings["spawn_xy"][0], settled[1] - settings["spawn_xy"][1]) > 0.2:
+        raise RuntimeError("Rover start is unstable on this reconstructed surface")
+    timeline.pause()
+    app.update()
+    startup_phase("ready")
+    baseline = SimulationManager.get_num_physics_steps()
+    range_sensor = RangeSensor()
+    from k1link.simulation.ai_polygon.mission_policy import WaypointMission
+
+    # The episode owns the mission. Reconnecting a failed model client must not
+    # rewind its route cursor, recovery budget or latched terminal condition.
+    mission = WaypointMission(settings.get("route_xy", []))
+    ai = LatestInference(
+        lambda: ComposedInference(
+            root / "simulation/ai-polygon", run, directory / "camera", mission=mission
+        ),
+        None,
+        directory / "camera",
+    )
+    ai.start()
+    control = {"control": "pause", "control_sequence": 0, "camera": "follow"}
+    start = last_report = last_sensor = last_control = time.monotonic()
+    render_frames = sensor_frames = sequence = 0
+    prior_metrics = (start, 0, 0, 0, baseline)
+    speed = 0.0
+    had_ai = False
+    unstable = False
+    terminal_at = None
+    envelope = DriveEnvelope(track_width=PROFILE["track_width_m"])
+    state = {"velocity": 0.0, "yaw_rate": 0.0, "stop_reason": "paused", "decision": None}
+    physics_rows = []
+
+    def control_step(dt, _context):
+        velocity, yaw_rate, reason, decision = ai.command(
+            time.monotonic(), frame_deadline=FRAME_DEADLINE_SECONDS
+        )
+        terminal = unstable or (
+            decision and decision["decision"]["reason"] in {"unstable", "stuck", "goal-reached"}
+        )
+        velocity, yaw_rate = envelope.step(
+            velocity, yaw_rate, dt, stop=bool(terminal) or reason != "none"
+        )
+        drive.command(velocity, yaw_rate)
+        state.update(
+            velocity=velocity,
+            yaw_rate=yaw_rate,
+            stop_reason="unstable" if unstable else reason,
+            decision=decision,
+        )
+
+    def update_cameras():
+        transform = body_transform()
+        position = transform.ExtractTranslation()
+        forward = transform.TransformDir(Gf.Vec3d(1, 0, 0)).GetNormalized()
+        yaw = math.atan2(forward[1], forward[0])
+        # Both sensor position and attitude follow the articulated body.
+        # Mounting height is measured from nominal wheel contact.
+        eye = transform.Transform(
+            Gf.Vec3d(
+                PROFILE["camera_forward_m"],
+                0,
+                settings["camera_height_m"] - PROFILE["body_contact_height_m"],
+            )
+        )
+        pitch = math.radians(PROFILE["camera_pitch_degrees"])
+        camera_axes = [
+            transform.TransformDir(Gf.Vec3d(*axis)).GetNormalized()
+            for axis in (
+                (math.cos(pitch), 0, math.sin(pitch)),
+                (0, 1, 0),
+                (-math.sin(pitch), 0, math.cos(pitch)),
+            )
+        ]
+
+        def look_at(pos, rotation, origin, target, up=None):
+            if up is None:
+                up = Gf.Vec3d(0, 0, 1)
+            pos.Set(origin)
+            rotation.Set(
+                Gf.Quatf(
+                    Gf.Matrix4d().SetLookAt(origin, target, up).GetInverse().ExtractRotationQuat()
+                )
+            )
+
+        look_at(
+            sensor_pos,
+            sensor_rot,
+            eye,
+            eye + camera_axes[0],
+            camera_axes[2],
+        )
+        mode = control.get("camera", "follow")
+        if mode == "camera":
+            viewport.camera_path = str(sensor.GetPath())
+        else:
+            viewport.camera_path = str(observer.GetPath())
+            offset = (
+                Gf.Vec3d(-math.cos(yaw) * 2.5, -math.sin(yaw) * 2.5, 1.5)
+                if mode == "follow"
+                else Gf.Vec3d(-3, -3, 5)
+            )
+            look_at(observer_pos, observer_rot, position + offset, position + Gf.Vec3d(0.1, 0, 0.2))
+        linear, angular = articulation.get_velocities()
+        q = transform.ExtractRotationQuat()
+        state.update(
+            transform=transform,
+            position=position,
+            yaw=yaw,
+            eye=eye,
+            camera_axes=camera_axes,
+            mode=mode,
+            pose=list(position) + list(q.GetImaginary()) + [q.GetReal()],
+            speed=float(np.linalg.norm(linear.numpy()[0, :2])),
+            body_velocity=linear.numpy()[0].tolist(),
+            body_angular_velocity=angular.numpy()[0].tolist(),
+        )
+
+    def physics_step(dt, _context):
+        global unstable
+        update_cameras()
+        unstable |= inclination(state["pose"]) >= PROFILE["stop_tilt_degrees"]
+        physics_rows.append(
+            {
+                "monotonic": time.monotonic(),
+                "dt": dt,
+                "pose": state["pose"],
+                "velocity": state["body_velocity"],
+                "angular_velocity": state["body_angular_velocity"],
+                "command": [state["velocity"], state["yaw_rate"]],
+                "stop_reason": state["stop_reason"],
+            }
+        )
+
+    callback_errors = []
+
+    def guarded(callback):
+        def invoke(dt, context):
+            try:
+                callback(dt, context)
+            except Exception:
+                # Native event dispatch logs Python exceptions and continues.
+                # A failed control/camera callback must instead end this run.
+                callback_errors.append(traceback.format_exc())
+                ai.enable(False)
+                drive.command(0.0, 0.0)
+
+        return invoke
+
+    update_cameras()
+    physics_callbacks.extend(
+        [
+            SimulationManager.register_callback(
+                guarded(control_step), SimulationEvent.PHYSICS_PRE_STEP
+            ),
+            SimulationManager.register_callback(
+                guarded(physics_step), SimulationEvent.PHYSICS_POST_STEP
+            ),
+        ]
+    )
+    with (
+        (directory / "motion.jsonl").open("a", encoding="utf-8") as motion,
+        (directory / "physics-motion.jsonl").open("a", encoding="utf-8") as physics_motion,
+    ):
+        while app.is_running():
+            now = time.monotonic()
+            if now - last_control >= 0.1:
+                control = channel.read("control", control)
+                last_control = now
+            if control["control"] == "stop":
+                result = {"outcome": "stopped", "message": "Симуляция и inference остановлены."}
+                break
+            if now - start >= run["request"]["duration_seconds"]:
+                result = {"outcome": "completed", "message": "Время прогона завершено."}
+                break
+            playing = control["control"] == "play"
+            ai.enable(playing)
+            had_ai |= playing
+            if playing and not timeline.is_playing():
+                timeline.play()
+            elif not playing and timeline.is_playing():
+                timeline.pause()
+            velocity, yaw_rate, stop_reason, decision = ai.command(
+                now, frame_deadline=FRAME_DEADLINE_SECONDS
+            )
+            transform = body_transform()
+            q = transform.ExtractRotationQuat()
+            pose = list(transform.ExtractTranslation()) + list(q.GetImaginary()) + [q.GetReal()]
+            unstable |= inclination(pose) >= PROFILE["stop_tilt_degrees"]
+            if unstable:
+                velocity, yaw_rate, stop_reason = 0.0, 0.0, "unstable"
+            terminal_reason = (
+                "unstable" if unstable else (decision["decision"]["reason"] if decision else None)
+            )
+            if terminal_reason in {"unstable", "stuck", "goal-reached"}:
+                velocity, yaw_rate = 0.0, 0.0
+                terminal_at = now if terminal_at is None else terminal_at
+                if now - terminal_at >= 2:
+                    result = {
+                        "outcome": "completed" if terminal_reason == "goal-reached" else "failed",
+                        "message": {
+                            "goal-reached": "Ровер достиг цели маршрута.",
+                            "stuck": "Ровер не нашёл проезд после трёх попыток. Прогон завершён.",
+                            "unstable": "Прогон завершён из-за опасного наклона ровера.",
+                        }[terminal_reason],
+                    }
+                    break
+            # The physics tensor API applies live drive targets in radians/s;
+            # USD authoring attributes are only the initial scene configuration.
+            if not playing:
+                drive.command(*envelope.step(0.0, 0.0, 0.0, stop=True))
+                update_cameras()
+                state.update(
+                    velocity=0.0, yaw_rate=0.0, stop_reason="paused", speed=0.0, decision=decision
+                )
+            app.update()  # Never waits for inference, a network request, or an operator ACK.
+            if callback_errors:
+                raise RuntimeError(callback_errors[0])
+            render_frames += 1
+            captured_at = time.monotonic()
+            transform, position = state["transform"], state["position"]
+            eye, camera_axes = state["eye"], state["camera_axes"]
+            pose, yaw, speed, mode = state["pose"], state["yaw"], state["speed"], state["mode"]
+            velocity, yaw_rate = state["velocity"], state["yaw_rate"]
+            stop_reason, decision = state["stop_reason"], state["decision"]
+            if physics_rows:
+                physics_motion.writelines(
+                    json.dumps(row, allow_nan=False) + "\n" for row in physics_rows
+                )
+                physics_rows.clear()
+                physics_motion.flush()
+            physics = SimulationManager.get_num_physics_steps()
+            sim_ns = round((physics - baseline) * 1e9 / 60)
+            if playing and captured_at - last_sensor >= 1 / 5:
+                rgb = annotator.get_data()
+                if isinstance(rgb, np.ndarray) and rgb.shape[:2] == (600, 800):
+                    sensor_frames += 1
+                    rear_eye = transform.Transform(
+                        Gf.Vec3d(
+                            PROFILE["rear_range_forward_m"],
+                            0,
+                            settings["camera_height_m"] - PROFILE["body_contact_height_m"],
+                        )
+                    )
+                    points = range_sensor.capture(eye, transform, rear_eye)
+                    quaternion = transform.ExtractRotationQuat()
+                    imaginary = quaternion.GetImaginary()
+                    # Columns are the camera's forward/left/up axes in world.
+                    observation = {
+                        "simulation_time_ns": sim_ns,
+                        "points": points,
+                        "pose": [float(v) for v in position]
+                        + [float(v) for v in imaginary]
+                        + [float(quaternion.GetReal())],
+                        "calibration": {
+                            "origin": list(eye),
+                            "rotation": np.array(camera_axes).T.reshape(-1).tolist(),
+                            "intrinsics": [focal_pixels, focal_pixels, 400.0, 300.0],
+                            "body_contact_height_m": PROFILE["body_contact_height_m"],
+                            "range_origins": [list(eye), list(rear_eye)],
+                        },
+                    }
+                    ai.submit(
+                        np.array(rgb[:, :, :3], copy=True, order="C"),
+                        sensor_frames,
+                        captured_at,
+                        sim_ns,
+                        observation,
+                    )
+                last_sensor = captured_at
+            if captured_at - last_report >= 0.25:
+                prior_at, prior_render, prior_sensor, prior_ai, prior_physics = prior_metrics
+                elapsed = captured_at - prior_at
+                report = dict(
+                    sequence=sequence,
+                    control_sequence=control["control_sequence"],
+                    state="running" if playing else "paused" if had_ai else "ready",
+                    phase="models" if playing and not ai.ever_ready else "running",
+                    simulation_time_ns=sim_ns,
+                    wall_elapsed_seconds=captured_at - start,
+                    physics_steps=physics - baseline,
+                    render_frames=render_frames,
+                    sensor_frames=sensor_frames,
+                    inference_count=ai.count,
+                    dropped_frames=ai.dropped,
+                    rtf=(physics - prior_physics) / 60 / elapsed,
+                    render_fps=(render_frames - prior_render) / elapsed,
+                    sensor_fps=(sensor_frames - prior_sensor) / elapsed,
+                    ai_hz=(ai.count - prior_ai) / elapsed,
+                    inference_ms=decision["inference_ms"] if decision else None,
+                    frame_age_ms=(captured_at - decision["captured_at"]) * 1000
+                    if decision
+                    else None,
+                    command_age_ms=(captured_at - decision["completed_at"]) * 1000
+                    if decision
+                    else None,
+                    pose_xy=[float(position[0]), float(position[1])],
+                    pose_yaw=yaw,
+                    speed_mps=speed,
+                    applied_speed_mps=velocity,
+                    applied_yaw_rate_rps=yaw_rate,
+                    decision=decision["decision"] if decision else None,
+                    stop_reason=stop_reason,
+                    ai_ready=ai.ready,
+                    stream_ready=True,
+                    camera=mode,
+                )
+                channel.write("snapshot", report)
+                motion.write(
+                    json.dumps(
+                        {
+                            **report,
+                            "monotonic": captured_at,
+                            "pose_xyz_xyzw": pose,
+                            "tilt_degrees": inclination(pose),
+                            "wheel_velocity_rps": articulation.get_dof_velocities(
+                                dof_indices=wheel_indices
+                            )
+                            .numpy()
+                            .tolist()[0],
+                            "usd_position_xyz": list(
+                                UsdGeom.XformCache()
+                                .GetLocalToWorldTransform(
+                                    stage.GetPrimAtPath("/World/Rover/chassis")
+                                )
+                                .ExtractTranslation()
+                            ),
+                        },
+                        allow_nan=False,
+                    )
+                    + "\n"
+                )
+                motion.flush()
+                last_report = captured_at
+                sequence += 1
+                if elapsed >= 1:
+                    prior_metrics = (captured_at, render_frames, sensor_frames, ai.count, physics)
+                if shutil.disk_usage(directory).free < 1024**3:
+                    raise RuntimeError("Worker evidence disk is full")
+except Exception as exc:
+    message = str(exc)
+    placement_failure = any(
+        text in message
+        for text in (
+            "Configured spawn",
+            "Rover start",
+            "reconstructed ground",
+            "reconstructed support",
+        )
+    )
+    result = {
+        "outcome": "failed",
+        "message": "Не удалось устойчиво разместить ровер на грунте. Измените положение старта."
+        if placement_failure
+        else "Симуляция прервана из-за ошибки на Worker. Журнал сохранён.",
+    }
+    traceback.print_exc()
+finally:
+    for callback_id in physics_callbacks:
+        SimulationManager.deregister_callback(callback_id)
+    channel.close()
+    if ai is not None:
+        ai.close()
+    write_json(directory / "result.json", result)
+    if app is not None:
+        app.close()
diff --git a/simulation/ai-polygon/run_scene.py b/simulation/ai-polygon/run_scene.py
new file mode 100644
index 0000000..13c0b0f
--- /dev/null
+++ b/simulation/ai-polygon/run_scene.py
@@ -0,0 +1,206 @@
+"""One native Isaac episode. Receives pixels from physics; emits virtual commands only."""
+
+import argparse
+import base64
+import io
+import json
+import math
+import sys
+import time
+from pathlib import Path
+
+from core_client import CoreClient
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--run", type=Path, required=True)
+parser.add_argument("--source", type=Path, required=True)
+parser.add_argument("--core", required=True)
+parser.add_argument("--token-file", type=Path, required=True)
+parser.add_argument("--instance", required=True)
+args = parser.parse_args()
+root = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(root / "src"))
+run = json.loads(args.run.read_text())
+settings = run["world"]["settings"]
+client = CoreClient(args.core, args.token_file, args.instance)
+result = {"outcome": "failed", "message": "Симуляция не завершена."}
+app = model = robot = controller = None
+try:
+    from isaacsim import SimulationApp
+
+    app = SimulationApp({"headless": True, "multi_gpu": False, "width": 800, "height": 600})
+    import numpy as np
+
+    # Isaac bundles the converter compatible with its own USD/ParticleField schema.
+    import omni.kit.app
+    import omni.replicator.core as rep
+    import omni.usd
+    from isaacsim.core.experimental.utils import app as app_utils
+    from isaacsim.core.simulation_manager import SimulationManager
+    from isaacsim.robot.experimental.wheeled_robots.controllers import DifferentialController
+    from isaacsim.robot.experimental.wheeled_robots.robots import WheeledRobot
+    from PIL import Image
+    from pxr import Gf, UsdGeom, UsdLux, UsdPhysics
+
+    from k1link.simulation.ai_polygon.inference import ModelInference
+    from k1link.simulation.ai_polygon.policy import RoadPolicy
+
+    omni.kit.app.get_app().get_extension_manager().set_extension_enabled_immediate(
+        "omni.kit.converter.gsplat", True
+    )
+    from usd_convert_gsplat import read_ply, write_gaussian_splat_usd
+
+    converted = args.source.with_suffix(".usd")
+    if not converted.exists():
+        temporary = converted.with_suffix(".part.usd")
+        write_gaussian_splat_usd(
+            read_ply(str(args.source)),
+            str(temporary),
+            source_file=str(args.source),
+            prim_name="Gaussians",
+            up_axis="Z",
+        )
+        temporary.replace(converted)
+    stage = omni.usd.get_context().get_stage()
+    UsdGeom.SetStageUpAxis(stage, UsdGeom.Tokens.z)
+    UsdGeom.SetStageMetersPerUnit(stage, 1)
+    visual = UsdGeom.Xform.Define(stage, "/World/Scan")
+    visual.AddRotateXYZOp().Set(Gf.Vec3f(*settings["rotation_degrees"]))
+    visual.AddScaleOp().Set(Gf.Vec3f(settings["meters_per_unit"]))
+    stage.DefinePrim("/World/Scan/Gaussians").GetReferences().AddReference(str(converted))
+    # Only a hidden rigid ground plane. Vegetation in the scan is visual-only.
+    ground = UsdGeom.Plane.Define(stage, "/World/Ground")
+    ground.CreateAxisAttr("Z")
+    ground.CreateWidthAttr(20000)
+    ground.CreateLengthAttr(20000)
+    ground.AddTranslateOp().Set(Gf.Vec3d(0, 0, settings["ground_z"]))
+    UsdPhysics.CollisionAPI.Apply(ground.GetPrim())
+    ground.CreateVisibilityAttr(UsdGeom.Tokens.invisible)
+    UsdLux.DomeLight.Define(stage, "/World/Light").CreateIntensityAttr(500)
+    robot_asset = Path("D:/NDC_MISSIONCORE/runtime/simulation/assets/jetbot-6.1-v1/jetbot.usda")
+    if not robot_asset.is_file():
+        raise RuntimeError("Prepare the pinned stock Jetbot asset cache first")
+    heading = math.radians(settings["heading_degrees"])
+    robot = WheeledRobot(
+        paths="/World/Rover",
+        wheel_dof_names=["left_wheel_joint", "right_wheel_joint"],
+        usd_path=str(robot_asset),
+        positions=[*settings["spawn_xy"], settings["ground_z"] + 0.05],
+        orientations=[math.cos(heading / 2), 0, 0, math.sin(heading / 2)],
+    )
+    controller = DifferentialController(wheel_radius=0.03, wheel_base=0.1125)
+    camera = UsdGeom.Camera.Define(stage, "/World/Camera")
+    camera.CreateFocalLengthAttr(24)
+    camera.CreateHorizontalApertureAttr(36)
+    camera.CreateVerticalApertureAttr(27)
+    camera.CreateClippingRangeAttr(Gf.Vec2f(0.02, 1000))
+    camera_pos = camera.AddTranslateOp()
+    camera_rot = camera.AddOrientOp()
+    render_product = rep.create.render_product(str(camera.GetPath()), (800, 600))
+    annotator = rep.AnnotatorRegistry.get_annotator("rgb")
+    annotator.attach(render_product)
+    SimulationManager.setup_simulation(dt=1 / 60, device="cpu")
+    app_utils.play()
+    app_utils.update_app(steps=20)
+    app_utils.pause()
+    baseline = SimulationManager.get_num_physics_steps()
+    profile = json.loads((root / "simulation/ai-polygon/models.worker-006.json").read_text())
+    model = ModelInference(
+        "http://127.0.0.1:18092", Path(profile["labels"]), "http://127.0.0.1:18091"
+    )
+    model.ready()
+    policy = RoadPolicy(settings["max_speed_mps"])
+
+    def position():
+        positions, rotations = robot.get_world_poses()
+        xyz, q = positions.numpy()[0], rotations.numpy()[0]
+        yaw = math.atan2(2 * (q[0] * q[3] + q[1] * q[2]), 1 - 2 * (q[2] ** 2 + q[3] ** 2))
+        return xyz, yaw
+
+    sequence = 0
+    last_cycle_ms = 0.0
+    while sequence < run["request"]["max_steps"]:
+        cycle_started = time.monotonic()
+        transport_ms = 0.0
+        transport_started = time.monotonic()
+        action = client.request(
+            "/worker/poll", {"instance_id": args.instance, "run_id": run["run_id"]}
+        )["action"]
+        transport_ms += (time.monotonic() - transport_started) * 1000
+        if action == "stop":
+            result = {"outcome": "stopped", "message": "Прогон остановлен."}
+            break
+        if action == "pause":
+            robot.apply_wheel_actions(controller.forward([0, 0]))
+            time.sleep(0.1)
+            continue
+        if action not in ("play", "step"):
+            raise RuntimeError("Core withdrew simulation ownership")
+        xyz, yaw = position()
+        # Pose is used solely to place the virtual sensor, never passed to policy.
+        eye = Gf.Vec3d(float(xyz[0]), float(xyz[1]), float(xyz[2] + settings["camera_height_m"]))
+        camera_pos.Set(eye)
+        # USD cameras look along local -Z with +Y up. Invert a world-to-camera
+        # look-at transform to keep the optical horizon level for every heading.
+        view = Gf.Matrix4d().SetLookAt(
+            eye, eye + Gf.Vec3d(math.cos(yaw), math.sin(yaw), 0), Gf.Vec3d(0, 0, 1)
+        )
+        camera_rot.Set(Gf.Quatf(view.GetInverse().ExtractRotationQuat()))
+        before = SimulationManager.get_num_physics_steps()
+        render_started = time.monotonic()
+        rep.orchestrator.step(rt_subframes=1, delta_time=0.0, pause_timeline=True)
+        if SimulationManager.get_num_physics_steps() != before:
+            raise RuntimeError("Rendering advanced the simulation clock")
+        render_ms = (time.monotonic() - render_started) * 1000
+        rgb = np.ascontiguousarray(annotator.get_data()[:, :, :3])
+        started = time.monotonic_ns()
+        road, boxes = model.infer(rgb)
+        decision = policy.decide(road, boxes)
+        inference_ms = (time.monotonic_ns() - started) / 1e6
+        image = io.BytesIO()
+        Image.fromarray(rgb).save(image, "JPEG", quality=85)
+        transport_started = time.monotonic()
+        client.request(
+            "/worker/runs/" + run["run_id"] + "/samples",
+            {
+                "sequence": sequence,
+                "simulation_time_ns": sequence * run["step_ns"],
+                "inference_ms": inference_ms,
+                "pose_xy": [float(xyz[0]), float(xyz[1])],
+                "decision": decision.model_dump(),
+                "image_jpeg_base64": base64.b64encode(image.getvalue()).decode(),
+            },
+        )
+        transport_ms += (time.monotonic() - transport_started) * 1000
+        robot.apply_wheel_actions(controller.forward([decision.speed_mps, decision.yaw_rate_rps]))
+        SimulationManager.step(steps=6)
+        if SimulationManager.get_num_physics_steps() != baseline + (sequence + 1) * 6:
+            raise RuntimeError("Physics clock left lockstep")
+        xyz, yaw = position()
+        transport_started = time.monotonic()
+        client.request(
+            "/worker/runs/" + run["run_id"] + "/applied",
+            {
+                "sequence": sequence,
+                "simulation_time_ns": (sequence + 1) * run["step_ns"],
+                "physics_steps": 6,
+                "pose_yaw": float(yaw),
+                "cycle_ms": last_cycle_ms or (time.monotonic() - cycle_started) * 1000,
+                "render_ms": render_ms,
+                "transport_ms": transport_ms,
+                "pose_xy": [float(xyz[0]), float(xyz[1])],
+            },
+        )
+        last_cycle_ms = (time.monotonic() - cycle_started) * 1000
+        sequence += 1
+    else:
+        result = {"outcome": "completed", "message": "Прогон завершён."}
+except Exception as exc:
+    result = {"outcome": "failed", "message": "Прогон прерван. Подробности сохранены на Worker."}
+    print(type(exc).__name__ + ": " + str(exc), file=sys.stderr, flush=True)
+finally:
+    if model is not None:
+        model.close()
+    args.run.with_name("result.json").write_text(json.dumps(result))
+    if app is not None:
+        app.close()
diff --git a/simulation/ai-polygon/segformer/Dockerfile b/simulation/ai-polygon/segformer/Dockerfile
new file mode 100644
index 0000000..03193bb
--- /dev/null
+++ b/simulation/ai-polygon/segformer/Dockerfile
@@ -0,0 +1,10 @@
+FROM ndc/mission-core-ai-module-ddrnet:20260904-v8
+RUN /opt/conda/envs/goose/bin/python -m pip install --no-cache-dir --no-deps \
+    transformers==4.44.2 tokenizers==0.19.1 safetensors==0.4.5 \
+    huggingface-hub==0.24.6 regex==2024.9.11 packaging==24.1 \
+    filelock==3.16.1 fsspec==2024.9.0 PyYAML==6.0.2 \
+    requests==2.32.3 tqdm==4.66.5 typing-extensions==4.12.2
+ENV HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 TOKENIZERS_PARALLELISM=false
+LABEL com.nodedc.product=mission-core com.nodedc.stack=ai-polygon \
+    com.nodedc.role=ai-module com.nodedc.managed-by=ai-polygon-worker
+ENTRYPOINT ["/opt/conda/envs/goose/bin/python", "-B", "/adapter/segformer/server.py"]
diff --git a/simulation/ai-polygon/segformer/Prepare-Segformer.ps1 b/simulation/ai-polygon/segformer/Prepare-Segformer.ps1
new file mode 100644
index 0000000..be0c5ee
--- /dev/null
+++ b/simulation/ai-polygon/segformer/Prepare-Segformer.ps1
@@ -0,0 +1,37 @@
+param([string]$Root = 'D:\NDC_MISSIONCORE\runtime\simulation')
+$ErrorActionPreference='Stop'
+$ProgressPreference='SilentlyContinue'
+[Console]::OutputEncoding=[System.Text.Encoding]::UTF8
+$revision='de01bae28967510f9ddd496c60a969357195400c'
+$out=Join-Path $Root ('assets\segformer-b2-ade\'+$revision)
+New-Item -ItemType Directory -Force $out | Out-Null
+$files=@{
+    'config.json'='ee7400840fdb1e5045f0b2eba78bf053df8e33a309c4acec31705a48c8cc5c00'
+    'preprocessor_config.json'='8039d1d210abaa7117ad78e58cdfd6141a2ec72c03dae891b3cd76737e422c6c'
+    'README.md'='7b532a0053fc1769553386090fbc928ed8d0f5b5d5b20cfa8f51e46f56ef3c6d'
+    'pytorch_model.bin'='187ca07bea003a5717c63d04ea90b07f33cd033c0ebf44b4b89fce5070d6c8f3'
+}
+foreach($name in $files.Keys) {
+    $path=Join-Path $out $name
+    if (!(Test-Path $path)) {
+        $part=$path+'.part'
+        Invoke-WebRequest -UseBasicParsing -Uri "https://huggingface.co/nvidia/segformer-b2-finetuned-ade-512-512/resolve/$revision/$name" -OutFile $part
+        if ((Get-FileHash $part -Algorithm SHA256).Hash.ToLower() -ne $files[$name]) { throw "Downloaded model checksum mismatch: $name" }
+        Move-Item $part $path
+    }
+    if ((Get-FileHash $path -Algorithm SHA256).Hash.ToLower() -ne $files[$name]) { throw "Installed model checksum mismatch: $name" }
+}
+$ErrorActionPreference='Continue'
+$base=docker image inspect ndc/mission-core-ai-module-ddrnet:20260904-v8 --format '{{.Id}}'
+if ($LASTEXITCODE -ne 0 -or $base.Trim() -ne 'sha256:a3b7d22f5d3bfdf2d84444b936c8b01abf8243be652387d7e2024ba7bda587f5') {throw 'Pinned base image changed'}
+$tag='ndc/mission-core-ai-module-segformer:de01bae2-v1'
+$image=docker image inspect $tag --format '{{.Id}}' 2>$null
+if ($LASTEXITCODE -ne 0) {
+    docker build --pull=false --progress plain -t $tag -f (Join-Path $PSScriptRoot 'Dockerfile') $PSScriptRoot
+    if ($LASTEXITCODE -ne 0) {throw 'SegFormer image build failed'}
+    $image=docker image inspect $tag --format '{{.Id}}'
+    if ($LASTEXITCODE -ne 0) {throw 'SegFormer image unavailable'}
+}
+$receipt=@{schema_version='missioncore.ai-polygon-segformer-install/v1';revision=$revision;image=$image.Trim();tag=$tag;files=$files;assets=$out;installed_at=[DateTime]::UtcNow.ToString('o');license='NVIDIA SegFormer research/evaluation; see upstream model card'}
+[IO.File]::WriteAllText((Join-Path $out 'installation.json'),($receipt|ConvertTo-Json -Depth 4),[Text.UTF8Encoding]::new($false))
+$receipt|ConvertTo-Json -Depth 4 -Compress
diff --git a/simulation/ai-polygon/segformer/qualify.py b/simulation/ai-polygon/segformer/qualify.py
new file mode 100644
index 0000000..0678ba7
--- /dev/null
+++ b/simulation/ai-polygon/segformer/qualify.py
@@ -0,0 +1,52 @@
+"""Worker-only offline SegFormer comparison; no live control authority."""
+
+import argparse
+import hashlib
+import json
+import time
+from pathlib import Path
+
+import numpy as np
+import torch
+from PIL import Image
+from transformers import SegformerForSemanticSegmentation, SegformerImageProcessor
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--image", type=Path, required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    args = parser.parse_args()
+    processor = SegformerImageProcessor.from_pretrained("/assets", local_files_only=True)
+    model = (
+        SegformerForSemanticSegmentation.from_pretrained("/assets", local_files_only=True)
+        .cuda()
+        .eval()
+    )
+    image = Image.open(args.image).convert("RGB").crop((100, 0, 700, 600))
+    tensor = processor(images=image, return_tensors="pt")["pixel_values"].cuda()
+    times = []
+    with torch.inference_mode():
+        for _ in range(6):
+            torch.cuda.synchronize()
+            start = time.monotonic()
+            logits = model(tensor).logits
+            logits = torch.nn.functional.interpolate(
+                logits, size=(512, 512), mode="bilinear", align_corners=False
+            )
+            labels = logits.argmax(1)[0].cpu().numpy().astype(np.uint8)
+            times.append((time.monotonic() - start) * 1000)
+    args.output.mkdir(parents=True, exist_ok=True)
+    Image.fromarray(labels).save(args.output / "labels.png")
+    ids, counts = np.unique(labels, return_counts=True)
+    report = {
+        "source_sha256": hashlib.sha256(args.image.read_bytes()).hexdigest(),
+        "classes": {model.config.id2label[int(i)]: int(counts[n]) for n, i in enumerate(ids)},
+        "inference_ms": times,
+    }
+    (args.output / "report.json").write_text(json.dumps(report, indent=2), encoding="utf-8")
+    print(json.dumps(report))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/simulation/ai-polygon/segformer/server.py b/simulation/ai-polygon/segformer/server.py
new file mode 100644
index 0000000..ff1578a
--- /dev/null
+++ b/simulation/ai-polygon/segformer/server.py
@@ -0,0 +1,87 @@
+"""Resident, pinned ADE20K surface provider. RGB only; no scene truth input."""
+
+import hashlib
+import json
+from http.server import BaseHTTPRequestHandler, HTTPServer
+from pathlib import Path
+
+import numpy as np
+import torch
+from PIL import Image
+from transformers import SegformerForSemanticSegmentation, SegformerImageProcessor
+
+# These are surface candidates, not permission to drive. The metric terrain
+# planner still checks step height, slope and the full rover footprint.
+SURFACES = (3, 6, 9, 11, 13, 29, 34, 46, 52, 91)
+FILES = {
+    "config.json": "ee7400840fdb1e5045f0b2eba78bf053df8e33a309c4acec31705a48c8cc5c00",
+    "preprocessor_config.json": "8039d1d210abaa7117ad78e58cdfd6141a2ec72c03dae891b3cd76737e422c6c",
+    "pytorch_model.bin": "187ca07bea003a5717c63d04ea90b07f33cd033c0ebf44b4b89fce5070d6c8f3",
+}
+
+
+def main():
+    for name, expected in FILES.items():
+        if hashlib.sha256((Path("/assets") / name).read_bytes()).hexdigest() != expected:
+            raise RuntimeError("SegFormer asset identity changed: " + name)
+    processor = SegformerImageProcessor.from_pretrained("/assets", local_files_only=True)
+    model = (
+        SegformerForSemanticSegmentation.from_pretrained("/assets", local_files_only=True)
+        .cuda()
+        .eval()
+    )
+    torch.set_num_threads(2)
+
+    def infer(rgb):
+        image = Image.fromarray(rgb).crop((100, 0, 700, 600))
+        tensor = processor(images=image, return_tensors="pt")["pixel_values"].cuda()
+        with torch.inference_mode():
+            logits = model(tensor).logits
+            logits = torch.nn.functional.interpolate(
+                logits, size=(512, 512), mode="bilinear", align_corners=False
+            )
+            confidence, labels = logits.softmax(1).max(1)
+        labels = labels[0].cpu().numpy().astype(np.uint8)
+        candidate = np.isin(labels, SURFACES) & (confidence[0].cpu().numpy() >= 0.55)
+        return labels.tobytes() + candidate.astype(np.uint8).tobytes()
+
+    infer(np.zeros((600, 800, 3), dtype=np.uint8))
+
+    class Handler(BaseHTTPRequestHandler):
+        def setup(self):
+            super().setup()
+            self.connection.settimeout(10)
+
+        def reply(self, status, body):
+            self.send_response(status)
+            self.send_header("Content-Type", "application/octet-stream")
+            self.send_header("Content-Length", str(len(body)))
+            self.end_headers()
+            self.wfile.write(body)
+
+        def do_GET(self):
+            self.reply(
+                200 if self.path == "/ready" else 404,
+                json.dumps({"model": "segformer-b2-ade150"}).encode(),
+            )
+
+        def do_POST(self):
+            if self.path != "/infer" or self.headers.get("Content-Length") != "1440000":
+                self.reply(400, b"Expected 800x600 raw RGB uint8")
+                return
+            try:
+                raw = self.rfile.read(1440000)
+                if len(raw) != 1440000:
+                    raise ValueError("Incomplete camera frame")
+                self.reply(200, infer(np.frombuffer(raw, np.uint8).reshape(600, 800, 3)))
+            except (TimeoutError, ValueError, RuntimeError):
+                self.reply(500, b"Surface inference failed")
+
+        def log_message(self, *_):
+            pass
+
+    HTTPServer(("0.0.0.0", 8010), Handler).serve_forever()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/simulation/ai-polygon/spawn_clearance.py b/simulation/ai-polygon/spawn_clearance.py
new file mode 100644
index 0000000..31a3e0a
--- /dev/null
+++ b/simulation/ai-polygon/spawn_clearance.py
@@ -0,0 +1,54 @@
+"""Scene-authoring admission for a full rover footprint, never an AI map.
+
+Ground-height samples can miss a narrow trunk between sample rays. Test mesh
+triangles against the occupied prism using the separating-axis theorem, which
+also catches a face crossing the body when all of its vertices lie outside.
+"""
+
+import math
+
+import numpy as np
+
+
+def obstructing_triangles(vertices, faces, xy, heading, plane, *, step=0.1, height=1.1):
+    """Count triangles above qualified step height inside the 1 x 1 m start.
+
+    ``plane`` is z = a*(x-xy[0]) + b*(y-xy[1]) + c, fitted to the start's support.
+    This is a conservative preparation gate, not a claim of route traversability.
+    It neither changes the collider nor supplies privileged geometry to inference.
+    """
+    triangles = np.asarray(vertices)[faces]
+    center_xy = np.asarray(xy)
+    selected = (triangles[:, :, :2].min(axis=1) <= center_xy + 0.71).all(axis=1) & (
+        triangles[:, :, :2].max(axis=1) >= center_xy - 0.71
+    ).all(axis=1)
+    triangles = triangles[selected].astype(np.float64)
+    if not len(triangles):
+        return 0
+    delta = triangles[:, :, :2] - center_xy
+    angle = math.radians(heading)
+    rotation = np.array([[math.cos(angle), -math.sin(angle)], [math.sin(angle), math.cos(angle)]])
+    triangles[:, :, 2] -= delta @ np.asarray(plane[:2]) + plane[2]
+    triangles[:, :, :2] = delta @ rotation
+    low = step + 1e-4  # Same centimetre-scale capability boundary as navigation.
+    half = np.array([0.5, 0.5, (height - low) / 2])
+    triangles[:, :, 2] -= (height + low) / 2
+    selected = (triangles.min(axis=1) <= half).all(axis=1) & (triangles.max(axis=1) >= -half).all(
+        axis=1
+    )
+    triangles = triangles[selected]
+    if not len(triangles):
+        return 0
+    edges = np.roll(triangles, -1, axis=1) - triangles
+    axes = [np.cross(edges[:, 0], edges[:, 1])]
+    for edge in range(3):
+        for box_axis in np.eye(3):
+            axes.append(np.cross(edges[:, edge], box_axis))
+    overlaps = np.ones(len(triangles), dtype=bool)
+    for axis in axes:
+        projections = np.einsum("nvi,ni->nv", triangles, axis)
+        radius = np.abs(axis) @ half
+        overlaps &= (projections.min(axis=1) <= radius + 1e-10) & (
+            projections.max(axis=1) >= -radius - 1e-10
+        )
+    return int(overlaps.sum())
diff --git a/simulation/ai-polygon/splat-tools/package-lock.json b/simulation/ai-polygon/splat-tools/package-lock.json
new file mode 100644
index 0000000..5efc583
--- /dev/null
+++ b/simulation/ai-polygon/splat-tools/package-lock.json
@@ -0,0 +1,101 @@
+{
+  "name": "ndc-ai-polygon-splat-tools",
+  "version": "1.0.0",
+  "lockfileVersion": 3,
+  "requires": true,
+  "packages": {
+    "": {
+      "name": "ndc-ai-polygon-splat-tools",
+      "version": "1.0.0",
+      "dependencies": {
+        "@playcanvas/splat-transform": "3.4.0"
+      }
+    },
+    "node_modules/@adobe/spz": {
+      "version": "0.2.3",
+      "resolved": "https://registry.npmjs.org/@adobe/spz/-/spz-0.2.3.tgz",
+      "integrity": "sha512-iMiIB+FUxQ2vaNxP012fdIVd4yugyq0PUXBaknGfYDE3xpTNU5PrhbLtnDTGFE9anKOwjneOuanSKqaaNpC68Q==",
+      "license": "ISC"
+    },
+    "node_modules/@playcanvas/splat-transform": {
+      "version": "3.4.0",
+      "resolved": "https://registry.npmjs.org/@playcanvas/splat-transform/-/splat-transform-3.4.0.tgz",
+      "integrity": "sha512-I/lZFfHNzonye1QIne/2nF+wDbMx0xHK4k1IpzJmD3TfNc8ftmuVsWWKszYPkScWDG1vq2n9QjMXS9DARUPmcQ==",
+      "license": "MIT",
+      "dependencies": {
+        "@adobe/spz": "0.2.3",
+        "webgpu": "0.6.0"
+      },
+      "bin": {
+        "splat-transform": "bin/cli.mjs"
+      },
+      "engines": {
+        "node": ">=22.0.0"
+      },
+      "peerDependencies": {
+        "playcanvas": "^2.0.0"
+      }
+    },
+    "node_modules/@types/webxr": {
+      "version": "0.5.24",
+      "resolved": "https://registry.npmjs.org/@types/webxr/-/webxr-0.5.24.tgz",
+      "integrity": "sha512-h8fgEd/DpoS9CBrjEQXR+dIDraopAEfu4wYVNY2tEPwk60stPWhvZMf4Foo5FakuQ7HFZoa8WceaWFervK2Ovg==",
+      "license": "MIT",
+      "peer": true
+    },
+    "node_modules/@webgpu/types": {
+      "version": "0.1.74",
+      "resolved": "https://registry.npmjs.org/@webgpu/types/-/types-0.1.74.tgz",
+      "integrity": "sha512-lgiI4hbuLcI9unnm2cL/tvCaQU45dp0xcLWJh5uB/9MBGvIA4F8XIk7nSiBeg+K4xWGYwgZKn80LmERI5CxoTA==",
+      "license": "BSD-3-Clause"
+    },
+    "node_modules/debug": {
+      "version": "4.4.3",
+      "resolved": "https://registry.npmjs.org/debug/-/debug-4.4.3.tgz",
+      "integrity": "sha512-RGwwWnwQvkVfavKVt22FGLw+xYSdzARwm0ru6DhTVA3umU5hZc28V3kO4stgYryrTlLpuvgI9GiijltAjNbcqA==",
+      "license": "MIT",
+      "dependencies": {
+        "ms": "^2.1.3"
+      },
+      "engines": {
+        "node": ">=6.0"
+      },
+      "peerDependenciesMeta": {
+        "supports-color": {
+          "optional": true
+        }
+      }
+    },
+    "node_modules/ms": {
+      "version": "2.1.3",
+      "resolved": "https://registry.npmjs.org/ms/-/ms-2.1.3.tgz",
+      "integrity": "sha512-6FlzubTLZG3J2a/NVCAleEhjzq5oxgHyaCU9yYXvcLsvoVaHJq/s5xXI6/XXP6tz7R9xAOtHnSO/tXtF3WRTlA==",
+      "license": "MIT"
+    },
+    "node_modules/playcanvas": {
+      "version": "2.22.3",
+      "resolved": "https://registry.npmjs.org/playcanvas/-/playcanvas-2.22.3.tgz",
+      "integrity": "sha512-HjBbJmUqwqYk678nHu/hUniep0YgziAJ9oqwi13ye/rlDjIN2/bbmnCZexKa5IZDINSMkjqfhq+jsJhPL4wBNA==",
+      "license": "MIT",
+      "peer": true,
+      "dependencies": {
+        "@types/webxr": "^0.5.24",
+        "@webgpu/types": "^0.1.70"
+      },
+      "engines": {
+        "node": ">=18.3.0"
+      }
+    },
+    "node_modules/webgpu": {
+      "version": "0.6.0",
+      "resolved": "https://registry.npmjs.org/webgpu/-/webgpu-0.6.0.tgz",
+      "integrity": "sha512-z44ZU/+ypPELgQrv+um3/5Eqcq7gOtVZAmaRfwujJMwGMRSa6+CmjIAvEWw7O68Y95cnEsywhbYlcChQTX+IQg==",
+      "hasInstallScript": true,
+      "license": "MIT",
+      "dependencies": {
+        "@webgpu/types": "^0.1.72",
+        "debug": "^4.4.0"
+      }
+    }
+  }
+}
diff --git a/simulation/ai-polygon/splat-tools/package.json b/simulation/ai-polygon/splat-tools/package.json
new file mode 100644
index 0000000..8936979
--- /dev/null
+++ b/simulation/ai-polygon/splat-tools/package.json
@@ -0,0 +1,6 @@
+{
+  "name": "ndc-ai-polygon-splat-tools",
+  "version": "1.0.0",
+  "private": true,
+  "dependencies": { "@playcanvas/splat-transform": "3.4.0" }
+}
diff --git a/simulation/ai-polygon/start_worker.py b/simulation/ai-polygon/start_worker.py
new file mode 100644
index 0000000..c3e1e00
--- /dev/null
+++ b/simulation/ai-polygon/start_worker.py
@@ -0,0 +1,44 @@
+"""Start the prepared Worker and its private tunnel in one foreground SSH session."""
+
+import argparse
+import base64
+import os
+import re
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--release", required=True, help="16-character build_bundle identity prefix")
+parser.add_argument("--once", action="store_true", help="Exit after one episode")
+args = parser.parse_args()
+if not re.fullmatch(r"[a-f0-9]{16}", args.release):
+    parser.error("release must be a 16-character lowercase hexadecimal bundle identity")
+script = r"""
+$ErrorActionPreference='Stop'
+$ProgressPreference='SilentlyContinue'
+[Console]::OutputEncoding=[System.Text.Encoding]::UTF8
+$b='D:\NDC_MISSIONCORE\runtime\simulation'
+$worker="$b\releases\ai-polygon-RELEASE\simulation\ai-polygon\worker.py"
+& "$b\isaac-sim-6.1.0\kit\python\python.exe" -u $worker `
+    --token-file "$b\private\worker.token" --state "$b\state" `
+    --isaac "$b\isaac-sim-6.1.0" ONCE
+exit $LASTEXITCODE
+""".replace("RELEASE", args.release).replace("ONCE", "--once" if args.once else "")
+encoded = base64.b64encode(script.encode("utf-16-le")).decode()
+os.execvp(
+    "ssh",
+    [
+        "ssh",
+        "-T",
+        "-o",
+        "BatchMode=yes",
+        "-o",
+        "ExitOnForwardFailure=yes",
+        "-o",
+        "ServerAliveInterval=5",
+        "-o",
+        "ServerAliveCountMax=3",
+        "-R",
+        "127.0.0.1:18081:127.0.0.1:8000",
+        "mission-gpu",
+        "powershell.exe -NoProfile -NonInteractive -EncodedCommand " + encoded,
+    ],
+)
diff --git a/simulation/ai-polygon/terrain.py b/simulation/ai-polygon/terrain.py
new file mode 100644
index 0000000..6560515
--- /dev/null
+++ b/simulation/ai-polygon/terrain.py
@@ -0,0 +1,202 @@
+"""Metric collision proxy and occlusion-aware virtual range sensor on Worker."""
+
+import hashlib
+import json
+import math
+import struct
+from pathlib import Path
+
+import numpy as np
+
+
+def load_glb(path):
+    data = Path(path).read_bytes()
+    magic, version, size = struct.unpack_from("<III", data)
+    if (magic, version, size) != (0x46546C67, 2, len(data)):
+        raise ValueError("Expected a complete glTF 2 binary collision mesh")
+    offset, document, binary = 12, None, None
+    while offset < len(data):
+        length, kind = struct.unpack_from("<II", data, offset)
+        chunk = data[offset + 8 : offset + 8 + length]
+        if kind == 0x4E4F534A:
+            document = json.loads(chunk)
+        elif kind == 0x004E4942:
+            binary = chunk
+        offset += 8 + length
+    if document is None or binary is None:
+        raise ValueError("Missing mesh buffers")
+
+    def accessor(index):
+        spec = document["accessors"][index]
+        view = document["bufferViews"][spec["bufferView"]]
+        if "sparse" in spec or view.get("buffer", 0) != 0:
+            raise ValueError("Unsupported sparse/external collision buffer")
+        dtype = {5123: "<u2", 5125: "<u4", 5126: "<f4"}[spec["componentType"]]
+        width = {"SCALAR": 1, "VEC3": 3}[spec["type"]]
+        item = np.dtype(dtype).itemsize
+        start = view.get("byteOffset", 0) + spec.get("byteOffset", 0)
+        return np.ndarray(
+            (spec["count"], width),
+            dtype=dtype,
+            buffer=binary,
+            offset=start,
+            strides=(view.get("byteStride", width * item), item),
+        ).copy()
+
+    vertices, faces, base = [], [], 0
+    # SplatTransform collision exports bake coordinates; reject transforms so
+    # we never silently misregister contact geometry against the Gaussian view.
+    for node in document.get("nodes", []):
+        if any(k in node for k in ("matrix", "translation", "rotation", "scale")):
+            raise ValueError("Collision node transform must be baked")
+    for mesh in document["meshes"]:
+        for primitive in mesh["primitives"]:
+            if primitive.get("mode", 4) != 4:
+                raise ValueError("Collision mesh must contain triangles")
+            points = accessor(primitive["attributes"]["POSITION"])
+            triangles = accessor(primitive["indices"]).reshape(-1, 3).astype(np.int32)
+            if not np.isfinite(points).all() or triangles.max() >= len(points):
+                raise ValueError("Invalid collision vertices")
+            vertices.append(points[:, [0, 2, 1]] * np.array([1, -1, 1], np.float32))
+            faces.append(triangles + base)
+            base += len(points)
+    return np.concatenate(vertices), np.concatenate(faces)
+
+
+def ground_intersections(vertices, faces, x, y):
+    """Vertical mesh intersections, used only to place the initial rigid body."""
+    triangles = vertices[faces]
+    low, high = triangles[:, :, :2].min(axis=1), triangles[:, :, :2].max(axis=1)
+    candidate = triangles[
+        (low[:, 0] <= x) & (high[:, 0] >= x) & (low[:, 1] <= y) & (high[:, 1] >= y)
+    ]
+    if not len(candidate):
+        return np.empty(0)
+    a, b, c = candidate[:, 0], candidate[:, 1], candidate[:, 2]
+    den = (b[:, 1] - c[:, 1]) * (a[:, 0] - c[:, 0]) + (c[:, 0] - b[:, 0]) * (a[:, 1] - c[:, 1])
+    valid = np.abs(den) > 1e-8
+    a, b, c, den = a[valid], b[valid], c[valid], den[valid]
+    u = ((b[:, 1] - c[:, 1]) * (x - c[:, 0]) + (c[:, 0] - b[:, 0]) * (y - c[:, 1])) / den
+    v = ((c[:, 1] - a[:, 1]) * (x - c[:, 0]) + (a[:, 0] - c[:, 0]) * (y - c[:, 1])) / den
+    inside = (u >= -1e-6) & (v >= -1e-6) & (u + v <= 1 + 1e-6)
+    return (u * a[:, 2] + v * b[:, 2] + (1 - u - v) * c[:, 2])[inside]
+
+
+def install_terrain(stage, manifest_path, world):
+    from pxr import UsdGeom, UsdPhysics, UsdShade
+    from rover_profile import PROFILE
+
+    from k1link.simulation.ai_polygon.terrain_contract import terrain_matches
+
+    manifest = json.loads(Path(manifest_path).read_text(encoding="utf-8-sig"))
+    if not terrain_matches(manifest, world):
+        raise ValueError("Terrain does not match the world calibration")
+    path = Path(manifest["collider"])
+    if hashlib.sha256(path.read_bytes()).hexdigest() != manifest["collider_sha256"]:
+        raise ValueError("Collision mesh identity changed")
+    points, triangles = load_glb(path)
+    settings = world["settings"]
+    heights = ground_intersections(points, triangles, *settings["spawn_xy"])
+    candidates = heights[np.abs(heights - settings["ground_z"]) < 0.8]
+    if not len(candidates):
+        raise ValueError("No reconstructed ground at the configured spawn")
+    # Place the whole footprint above nearby ground, without embedding a wheel
+    # in a stone beside the centre ray. This is setup, never a planner input.
+    angle = math.radians(settings["heading_degrees"])
+    support, offsets = [], []
+    for dx in (-0.5, 0, 0.5):
+        for dy in (-0.5, 0, 0.5):
+            x = settings["spawn_xy"][0] + dx * math.cos(angle) - dy * math.sin(angle)
+            y = settings["spawn_xy"][1] + dx * math.sin(angle) + dy * math.cos(angle)
+            intersections = ground_intersections(points, triangles, x, y)
+            nearby = intersections[np.abs(intersections - settings["ground_z"]) < 0.8]
+            if not len(nearby):
+                raise ValueError("Rover footprint has no reconstructed support")
+            support.append(float(nearby.max()))
+            offsets.append([x - settings["spawn_xy"][0], y - settings["spawn_xy"][1], 1])
+    plane = np.linalg.lstsq(np.asarray(offsets), np.asarray(support), rcond=None)[0]
+    residual = np.asarray(support) - np.asarray(offsets) @ plane
+    normal = np.array([-plane[0], -plane[1], 1.0])
+    normal /= np.linalg.norm(normal)
+    if np.max(np.abs(residual)) > 0.12 or normal[2] < math.cos(math.radians(25)):
+        raise ValueError("Configured spawn is too uneven for the metre-wide rover")
+    from spawn_clearance import obstructing_triangles
+
+    if obstructing_triangles(
+        points,
+        triangles,
+        settings["spawn_xy"],
+        settings["heading_degrees"],
+        plane,
+        step=PROFILE["max_step_m"],
+    ):
+        raise ValueError("Configured spawn contains an obstacle inside the rover footprint")
+    # Align to the local support plane before gravity settles the suspensionless
+    # lab chassis. Wheel centres start 4 cm above the highest residual contact.
+    wheel_center_z = float(
+        plane[2] + (PROFILE["wheel_radius_m"] + 0.04) / normal[2] + max(0, residual.max())
+    )
+    mesh = UsdGeom.Mesh.Define(stage, "/World/Terrain")
+    mesh.CreatePointsAttr(points.tolist())
+    mesh.CreateFaceVertexCountsAttr([3] * len(triangles))
+    mesh.CreateFaceVertexIndicesAttr(triangles.reshape(-1).tolist())
+    mesh.CreateSubdivisionSchemeAttr(UsdGeom.Tokens.none)
+    mesh.CreateDoubleSidedAttr(True)
+    mesh.CreateVisibilityAttr(UsdGeom.Tokens.invisible)
+    UsdPhysics.CollisionAPI.Apply(mesh.GetPrim())
+    UsdPhysics.MeshCollisionAPI.Apply(mesh.GetPrim()).CreateApproximationAttr("none")
+    material = UsdShade.Material.Define(stage, "/World/Materials/Terrain")
+    surface = UsdPhysics.MaterialAPI.Apply(material.GetPrim())
+    surface.CreateStaticFrictionAttr(0.9)
+    surface.CreateDynamicFrictionAttr(0.8)
+    surface.CreateRestitutionAttr(0)
+    UsdShade.MaterialBindingAPI.Apply(mesh.GetPrim()).Bind(
+        material, UsdShade.Tokens.weakerThanDescendants, "physics"
+    )
+    return wheel_center_z, dict(
+        manifest,
+        vertex_count=len(points),
+        triangle_count=len(triangles),
+        initial_contact_z=float(plane[2]),
+        initial_ground_normal=normal.tolist(),
+        support_residual_m=float(np.max(np.abs(residual))),
+    )
+
+
+class RangeSensor:
+    """Front 360-degree LiDAR plus rear near-field fan; first physical hit.
+
+    Separate mounts observe the ground beyond each bumper without seeing
+    through the chassis. The rear fan adds 660 rays to the 2160 front rays.
+    """
+
+    def __init__(self):
+        import omni.physx
+
+        self.query = omni.physx.get_physx_scene_query_interface()
+        elevations = (-80, -75, -70, -60, -45, -35, -28, -22, -18, -14, -10, -5, 0, 5, 15)
+        self.directions = [
+            (math.cos(e) * math.cos(a), math.cos(e) * math.sin(a), math.sin(e))
+            for e in np.radians(elevations)
+            for a in np.radians(np.arange(-180, 180, 2.5))
+        ]
+        self.rear_directions = [
+            (math.cos(e) * math.cos(a), math.cos(e) * math.sin(a), math.sin(e))
+            for e in np.radians(np.arange(-80, 16, 5))
+            for a in np.radians(np.arange(100, 261, 5))
+        ]
+
+    def capture(self, origin, rotation, rear_origin=None):
+        from pxr import Gf
+
+        points = []
+        mounts = [(origin, self.directions)]
+        if rear_origin is not None:
+            mounts.append((rear_origin, self.rear_directions))
+        for mount, directions in mounts:
+            for direction in directions:
+                ray = rotation.TransformDir(Gf.Vec3d(*direction)).GetNormalized()
+                hit = self.query.raycast_closest(tuple(mount), tuple(ray), 8.0, bothSides=True)
+                if hit["hit"] and not str(hit.get("rigidBody", "")).startswith("/World/Rover"):
+                    points.append(hit["position"])
+        return np.asarray(points, dtype=np.float32).reshape(-1, 3)
diff --git a/simulation/ai-polygon/worker.py b/simulation/ai-polygon/worker.py
new file mode 100644
index 0000000..2ab7bb7
--- /dev/null
+++ b/simulation/ai-polygon/worker.py
@@ -0,0 +1,212 @@
+"""Single-owner Windows coordinator; no GPU work while idle.
+
+Core owns admission. Each episode gets a fresh native Isaac process and the
+additive model stack. A missing Core heartbeat terminates this episode; it never
+resumes an old command. A leftover active.json requires explicit reconciliation.
+"""
+
+import argparse
+import hashlib
+import json
+import os
+import re
+import subprocess
+import threading
+import time
+from pathlib import Path
+from uuid import uuid4
+
+from core_client import CoreClient
+from model_stack import ROOT, ModelStack, sha256
+
+
+def terminate_episode(child):
+    """A concurrent normal exit is success; only a still-running child is failure."""
+    if child.poll() is None:
+        subprocess.run(
+            ["taskkill", "/PID", str(child.pid), "/T", "/F"],
+            capture_output=True,
+            check=False,
+            timeout=30,
+        )
+    child.wait(timeout=30)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--core", default="http://127.0.0.1:18081")
+    parser.add_argument("--token-file", type=Path, required=True)
+    parser.add_argument("--state", type=Path, required=True)
+    parser.add_argument("--isaac", type=Path, required=True)
+    parser.add_argument("--once", action="store_true")
+    args = parser.parse_args()
+    args.state.mkdir(parents=True, exist_ok=True)
+    active_file = args.state / "active.json"
+    if active_file.exists():
+        raise RuntimeError("Reconcile the previous simulation before starting a new worker")
+    lock = args.state / "worker.lock"
+    instance = uuid4().hex
+    client = CoreClient(args.core, args.token_file, instance)
+    stack = ModelStack()
+    stack.preflight()
+    robot_root = args.isaac.parent / "assets/jetbot-6.1-v1"
+    robot_manifest = robot_root / "asset-manifest.json"
+    assets = json.loads(robot_manifest.read_text(encoding="utf-8-sig"))
+    for asset in assets["files"]:
+        if sha256(robot_root / asset["path"]) != asset["sha256"]:
+            raise RuntimeError("Prepared Jetbot asset changed")
+    sources = {
+        "worker": sha256(ROOT / "worker.py"),
+        "scene": sha256(ROOT / "run_scene.py"),
+        "models": sha256(stack.profile_path),
+        "robot": sha256(robot_manifest),
+    }
+    hello = {
+        "worker_id": "worker-006-ai-polygon",
+        "instance_id": instance,
+        "runtime": "isaac-sim-6.1",
+        "model_ids": [m["id"] for m in stack.profile["models"]],
+        "profile_sha256": hashlib.sha256(json.dumps(sources, sort_keys=True).encode()).hexdigest(),
+        "runtime_sources": sources,
+    }
+    lost = threading.Event()
+    finished = threading.Event()
+    child = None
+
+    def heartbeat():
+        while not finished.wait(3):
+            try:
+                client.request("/worker/heartbeat", {"instance_id": instance})
+            except Exception:
+                lost.set()
+                return
+
+    fd = os.open(lock, os.O_WRONLY | os.O_CREAT | os.O_EXCL)
+    os.close(fd)
+    try:
+        client.request("/worker/register", hello)
+        threading.Thread(target=heartbeat, daemon=True).start()
+        while not lost.is_set():
+            polled = client.request("/worker/poll", {"instance_id": instance})
+            if polled["action"] == "idle":
+                time.sleep(0.5)
+                continue
+            run = polled["run"]
+            if not re.fullmatch(r"airun-[a-f0-9]{32}", run["run_id"]):
+                raise ValueError("Invalid run identity")
+            episode = args.state / run["run_id"]
+            episode.mkdir()
+            run_file = episode / "run.json"
+            run_file.write_text(json.dumps(run))
+            active_file.write_text(
+                json.dumps(
+                    {
+                        "run_id": run["run_id"],
+                        "instance_id": instance,
+                        "profile_sha256": hello["profile_sha256"],
+                    }
+                )
+            )
+            result = {"outcome": "failed", "message": "Worker прервал прогон."}
+            released = False
+            try:
+                cache = args.state / "worlds"
+                cache.mkdir(exist_ok=True)
+                source = cache / (run["world"]["sha256"] + ".ply")
+                client.download(run["world"], source)
+                if lost.is_set():
+                    raise RuntimeError("Core connection was lost during scene preparation")
+                control = client.request(
+                    "/worker/runs/" + run["run_id"] + "/progress", {"phase": "models"}
+                )
+                if control["control"] == "stop":
+                    raise InterruptedError("Stopped before model startup")
+                stack.start(
+                    cancelled=lambda run_id=run["run_id"]: (
+                        lost.is_set()
+                        or client.request("/runs/" + run_id)["control"] == "stop"
+                    )
+                )
+                if lost.is_set():
+                    raise RuntimeError("Core connection was lost during model startup")
+                control = client.request(
+                    "/worker/runs/" + run["run_id"] + "/progress", {"phase": "scene"}
+                )
+                if control["control"] == "stop":
+                    raise InterruptedError("Stopped before scene startup")
+                command = [
+                    str(args.isaac / "python.bat"),
+                    str(ROOT / "run_scene.py"),
+                    "--run",
+                    str(run_file),
+                    "--source",
+                    str(source),
+                    "--core",
+                    args.core,
+                    "--token-file",
+                    str(args.token_file),
+                    "--instance",
+                    instance,
+                ]
+                with (episode / "isaac.log").open("wb") as output:
+                    child = subprocess.Popen(command, stdout=output, stderr=subprocess.STDOUT)
+                    stopped = False
+                    next_control_check = 0.0
+                    deadline = time.monotonic() + 300 + run["request"]["max_steps"] * 15
+                    while (
+                        child.poll() is None and not lost.is_set() and time.monotonic() < deadline
+                    ):
+                        if time.monotonic() >= next_control_check:
+                            state = client.request("/runs/" + run["run_id"])
+                            if state["control"] == "stop" or state["state"] == "failed":
+                                stopped = state["control"] == "stop"
+                                break
+                            next_control_check = time.monotonic() + 1
+                        time.sleep(0.2)
+                    if child.poll() is None:
+                        terminate_episode(child)
+                    child.wait(timeout=30)
+                    child = None
+                result_file = episode / "result.json"
+                if stopped:
+                    result = {"outcome": "stopped", "message": "Прогон остановлен."}
+                elif result_file.exists() and not lost.is_set():
+                    result = json.loads(result_file.read_text())
+            except Exception as exc:
+                (episode / "worker-error.txt").write_text(type(exc).__name__ + ": " + str(exc))
+            finally:
+                if child is not None and child.poll() is None:
+                    terminate_episode(child)
+                    child.wait(timeout=30)
+                stack.stop()
+                released = True
+            # A Stop may race with the final sample/normal native exit. It remains a
+            # successful cancellation only after native and model cleanup above.
+            if not lost.is_set():
+                state = client.request("/runs/" + run["run_id"])
+                if state["control"] == "stop" and state["state"] == "stopping":
+                    result = {"outcome": "stopped", "message": "Движение и inference остановлены."}
+            # Do not release the reservation until native process AND GPU containers are gone.
+            for attempt in range(10):
+                try:
+                    if lost.is_set():
+                        client.request("/worker/register", hello)
+                    client.request(
+                        "/worker/runs/" + run["run_id"] + "/finish",
+                        {**result, "instance_id": instance, "resources_released": released},
+                    )
+                    active_file.unlink()
+                    break
+                except Exception:
+                    if attempt == 9:
+                        raise
+                    time.sleep(1)
+            if args.once or lost.is_set():
+                break
+    finally:
+        finished.set()
+        lock.unlink(missing_ok=True)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/src/k1link/artifacts.py b/src/k1link/artifacts.py
index 3a9dab6..9a2e0ce 100644
--- a/src/k1link/artifacts.py
+++ b/src/k1link/artifacts.py
@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import hashlib
 import json
 import os
 import tempfile
@@ -8,6 +9,14 @@ from pathlib import Path
 from typing import Any
 
 
+def canonical_json_sha256(value: object) -> str:
+    """Platform-independent identity for shared Core/Worker JSON contracts."""
+    payload = json.dumps(
+        value, ensure_ascii=False, allow_nan=False, separators=(",", ":"), sort_keys=True
+    ).encode("utf-8")
+    return hashlib.sha256(payload).hexdigest()
+
+
 def utc_now_iso() -> str:
     """Return a stable UTC timestamp for manifests and capture artifacts."""
     return datetime.now(UTC).isoformat(timespec="milliseconds").replace("+00:00", "Z")
diff --git a/src/k1link/observatory/__init__.py b/src/k1link/observatory/__init__.py
index 42211f8..766c2dc 100644
--- a/src/k1link/observatory/__init__.py
+++ b/src/k1link/observatory/__init__.py
@@ -1,33 +1,41 @@
-"""Observation-only laboratory orchestration contracts."""
+"""Observation contracts; platform-specific services are loaded only on demand.
 
-from k1link.observatory.canonical_result import (
-    is_admitted_observatory_recorded_result,
-)
-from k1link.observatory.run_preparations import (
-    MAX_RUN_PREPARATION_RECORDS,
-    MAX_RUN_PREPARATION_STORAGE_BYTES,
-    OBSERVATORY_RUN_PREPARATION_REQUEST_SCHEMA,
-    OBSERVATORY_RUN_PREPARATION_SCHEMA,
-    RUN_PREPARATION_DATABASE_NAME,
-    RUN_PREPARATION_SQLITE_LOCK_TIMEOUT_SECONDS,
-    ObservatoryRunPreparation,
-    ObservatoryRunPreparationCapacityError,
-    ObservatoryRunPreparationConflictError,
-    ObservatoryRunPreparationError,
-    ObservatoryRunPreparationIntegrityError,
-    ObservatoryRunPreparationIntent,
-    ObservatoryRunPreparationLedger,
-    ObservatoryRunPreparationNotFoundError,
-    load_observatory_run_preparation_ledger,
-    observatory_run_preparation_request_sha256,
-)
-from k1link.observatory.setups import (
-    LABORATORY_SETUP_CATALOG_SCHEMA,
-    LABORATORY_SETUP_REGISTRY_SCHEMA,
-    OBSERVATORY_CALCULATION_PROFILE_SCHEMA,
-    LaboratorySetupRegistry,
-    LaboratorySetupRegistryError,
-)
+Importing a portable composition on Windows must not import Core's POSIX
+artifact gateway or eagerly initialize session/recording infrastructure.
+"""
+
+from importlib import import_module
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from k1link.observatory.canonical_result import (
+        is_admitted_observatory_recorded_result,
+    )
+    from k1link.observatory.run_preparations import (
+        MAX_RUN_PREPARATION_RECORDS,
+        MAX_RUN_PREPARATION_STORAGE_BYTES,
+        OBSERVATORY_RUN_PREPARATION_REQUEST_SCHEMA,
+        OBSERVATORY_RUN_PREPARATION_SCHEMA,
+        RUN_PREPARATION_DATABASE_NAME,
+        RUN_PREPARATION_SQLITE_LOCK_TIMEOUT_SECONDS,
+        ObservatoryRunPreparation,
+        ObservatoryRunPreparationCapacityError,
+        ObservatoryRunPreparationConflictError,
+        ObservatoryRunPreparationError,
+        ObservatoryRunPreparationIntegrityError,
+        ObservatoryRunPreparationIntent,
+        ObservatoryRunPreparationLedger,
+        ObservatoryRunPreparationNotFoundError,
+        load_observatory_run_preparation_ledger,
+        observatory_run_preparation_request_sha256,
+    )
+    from k1link.observatory.setups import (
+        LABORATORY_SETUP_CATALOG_SCHEMA,
+        LABORATORY_SETUP_REGISTRY_SCHEMA,
+        OBSERVATORY_CALCULATION_PROFILE_SCHEMA,
+        LaboratorySetupRegistry,
+        LaboratorySetupRegistryError,
+    )
 
 __all__ = [
     "LABORATORY_SETUP_CATALOG_SCHEMA",
@@ -53,3 +61,41 @@ __all__ = [
     "load_observatory_run_preparation_ledger",
     "observatory_run_preparation_request_sha256",
 ]
+
+_EXPORTS = {
+    "is_admitted_observatory_recorded_result": "k1link.observatory.canonical_result",
+    "MAX_RUN_PREPARATION_RECORDS": "k1link.observatory.run_preparations",
+    "MAX_RUN_PREPARATION_STORAGE_BYTES": "k1link.observatory.run_preparations",
+    "OBSERVATORY_RUN_PREPARATION_REQUEST_SCHEMA": "k1link.observatory.run_preparations",
+    "OBSERVATORY_RUN_PREPARATION_SCHEMA": "k1link.observatory.run_preparations",
+    "RUN_PREPARATION_DATABASE_NAME": "k1link.observatory.run_preparations",
+    "RUN_PREPARATION_SQLITE_LOCK_TIMEOUT_SECONDS": "k1link.observatory.run_preparations",
+    "ObservatoryRunPreparation": "k1link.observatory.run_preparations",
+    "ObservatoryRunPreparationCapacityError": "k1link.observatory.run_preparations",
+    "ObservatoryRunPreparationConflictError": "k1link.observatory.run_preparations",
+    "ObservatoryRunPreparationError": "k1link.observatory.run_preparations",
+    "ObservatoryRunPreparationIntegrityError": "k1link.observatory.run_preparations",
+    "ObservatoryRunPreparationIntent": "k1link.observatory.run_preparations",
+    "ObservatoryRunPreparationLedger": "k1link.observatory.run_preparations",
+    "ObservatoryRunPreparationNotFoundError": "k1link.observatory.run_preparations",
+    "load_observatory_run_preparation_ledger": "k1link.observatory.run_preparations",
+    "observatory_run_preparation_request_sha256": "k1link.observatory.run_preparations",
+    "LABORATORY_SETUP_CATALOG_SCHEMA": "k1link.observatory.setups",
+    "LABORATORY_SETUP_REGISTRY_SCHEMA": "k1link.observatory.setups",
+    "OBSERVATORY_CALCULATION_PROFILE_SCHEMA": "k1link.observatory.setups",
+    "LaboratorySetupRegistry": "k1link.observatory.setups",
+    "LaboratorySetupRegistryError": "k1link.observatory.setups",
+}
+
+
+def __getattr__(name: str):
+    module = _EXPORTS.get(name)
+    if module is None:
+        raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+    value = getattr(import_module(module), name)
+    globals()[name] = value
+    return value
+
+
+def __dir__():
+    return sorted(set(globals()) | set(__all__))
diff --git a/src/k1link/observatory/modular_composition.py b/src/k1link/observatory/modular_composition.py
index b71465a..1241211 100644
--- a/src/k1link/observatory/modular_composition.py
+++ b/src/k1link/observatory/modular_composition.py
@@ -12,7 +12,7 @@ from dataclasses import dataclass
 from pathlib import Path
 from typing import Final, cast
 
-from k1link.observatory.portable_run_definitions import canonical_sha256
+from k1link.artifacts import canonical_json_sha256 as canonical_sha256
 
 COMPOSITION_SCHEMA: Final = "missioncore.observatory-ai-composition/v1"
 MODULE_SCHEMA: Final = "missioncore.observatory-ai-module/v1"
@@ -179,6 +179,11 @@ class CompositionSpec:
     """Source-independent graph, in deterministic topological execution order."""
 
     nodes: tuple[CompositionNode, ...]
+    execution_mode: str = "recorded-observation-only"
+
+    def __post_init__(self) -> None:
+        if self.execution_mode not in ("recorded-observation-only", "worker-local-simulation"):
+            raise CompositionError("unsupported composition execution mode")
 
     @property
     def source_capabilities(self) -> tuple[str, ...]:
@@ -212,7 +217,10 @@ class CompositionSpec:
             "nodes": [node.as_dict() for node in self.nodes],
             "source_capabilities": list(self.source_capabilities),
             "outputs": list(self.outputs),
-            "execution": {"max_parallel_nodes": 1, "mode": "recorded-observation-only"},
+            "execution": {
+                "max_parallel_nodes": 2 if self.execution_mode == "worker-local-simulation" else 1,
+                "mode": self.execution_mode,
+            },
         }
 
     @property
@@ -286,7 +294,9 @@ class ModuleRegistry:
             ],
         }
 
-    def compose(self, document: object) -> CompositionSpec:
+    def compose(
+        self, document: object, *, execution_mode: str = "recorded-observation-only"
+    ) -> CompositionSpec:
         root = _object(document, {"schema_version", "selections"})
         if root["schema_version"] != COMPOSITION_SCHEMA:
             raise CompositionError("unsupported composition schema")
@@ -381,7 +391,7 @@ class ModuleRegistry:
             for key in ready:
                 ordered.append(pending.pop(key))
                 emitted.add(key)
-        return CompositionSpec(tuple(ordered))
+        return CompositionSpec(tuple(ordered), execution_mode=execution_mode)
 
 
 def node_input_identity(
diff --git a/src/k1link/observatory/recorded_jobs.py b/src/k1link/observatory/recorded_jobs.py
index b8a3fbb..c863b3a 100644
--- a/src/k1link/observatory/recorded_jobs.py
+++ b/src/k1link/observatory/recorded_jobs.py
@@ -97,6 +97,11 @@ _AUTHORITY: Final = {
 }
 
 _SCHEMA_SQL = """
+CREATE TABLE IF NOT EXISTS simulation_worker_reservation (
+    singleton INTEGER PRIMARY KEY CHECK (singleton = 1),
+    owner_id TEXT NOT NULL,
+    created_at_utc TEXT NOT NULL
+);
 CREATE TABLE IF NOT EXISTS observatory_recorded_jobs (
     job_id TEXT PRIMARY KEY,
     idempotency_key TEXT NOT NULL UNIQUE,
@@ -1185,6 +1190,46 @@ class ObservatoryRecordedJobQueue:
         self._lock = threading.RLock()
         self._initialize()
 
+    def reserve_simulation(self, owner_id: str) -> None:
+        """Reserve this Worker's GPU atomically against recorded/live admission.
+
+        A lost heartbeat does not prove GPU release. This reservation survives
+        restarts and is removed only by the exact simulator's release receipt.
+        """
+        if not re.fullmatch(r"airun-[a-f0-9]{32}", owner_id):
+            raise ValueError("invalid simulation resource owner")
+        with self._transaction() as connection:
+            existing = connection.execute(
+                "SELECT owner_id FROM simulation_worker_reservation"
+            ).fetchone()
+            if existing is not None:
+                if existing["owner_id"] == owner_id:
+                    return
+                raise ObservatoryRecordedQueueBusyError("Worker занят другим прогоном симуляции.")
+            busy = connection.execute(
+                "SELECT job_id FROM observatory_recorded_jobs "
+                "WHERE state IN ('claimed', 'running', 'paused', 'preemption-pending', "
+                "'reconciliation-required') LIMIT 1"
+            ).fetchone()
+            if busy is not None or self._open_live_lease_row(connection) is not None:
+                raise ObservatoryRecordedQueueBusyError("Worker занят задачей AI Inference.")
+            connection.execute(
+                "INSERT INTO simulation_worker_reservation VALUES (1, ?, ?)",
+                (owner_id, self._timestamp()),
+            )
+
+    def release_simulation(self, owner_id: str) -> None:
+        """Called only after the trusted simulator reports all GPU work stopped."""
+        with self._transaction() as connection:
+            existing = connection.execute(
+                "SELECT owner_id FROM simulation_worker_reservation"
+            ).fetchone()
+            if existing is not None and existing["owner_id"] != owner_id:
+                raise ObservatoryRecordedQueueConflictError("simulation resource owner changed")
+            connection.execute(
+                "DELETE FROM simulation_worker_reservation WHERE owner_id = ?", (owner_id,)
+            )
+
     def resolve_definition(
         self,
         setup_id: str,
@@ -1444,7 +1489,10 @@ class ObservatoryRecordedJobQueue:
                     label="legacy claim receipt",
                 )
             row = None
-            if self._open_live_lease_row(connection) is None:
+            simulation = connection.execute(
+                "SELECT owner_id FROM simulation_worker_reservation"
+            ).fetchone()
+            if self._open_live_lease_row(connection) is None and simulation is None:
                 active_owner = connection.execute(
                     "SELECT job_id FROM observatory_recorded_jobs "
                     "WHERE state IN ('claimed', 'running', 'preemption-pending', "
@@ -2382,6 +2430,10 @@ class ObservatoryRecordedJobQueue:
         _validate_pattern(lease_id, _LEASE_ID, "live lease id")
         self.recover_stale_claims()
         with self._transaction() as connection:
+            if connection.execute("SELECT owner_id FROM simulation_worker_reservation").fetchone():
+                raise ObservatoryRecordedQueueBusyError(
+                    "simulation has not released the Worker GPU"
+                )
             lease = self._get_live_lease(connection, lease_id)
             if lease.state == "active":
                 return lease
diff --git a/src/k1link/simulation/__init__.py b/src/k1link/simulation/__init__.py
index 328c93b..516c319 100644
--- a/src/k1link/simulation/__init__.py
+++ b/src/k1link/simulation/__init__.py
@@ -1,78 +1,86 @@
-"""Mission Core qualification and simulation boundaries."""
+"""Mission Core simulation contracts, with platform services loaded on demand.
 
-from k1link.simulation.contracts import (
-    AckermannControlSetpoint,
-    AuthorityProfile,
-    CommandAuthorityScope,
-    ControlProfile,
-    ControlSetpoint,
-    DifferentialControlSetpoint,
-    ProviderPin,
-    QualificationArtifact,
-    QualificationEvent,
-    QualificationRun,
-    ReproducibilityTier,
-    RunKind,
-    RunState,
-    SimulationContractError,
-)
-from k1link.simulation.orchestrator import (
-    ActiveQualificationRunError,
-    SimulationApplicationService,
-    SimulationOrchestratorError,
-    SimulationWorkerPort,
-    WorkerStartResult,
-    WorkerStopResult,
-)
-from k1link.simulation.process_supervisor import (
-    OwnedProcess,
-    PosixProcessSupervisor,
-    ProcessSpec,
-    ProcessStopResult,
-    ProcessSupervisorError,
-)
-from k1link.simulation.provider_contract import (
-    PROVIDER_PROFILE_SCHEMA,
-    ProviderRole,
-    SimulationClockDescriptor,
-    SimulationProviderContractError,
-    SimulationProviderDescriptor,
-    SimulationProviderProfile,
-)
-from k1link.simulation.run_store import (
-    QualificationRunConflictError,
-    QualificationRunIntegrityError,
-    QualificationRunNotFoundError,
-    QualificationRunStore,
-    QualificationRunStoreError,
-    QualificationRunTransitionError,
-)
-from k1link.simulation.s0 import (
-    CheckStatus,
-    DoctorVerdict,
-    RuntimeAcceptance,
-    S0DoctorReport,
-    S0Profile,
-    S0ProfileError,
-    load_s0_profile,
-    run_s0_doctor,
-)
-from k1link.simulation.stock_rover import (
-    LIFECYCLE_PROFILE_SCHEMA,
-    StockRoverLifecycleProfile,
-    StockRoverProfileError,
-    StockRoverTargetPaths,
-    load_stock_rover_lifecycle_profile,
-    stock_rover_process_environment,
-    stock_rover_process_specs,
-)
-from k1link.simulation.worker import (
-    LocalProcessWorkerAdapter,
-    S0WorkerGuard,
-    SimulationWorldControl,
-    WorkerAdmission,
-    WorkerAdmissionError,
-)
+Portable AI compositions do not require the legacy S0/YAML or POSIX process
+supervisor when loaded by the Windows Worker coordinator.
+"""
+
+from importlib import import_module
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from k1link.simulation.contracts import (
+        AckermannControlSetpoint,
+        AuthorityProfile,
+        CommandAuthorityScope,
+        ControlProfile,
+        ControlSetpoint,
+        DifferentialControlSetpoint,
+        ProviderPin,
+        QualificationArtifact,
+        QualificationEvent,
+        QualificationRun,
+        ReproducibilityTier,
+        RunKind,
+        RunState,
+        SimulationContractError,
+    )
+    from k1link.simulation.orchestrator import (
+        ActiveQualificationRunError,
+        SimulationApplicationService,
+        SimulationOrchestratorError,
+        SimulationWorkerPort,
+        WorkerStartResult,
+        WorkerStopResult,
+    )
+    from k1link.simulation.process_supervisor import (
+        OwnedProcess,
+        PosixProcessSupervisor,
+        ProcessSpec,
+        ProcessStopResult,
+        ProcessSupervisorError,
+    )
+    from k1link.simulation.provider_contract import (
+        PROVIDER_PROFILE_SCHEMA,
+        ProviderRole,
+        SimulationClockDescriptor,
+        SimulationProviderContractError,
+        SimulationProviderDescriptor,
+        SimulationProviderProfile,
+    )
+    from k1link.simulation.run_store import (
+        QualificationRunConflictError,
+        QualificationRunIntegrityError,
+        QualificationRunNotFoundError,
+        QualificationRunStore,
+        QualificationRunStoreError,
+        QualificationRunTransitionError,
+    )
+    from k1link.simulation.s0 import (
+        CheckStatus,
+        DoctorVerdict,
+        RuntimeAcceptance,
+        S0DoctorReport,
+        S0Profile,
+        S0ProfileError,
+        load_s0_profile,
+        run_s0_doctor,
+    )
+    from k1link.simulation.stock_rover import (
+        LIFECYCLE_PROFILE_SCHEMA,
+        StockRoverLifecycleProfile,
+        StockRoverProfileError,
+        StockRoverTargetPaths,
+        load_stock_rover_lifecycle_profile,
+        stock_rover_process_environment,
+        stock_rover_process_specs,
+    )
+    from k1link.simulation.worker import (
+        LocalProcessWorkerAdapter,
+        S0WorkerGuard,
+        SimulationWorldControl,
+        WorkerAdmission,
+        WorkerAdmissionError,
+    )
 
 __all__ = [
     "AckermannControlSetpoint",
@@ -133,3 +141,76 @@ __all__ = [
     "stock_rover_process_environment",
     "stock_rover_process_specs",
 ]
+
+_EXPORTS = {
+    "AckermannControlSetpoint": "k1link.simulation.contracts",
+    "AuthorityProfile": "k1link.simulation.contracts",
+    "CommandAuthorityScope": "k1link.simulation.contracts",
+    "ControlProfile": "k1link.simulation.contracts",
+    "ControlSetpoint": "k1link.simulation.contracts",
+    "DifferentialControlSetpoint": "k1link.simulation.contracts",
+    "ProviderPin": "k1link.simulation.contracts",
+    "QualificationArtifact": "k1link.simulation.contracts",
+    "QualificationEvent": "k1link.simulation.contracts",
+    "QualificationRun": "k1link.simulation.contracts",
+    "ReproducibilityTier": "k1link.simulation.contracts",
+    "RunKind": "k1link.simulation.contracts",
+    "RunState": "k1link.simulation.contracts",
+    "SimulationContractError": "k1link.simulation.contracts",
+    "ActiveQualificationRunError": "k1link.simulation.orchestrator",
+    "SimulationApplicationService": "k1link.simulation.orchestrator",
+    "SimulationOrchestratorError": "k1link.simulation.orchestrator",
+    "SimulationWorkerPort": "k1link.simulation.orchestrator",
+    "WorkerStartResult": "k1link.simulation.orchestrator",
+    "WorkerStopResult": "k1link.simulation.orchestrator",
+    "OwnedProcess": "k1link.simulation.process_supervisor",
+    "PosixProcessSupervisor": "k1link.simulation.process_supervisor",
+    "ProcessSpec": "k1link.simulation.process_supervisor",
+    "ProcessStopResult": "k1link.simulation.process_supervisor",
+    "ProcessSupervisorError": "k1link.simulation.process_supervisor",
+    "PROVIDER_PROFILE_SCHEMA": "k1link.simulation.provider_contract",
+    "ProviderRole": "k1link.simulation.provider_contract",
+    "SimulationClockDescriptor": "k1link.simulation.provider_contract",
+    "SimulationProviderContractError": "k1link.simulation.provider_contract",
+    "SimulationProviderDescriptor": "k1link.simulation.provider_contract",
+    "SimulationProviderProfile": "k1link.simulation.provider_contract",
+    "QualificationRunConflictError": "k1link.simulation.run_store",
+    "QualificationRunIntegrityError": "k1link.simulation.run_store",
+    "QualificationRunNotFoundError": "k1link.simulation.run_store",
+    "QualificationRunStore": "k1link.simulation.run_store",
+    "QualificationRunStoreError": "k1link.simulation.run_store",
+    "QualificationRunTransitionError": "k1link.simulation.run_store",
+    "CheckStatus": "k1link.simulation.s0",
+    "DoctorVerdict": "k1link.simulation.s0",
+    "RuntimeAcceptance": "k1link.simulation.s0",
+    "S0DoctorReport": "k1link.simulation.s0",
+    "S0Profile": "k1link.simulation.s0",
+    "S0ProfileError": "k1link.simulation.s0",
+    "load_s0_profile": "k1link.simulation.s0",
+    "run_s0_doctor": "k1link.simulation.s0",
+    "LIFECYCLE_PROFILE_SCHEMA": "k1link.simulation.stock_rover",
+    "StockRoverLifecycleProfile": "k1link.simulation.stock_rover",
+    "StockRoverProfileError": "k1link.simulation.stock_rover",
+    "StockRoverTargetPaths": "k1link.simulation.stock_rover",
+    "load_stock_rover_lifecycle_profile": "k1link.simulation.stock_rover",
+    "stock_rover_process_environment": "k1link.simulation.stock_rover",
+    "stock_rover_process_specs": "k1link.simulation.stock_rover",
+    "LocalProcessWorkerAdapter": "k1link.simulation.worker",
+    "S0WorkerGuard": "k1link.simulation.worker",
+    "SimulationWorldControl": "k1link.simulation.worker",
+    "WorkerAdmission": "k1link.simulation.worker",
+    "WorkerAdmissionError": "k1link.simulation.worker",
+}
+
+
+def __getattr__(name: str):
+    module = _EXPORTS.get(name)
+    if module is None:
+        raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+    value = getattr(import_module(module), name)
+    globals()[name] = value
+    return value
+
+
+def __dir__():
+    return sorted(set(globals()) | set(__all__))
diff --git a/src/k1link/simulation/ai_polygon/__init__.py b/src/k1link/simulation/ai_polygon/__init__.py
new file mode 100644
index 0000000..2d948d5
--- /dev/null
+++ b/src/k1link/simulation/ai_polygon/__init__.py
@@ -0,0 +1 @@
+"""Independent, virtual-only AI polygon; never routes assets through the LCC pipeline."""
diff --git a/src/k1link/simulation/ai_polygon/composition.py b/src/k1link/simulation/ai_polygon/composition.py
new file mode 100644
index 0000000..e3f3d2e
--- /dev/null
+++ b/src/k1link/simulation/ai_polygon/composition.py
@@ -0,0 +1,194 @@
+"""Simulation providers use the same immutable typed composition as AI Inference.
+
+Pinhole camera contracts are intentionally distinct from recorded K1/KB4 ports.
+Only installed adapters are executable; selection JSON never carries code.
+"""
+
+import hashlib
+import json
+from pathlib import Path
+
+from k1link.observatory.modular_composition import (
+    COMPOSITION_SCHEMA,
+    CompositionError,
+    ModuleRegistry,
+    ModuleSpec,
+    canonical_bytes,
+)
+
+
+def digest(value):
+    return hashlib.sha256(canonical_bytes(value)).hexdigest()
+
+
+def registry(adapter_root: Path) -> ModuleRegistry:
+    profile = json.loads((adapter_root / "models.worker-006.json").read_text())
+    models = {m["id"]: m for m in profile["models"]}
+    implementation = hashlib.sha256(
+        Path(__file__).with_name("inference.py").read_bytes()
+    ).hexdigest()
+    ddr, detector = models["ddrnet-goose-pytorch-reference"], models["rf_detr_large"]
+    nav = profile["navigation"]
+    ade = models["segformer-b2-ade150"]
+    return ModuleRegistry(
+        (
+            ModuleSpec(
+                "simulation-ddrnet-goose",
+                "DDRNet · GOOSE 64",
+                "segmentation",
+                ddr["image"].removeprefix("sha256:"),
+                implementation,
+                ddr["checkpoint_sha256"],
+                digest(
+                    {
+                        "camera": profile["camera"],
+                        "preprocess": ddr["preprocess"],
+                        "labels": profile["labels_sha256"],
+                    }
+                ),
+                ("source.camera.rgb",),
+                ("segmentation.labels", "segmentation.surface"),
+            ),
+            ModuleSpec(
+                "simulation-segformer-ade",
+                "SegFormer · природные поверхности",
+                "segmentation",
+                ade["image"].removeprefix("sha256:"),
+                digest(
+                    {
+                        "client": implementation,
+                        "server": hashlib.sha256(
+                            (adapter_root / "segformer/server.py").read_bytes()
+                        ).hexdigest(),
+                    }
+                ),
+                ade["checkpoint_sha256"],
+                digest(
+                    {
+                        "camera": profile["camera"],
+                        "preprocess": ade["preprocess"],
+                        "config": ade["config_sha256"],
+                        "processor": ade["processor_sha256"],
+                        "output": "ade150-labels-and-bool-surface-candidate-square512-v1",
+                    }
+                ),
+                ("source.camera.rgb",),
+                ("segmentation.labels", "segmentation.surface"),
+            ),
+            ModuleSpec(
+                "simulation-rf-detr",
+                "RF-DETR · люди и животные",
+                "detection",
+                detector["image"].removeprefix("sha256:"),
+                implementation,
+                detector["checkpoint_sha256"],
+                digest(
+                    {
+                        "camera": profile["camera"],
+                        "preprocess": detector["preprocess"],
+                        "output": "normalized-xyxy-risk-boxes-v1",
+                    }
+                ),
+                ("source.camera.rgb",),
+                ("detection.boxes",),
+            ),
+            ModuleSpec(
+                "simulation-waypoint-mission",
+                "Маршрут · контроль продвижения",
+                "policy",
+                nav["image"].removeprefix("sha256:"),
+                digest(
+                    {
+                        "policy": hashlib.sha256(
+                            Path(__file__).with_name("mission_policy.py").read_bytes()
+                        ).hexdigest(),
+                        "adapter": hashlib.sha256(
+                            (adapter_root / "navigation_client.py").read_bytes()
+                        ).hexdigest(),
+                    }
+                ),
+                None,
+                digest(
+                    {
+                        "task": "operator-metric-waypoints-v1",
+                        "recovery": "three-observed-reverse-and-replan-attempts-v2",
+                    }
+                ),
+                (
+                    "segmentation.surface",
+                    "source.camera.calibration",
+                    "source.lidar",
+                    "source.pose",
+                    "source.simulation-time",
+                ),
+                ("navigation.goal", "navigation.intent"),
+                state_policy="causal-reset-at-source-start",
+            ),
+            ModuleSpec(
+                "simulation-cmu-navigation",
+                "CMU · рельеф и движение",
+                "motion",
+                nav["image"].removeprefix("sha256:"),
+                digest(
+                    {
+                        p.relative_to(adapter_root).as_posix(): hashlib.sha256(
+                            p.read_bytes()
+                        ).hexdigest()
+                        for p in (
+                            adapter_root / "navigation/server.py",
+                            adapter_root / "navigation/footprint.py",
+                            adapter_root / "navigation/terrain_costs.py",
+                            adapter_root / "navigation/terrain_connectivity.cpp",
+                            adapter_root / "navigation/fastdds.xml",
+                            adapter_root / "navigation_client.py",
+                        )
+                    }
+                ),
+                None,
+                digest(
+                    {
+                        "upstream": nav["upstream_commit"],
+                        "footprint": [1, 1],
+                        "inputs": "metric-occluded-range-and-rgb",
+                        "command": "signed-mps-rps-observed-recovery-v2",
+                    }
+                ),
+                (
+                    "detection.boxes",
+                    "navigation.goal",
+                    "navigation.intent",
+                    "segmentation.surface",
+                    "source.camera.calibration",
+                    "source.lidar",
+                    "source.pose",
+                ),
+                ("motion.command", "motion.path"),
+                state_policy="causal-reset-at-source-start",
+            ),
+        )
+    )
+
+
+def compose(adapter_root: Path, selection=None):
+    installed = registry(adapter_root)
+    if selection is None:
+        defaults = json.loads((adapter_root / "models.worker-006.json").read_text())[
+            "default_modules"
+        ]
+        selection = {
+            "schema_version": COMPOSITION_SCHEMA,
+            "selections": [
+                {
+                    "group": m.group,
+                    "module_id": m.module_id,
+                    "module_sha256": m.sha256,
+                    "parameters": {},
+                }
+                for m in installed.modules
+                if m.module_id in defaults
+            ],
+        }
+    result = installed.compose(selection, execution_mode="worker-local-simulation")
+    if "motion.command" not in result.outputs:
+        raise CompositionError("Для движения выберите модуль навигации и его зависимости.")
+    return result
diff --git a/src/k1link/simulation/ai_polygon/contracts.py b/src/k1link/simulation/ai_polygon/contracts.py
new file mode 100644
index 0000000..c3b8a11
--- /dev/null
+++ b/src/k1link/simulation/ai_polygon/contracts.py
@@ -0,0 +1,194 @@
+"""Versioned scene preparation and camera-driven run contracts."""
+
+from ipaddress import ip_address, ip_network
+from typing import Annotated, Literal
+
+from pydantic import BaseModel, ConfigDict, Field, HttpUrl, field_validator
+
+
+class Contract(BaseModel):
+    model_config = ConfigDict(extra="forbid", allow_inf_nan=False, str_strip_whitespace=True)
+
+
+class WorldCreate(Contract):
+    name: str = Field(min_length=1, max_length=120)
+    filename: str = Field(pattern=r"^[^/\\\x00-\x1f]{1,200}\.[pP][lL][yY]$")
+    byte_length: int = Field(gt=0, le=8 * 1024**3, strict=True)
+    author: str = Field(min_length=1, max_length=160)
+    license: str = Field(min_length=1, max_length=160)
+    source_url: HttpUrl | None = None
+
+
+class WorldSettings(Contract):
+    # Source -> metric Z-up world; collision proxy is prepared separately.
+    meters_per_unit: float = Field(default=1, ge=0.0001, le=1000)
+    rotation_degrees: tuple[
+        Annotated[float, Field(ge=-360, le=360)],
+        Annotated[float, Field(ge=-360, le=360)],
+        Annotated[float, Field(ge=-360, le=360)],
+    ] = (0, 0, 0)
+    ground_z: float = Field(default=0, ge=-10000, le=10000)
+    spawn_xy: tuple[
+        Annotated[float, Field(ge=-10000, le=10000)], Annotated[float, Field(ge=-10000, le=10000)]
+    ] = (0, 0)
+    heading_degrees: float = Field(default=0, ge=-360, le=360)
+    camera_height_m: float = Field(default=0.5, ge=0.1, le=3)
+    max_speed_mps: float = Field(default=0.3, ge=0.05, le=1)
+    prepared: bool = False
+    route_xy: list[
+        tuple[
+            Annotated[float, Field(ge=-10000, le=10000)],
+            Annotated[float, Field(ge=-10000, le=10000)],
+        ]
+    ] = Field(default_factory=list, max_length=32)
+
+
+class RunCreate(Contract):
+    world_id: str = Field(pattern=r"^aiworld-[a-f0-9]{32}$")
+    max_steps: int = Field(default=600, ge=1, le=3600, strict=True)
+    clock: Literal["lockstep", "realtime"] = "lockstep"
+    start_paused: bool = False
+    duration_seconds: int = Field(default=1800, ge=10, le=7200, strict=True)
+    composition: dict | None = None
+
+
+class WorkerWorldCreate(WorldCreate):
+    """Attestation from the connected Worker after validating its local files."""
+
+    sha256: str = Field(pattern=r"^[a-f0-9]{64}$")
+    collider_sha256: str = Field(pattern=r"^[a-f0-9]{64}$")
+    splat_count: int = Field(ge=1, le=20_000_000, strict=True)
+    settings: WorldSettings
+
+
+class StreamEndpoint(Contract):
+    server: str
+    signaling_port: Literal[49100] = 49100
+    media_port: Literal[47998] = 47998
+    width: Literal[1280] = 1280
+    height: Literal[720] = 720
+    fps: Literal[30] = 30
+
+    @field_validator("server")
+    @classmethod
+    def private_address(cls, value: str) -> str:
+        address = ip_address(value)
+        if (
+            address.version != 4
+            or not (address.is_private or address in ip_network("100.64.0.0/10"))
+            or address.is_unspecified
+            or address.is_multicast
+        ):
+            raise ValueError("Streaming requires a private IPv4 address")
+        return str(address)
+
+
+class WorkerHello(Contract):
+    worker_id: str = Field(pattern=r"^[a-zA-Z0-9_-]{1,80}$")
+    instance_id: str = Field(pattern=r"^[a-f0-9]{32}$")
+    runtime: Literal["isaac-sim-6.1"]
+    model_ids: list[Annotated[str, Field(min_length=1, max_length=100)]] = Field(
+        min_length=1, max_length=8
+    )
+    runtime_sources: dict[
+        Literal["worker", "scene", "models", "robot"],
+        Annotated[str, Field(pattern=r"^[a-f0-9]{64}$")],
+    ] = Field(min_length=4, max_length=4)
+    profile_sha256: str = Field(pattern=r"^[a-f0-9]{64}$")
+    execution_modes: list[Literal["lockstep", "realtime"]] = ["lockstep"]
+    stream: StreamEndpoint | None = None
+
+
+class Decision(Contract):
+    speed_mps: float = Field(ge=-1, le=1)
+    yaw_rate_rps: float = Field(ge=-1, le=1)
+    reason: Literal[
+        "road",
+        "obstacle",
+        "no-road",
+        "uncertain",
+        "inference-error",
+        "replanning",
+        "stuck",
+        "goal-reached",
+        "unstable",
+        "waiting",
+    ]
+    road_fraction: float = Field(ge=0, le=1)
+    obstacle_count: int = Field(ge=0, le=300, strict=True)
+
+
+class RunSample(Contract):
+    sequence: int = Field(ge=0, le=3600, strict=True)
+    simulation_time_ns: int = Field(ge=0, strict=True)
+    inference_ms: float = Field(ge=0, le=120000)
+    # The snapshot is an observation BEFORE the action below is applied.
+    pose_xy: tuple[float, float]
+    decision: Decision
+    image_jpeg_base64: str = Field(max_length=2 * 1024 * 1024)
+
+
+class WorkerPoll(Contract):
+    instance_id: str = Field(pattern=r"^[a-f0-9]{32}$")
+    run_id: str | None = Field(default=None, pattern=r"^airun-[a-f0-9]{32}$")
+
+
+class RunProgress(Contract):
+    phase: Literal["world", "models", "scene"]
+
+
+class RunApplied(Contract):
+    sequence: int = Field(ge=0, le=3600, strict=True)
+    simulation_time_ns: int = Field(ge=0, strict=True)
+    physics_steps: Literal[6]
+    pose_xy: tuple[float, float]
+    pose_yaw: float | None = None
+    cycle_ms: float | None = Field(default=None, ge=0, le=120000)
+    render_ms: float | None = Field(default=None, ge=0, le=120000)
+    transport_ms: float | None = Field(default=None, ge=0, le=120000)
+
+
+class WorkerResult(Contract):
+    instance_id: str = Field(pattern=r"^[a-f0-9]{32}$")
+    outcome: Literal["completed", "stopped", "failed"]
+    message: str = Field(default="", max_length=500)
+    resources_released: Literal[True]
+
+
+class ViewControl(Contract):
+    camera: Literal["follow", "overview", "camera"]
+
+
+class RealtimeSnapshot(Contract):
+    """Bounded observation of Worker-owned state; never a physics-step receipt."""
+
+    sequence: int = Field(ge=0, strict=True)
+    control_sequence: int = Field(ge=0, strict=True)
+    state: Literal["ready", "running", "paused", "stopping"]
+    phase: Literal["scene", "models", "running"]
+    simulation_time_ns: int = Field(ge=0, strict=True)
+    wall_elapsed_seconds: float = Field(ge=0)
+    physics_steps: int = Field(ge=0, strict=True)
+    render_frames: int = Field(ge=0, strict=True)
+    sensor_frames: int = Field(ge=0, strict=True)
+    inference_count: int = Field(ge=0, strict=True)
+    dropped_frames: int = Field(ge=0, strict=True)
+    rtf: float = Field(ge=0, le=100)
+    render_fps: float = Field(ge=0, le=1000)
+    sensor_fps: float = Field(ge=0, le=1000)
+    ai_hz: float = Field(ge=0, le=1000)
+    inference_ms: float | None = Field(default=None, ge=0)
+    frame_age_ms: float | None = Field(default=None, ge=0)
+    command_age_ms: float | None = Field(default=None, ge=0)
+    pose_xy: tuple[float, float]
+    pose_yaw: float
+    speed_mps: float
+    applied_speed_mps: float = Field(ge=-1, le=1)
+    applied_yaw_rate_rps: float = Field(ge=-1, le=1)
+    stop_reason: Literal[
+        "none", "paused", "stale-camera", "stale-command", "inference-error", "unstable"
+    ]
+    decision: Decision | None = None
+    ai_ready: bool
+    stream_ready: bool
+    camera: Literal["follow", "overview", "camera"]
diff --git a/src/k1link/simulation/ai_polygon/inference.py b/src/k1link/simulation/ai_polygon/inference.py
new file mode 100644
index 0000000..57b46f2
--- /dev/null
+++ b/src/k1link/simulation/ai_polygon/inference.py
@@ -0,0 +1,158 @@
+"""Adapters for the pinned reference DDRNet and RF-DETR model contracts.
+
+The simulation RGB pinhole profile is separate from the device's KB4 profile.
+No device valid-FOV mask or recorded LiDAR calibration is applied to simulator RGB.
+"""
+
+import csv
+import http.client
+from pathlib import Path
+from urllib.parse import urlsplit
+
+import numpy as np
+from PIL import Image
+
+from k1link.perception.rf_detr_object_detector import (
+    COCO_SPARSE_TO_CONTIGUOUS,
+    RISK_CLASS_IDS,
+    TritonRfDetrHttpInferenceBackend,
+    preprocess_raw_kb4_rf_detr,
+)
+
+MODEL_IDS = ["ddrnet-goose-pytorch-reference", "rf_detr_large"]
+GOOSE_SURFACE_IDS = (3, 5, 7, 9, 11, 18, 21, 23, 24, 31, 40, 50, 62)
+
+
+def local_endpoint(endpoint: str):
+    parsed = urlsplit(endpoint)
+    if (
+        parsed.scheme != "http"
+        or parsed.hostname not in {"127.0.0.1", "localhost", "::1"}
+        or parsed.path not in {"", "/"}
+        or parsed.query
+        or parsed.fragment
+        or parsed.username
+    ):
+        raise ValueError("Use a local Triton endpoint or a loopback tunnel")
+    return parsed
+
+
+class PillowResizer:
+    def resize(self, image, width, height):
+        return np.asarray(Image.fromarray(image).resize((width, height), Image.Resampling.BILINEAR))
+
+
+class ModelInference:
+    def __init__(
+        self,
+        endpoint: str,
+        goose_labels: Path,
+        ddrnet_endpoint: str,
+        *,
+        segmenter_id: str = "simulation-ddrnet-goose",
+    ):
+        if segmenter_id not in {"simulation-ddrnet-goose", "simulation-segformer-ade"}:
+            raise ValueError("Uninstalled surface provider")
+        self.segmenter_id = segmenter_id
+        parsed = local_endpoint(ddrnet_endpoint)
+        local_endpoint(endpoint)
+        with goose_labels.open(newline="") as stream:
+            labels = {int(row["label_key"]): row["class_name"] for row in csv.DictReader(stream)}
+        if set(labels) != set(range(64)):
+            raise ValueError("Expected the existing GOOSE fine-64 label table")
+        self.road_ids = [
+            i
+            for i, name in labels.items()
+            if name in {"asphalt", "bikeway", "cobble", "sidewalk", "gravel", "soil"}
+        ]
+        if len(self.road_ids) != 6:
+            raise ValueError("GOOSE road label mapping is incomplete")
+        self.connection = http.client.HTTPConnection(
+            parsed.hostname, parsed.port or 8000, timeout=10
+        )
+        self.detector = TritonRfDetrHttpInferenceBackend(endpoint, timeout_seconds=10)
+        self.detector_url = local_endpoint(endpoint)
+
+    def ready(self):
+        self.connection.request("GET", "/ready")
+        response = self.connection.getresponse()
+        response.read(65536)
+        if response.status != 200:
+            raise RuntimeError("Selected surface provider is unavailable")
+        connection = http.client.HTTPConnection(
+            self.detector_url.hostname, self.detector_url.port or 8000, timeout=10
+        )
+        try:
+            connection.request("GET", "/v2/models/rf_detr_large/versions/1/ready")
+            response = connection.getresponse()
+            response.read(65536)
+            if response.status != 200:
+                raise RuntimeError("RF-DETR is unavailable")
+        finally:
+            connection.close()
+
+    def segment(self, rgb: np.ndarray):
+        return self.surface(rgb)["segmentation.labels"]
+
+    def surface(self, rgb: np.ndarray):
+        if rgb.shape != (600, 800, 3) or rgb.dtype != np.uint8:
+            raise ValueError("Expected an 800x600 RGB simulation camera")
+        self.connection.request(
+            "POST",
+            "/infer",
+            body=np.ascontiguousarray(rgb).tobytes(),
+            headers={"Content-Type": "application/octet-stream"},
+        )
+        response = self.connection.getresponse()
+        ade = self.segmenter_id == "simulation-segformer-ade"
+        expected = 512**2 * (2 if ade else 1)
+        raw = response.read(expected + 1)
+        if response.status != 200 or len(raw) != expected:
+            raise RuntimeError("Surface provider response changed")
+        mask = np.frombuffer(raw[: 512**2], dtype=np.uint8).reshape(512, 512)
+        if np.any(mask >= (150 if ade else 64)):
+            raise RuntimeError("Surface label taxonomy changed")
+        if ade:
+            candidate = np.frombuffer(raw[512**2 :], dtype=np.uint8).reshape(512, 512)
+            if np.any(candidate > 1):
+                raise RuntimeError("Surface candidate raster changed")
+            candidate = candidate.astype(bool)
+        else:
+            candidate = np.isin(mask, GOOSE_SURFACE_IDS)
+        return {"segmentation.labels": mask.copy(), "segmentation.surface": candidate}
+
+    def detect(self, rgb: np.ndarray):
+        if rgb.shape != (600, 800, 3) or rgb.dtype != np.uint8:
+            raise ValueError("Expected an 800x600 RGB simulation camera")
+        bgr = np.ascontiguousarray(rgb[:, :, ::-1])
+        detector_tensor = preprocess_raw_kb4_rf_detr(
+            bgr, np.ones((600, 800), dtype=bool), resizer=PillowResizer()
+        )
+        output = self.detector.infer(detector_tensor)
+        if (
+            output.boxes.shape != (1, 300, 4)
+            or output.logits.shape != (1, 300, 91)
+            or not np.isfinite(output.boxes).all()
+            or not np.isfinite(output.logits).all()
+        ):
+            raise RuntimeError("RF-DETR output contract changed")
+        probabilities = 1 / (1 + np.exp(-np.clip(output.logits[0].astype(np.float32), -80, 80)))
+        classes = [
+            key for key, value in COCO_SPARSE_TO_CONTIGUOUS.items() if value in RISK_CLASS_IDS
+        ]
+        scores = probabilities[:, classes].max(axis=1)
+        boxes = []
+        for x, y, width, height in output.boxes[0][scores >= 0.25]:
+            if width <= 0 or height <= 0:
+                continue
+            # Unlike the recorded diagnostic filter, never discard a very large near obstacle.
+            box = np.clip([x - width / 2, y - height / 2, x + width / 2, y + height / 2], 0, 1)
+            boxes.append(tuple(float(value) for value in box))
+        return boxes
+
+    def infer(self, rgb: np.ndarray):
+        return np.isin(self.segment(rgb), self.road_ids), self.detect(rgb)
+
+    def close(self):
+        self.connection.close()
+        self.detector.close()
diff --git a/src/k1link/simulation/ai_polygon/mission_policy.py b/src/k1link/simulation/ai_polygon/mission_policy.py
new file mode 100644
index 0000000..9d77a21
--- /dev/null
+++ b/src/k1link/simulation/ai_polygon/mission_policy.py
@@ -0,0 +1,116 @@
+"""Bounded waypoint mission and progress recovery, driven only by observations.
+
+Operator waypoints specify the task. They are not a collision map or a motion
+trajectory. CMU retains authority to reject every requested local waypoint.
+"""
+
+import math
+
+import numpy as np
+
+
+def inclination(pose):
+    x, y = pose[3:5]
+    return math.degrees(math.acos(max(-1, min(1, 1 - 2 * (x * x + y * y)))))
+
+
+class WaypointMission:
+    def __init__(self, route):
+        self.route = [list(p) for p in route]
+        self.index = self.attempts = 0
+        self.anchor = self.anchor_time = None
+        self.last_time = None
+        self.state = "following"
+        self.failed_goals = []
+        self.goal = None
+        self.fault = None
+        self.best_distance = None
+        self.recovery_goal = None
+        self.recovery_started = None
+
+    def resume(self):
+        # Pause freezes the world clock. Preserve the mission cursor and any
+        # latched failure. The paused world has not moved: retain the observed
+        # goal entering the camera blind strip, but revalidate it before motion.
+        self.anchor = self.anchor_time = None
+
+    def update(self, pose, seconds, choose_goal, choose_recovery=None):
+        xy = np.asarray(pose[:2])
+        if self.last_time is not None and seconds < self.last_time:
+            raise ValueError("Mission observations must not rewind")
+        self.last_time = seconds
+        if inclination(pose) >= 30:
+            self.fault = "unstable"
+        if self.fault:
+            return None, self.intent(self.fault)
+        if self.route and np.linalg.norm(xy - self.route[self.index]) <= 0.4:
+            if self.index == len(self.route) - 1:
+                self.fault = "goal-reached"
+                return None, self.intent(self.fault)
+            self.index += 1
+            self.anchor = self.goal = None
+            self.best_distance = None
+            self.recovery_goal = None
+            self.attempts = 0
+        target = self.route[self.index] if self.route else None
+        if self.anchor is None:
+            self.anchor, self.anchor_time = xy.copy(), seconds
+        distance = float(np.linalg.norm(xy - target)) if target is not None else None
+        if self.best_distance is None:
+            self.best_distance = distance
+        progress = self.best_distance - distance if target is not None else 0.0
+        if target is None and self.goal is not None:
+            direction = np.asarray(self.goal[:2]) - self.anchor
+            progress = float(
+                np.dot(xy - self.anchor, direction) / max(np.linalg.norm(direction), 0.1)
+            )
+        if self.recovery_goal is not None:
+            remaining = np.linalg.norm(xy - self.recovery_goal[:2])
+            observed = choose_recovery(self.recovery_goal) if choose_recovery is not None else None
+            if remaining > 0.3 and seconds - self.recovery_started < 6 and observed is not None:
+                return self.recovery_goal, self.intent("reversing")
+            # Recovery is an attempt to escape, never route progress. The next
+            # forward attempt must beat the previous best distance to the task.
+            self.recovery_goal = self.goal = None
+            self.anchor_time = seconds
+            return None, self.intent("replanning")
+        if progress >= 0.1:
+            # A reverse/forward cycle cannot replenish the recovery budget.
+            self.anchor, self.anchor_time = xy.copy(), seconds
+            self.best_distance = distance
+            self.attempts = 0
+            self.failed_goals.clear()
+        stalled = seconds - self.anchor_time >= 8
+        if stalled:
+            self.attempts += 1
+            if self.goal is not None:
+                self.failed_goals.append(self.goal)
+            self.goal = None
+            self.anchor_time = seconds
+            if self.attempts > 3:
+                self.fault = "stuck"
+                return None, self.intent("stuck")
+            self.state = "replanning"
+            if choose_recovery is not None:
+                self.recovery_goal = choose_recovery(None)
+                if self.recovery_goal is not None:
+                    self.recovery_started = seconds
+                    return self.recovery_goal, self.intent("reversing")
+        observed = choose_goal(target, self.goal, self.failed_goals)
+        if observed is None:
+            # A rejected frame stops motion, not causal memory. Forgetting the
+            # last observed exact waypoint strands it behind the near camera
+            # boundary on the next frame. choose_goal must revalidate visible
+            # support, and live range/collision checks retain final authority.
+            return None, self.intent("no-road")
+        self.goal = observed
+        self.state = "replanning" if self.attempts else "following"
+        return self.goal, self.intent(self.state)
+
+    def intent(self, state):
+        return dict(
+            state=state,
+            waypoint=self.index,
+            waypoint_count=len(self.route),
+            recovery_attempt=self.attempts,
+        )
diff --git a/src/k1link/simulation/ai_polygon/policy.py b/src/k1link/simulation/ai_polygon/policy.py
new file mode 100644
index 0000000..90ffc4a
--- /dev/null
+++ b/src/k1link/simulation/ai_polygon/policy.py
@@ -0,0 +1,79 @@
+"""Camera-only laboratory road follower. No scene truth or actor coordinates enter here.
+
+This deliberately small baseline measures model-driven following/braking. It is
+not a route planner, metric obstacle-distance estimator, or field safety policy.
+"""
+
+import numpy as np
+
+from k1link.simulation.ai_polygon.contracts import Decision
+
+
+class RoadPolicy:
+    def __init__(self, max_speed_mps: float):
+        if not 0 < max_speed_mps <= 1:
+            raise ValueError("invalid laboratory speed")
+        self.max_speed = max_speed_mps
+        self.clear_frames = 0
+
+    def reset(self):
+        self.clear_frames = 0
+
+    def decide(
+        self, road_mask: np.ndarray, boxes: list[tuple[float, float, float, float]]
+    ) -> Decision:
+        if road_mask.shape != (512, 512) or road_mask.dtype != np.bool_:
+            raise ValueError("road mask must be the model's 512x512 boolean raster")
+        # Detector boxes are normalized in the full RGB camera, not the DDRNet crop.
+        if any(
+            not all(np.isfinite(box))
+            or not (0 <= box[0] <= box[2] <= 1 and 0 <= box[1] <= box[3] <= 1)
+            for box in boxes
+        ):
+            raise ValueError("invalid observation box")
+        obstacles = sum(x1 < 0.68 and x2 > 0.32 and y2 > 0.55 for x1, _, x2, y2 in boxes)
+        near_road = road_mask[320:500, 100:412]
+        fraction = float(near_road.mean())
+        stop = "obstacle" if obstacles else "no-road" if fraction < 0.35 else None
+        if stop:
+            self.clear_frames = 0
+            return Decision(
+                speed_mps=0,
+                yaw_rate_rps=0,
+                reason=stop,
+                road_fraction=fraction,
+                obstacle_count=obstacles,
+            )
+        self.clear_frames += 1
+        if self.clear_frames < 3:
+            return Decision(
+                speed_mps=0,
+                yaw_rate_rps=0,
+                reason="uncertain",
+                road_fraction=fraction,
+                obstacle_count=0,
+            )
+        # Choose a contiguous visible corridor, rather than averaging two disconnected roads.
+        columns = near_road.mean(axis=0) >= 0.6
+        edges = np.flatnonzero(np.diff(np.r_[False, columns, False].astype(np.int8)))
+        spans = [(a, b) for a, b in zip(edges[::2], edges[1::2], strict=True) if b - a >= 48]
+        if not spans:
+            self.clear_frames = 0
+            return Decision(
+                speed_mps=0,
+                yaw_rate_rps=0,
+                reason="uncertain",
+                road_fraction=fraction,
+                obstacle_count=0,
+            )
+        left, right = min(spans, key=lambda span: abs((span[0] + span[1]) / 2 - 156))
+        center = (left + right) / 2
+        yaw = float(np.clip((156 - center) / 156, -0.6, 0.6))
+        speed = self.max_speed * min(1, fraction / 0.65) * (1 - abs(yaw))
+        return Decision(
+            speed_mps=float(speed),
+            yaw_rate_rps=yaw,
+            reason="road",
+            road_fraction=fraction,
+            obstacle_count=0,
+        )
diff --git a/src/k1link/simulation/ai_polygon/runs.py b/src/k1link/simulation/ai_polygon/runs.py
new file mode 100644
index 0000000..4127839
--- /dev/null
+++ b/src/k1link/simulation/ai_polygon/runs.py
@@ -0,0 +1,412 @@
+"""Single-worker laboratory lifecycle and durable observation/decision journal."""
+
+import base64
+import hashlib
+import io
+import json
+import os
+import re
+import secrets
+import shutil
+import stat
+import threading
+import time
+from pathlib import Path
+from uuid import uuid4
+
+from PIL import Image
+
+from k1link.artifacts import utc_now_iso
+from k1link.observatory.recorded_jobs import ObservatoryRecordedJobQueue
+from k1link.simulation.ai_polygon.contracts import (
+    RealtimeSnapshot,
+    RunApplied,
+    RunCreate,
+    RunSample,
+    WorkerHello,
+)
+from k1link.simulation.ai_polygon.worlds import WorldStore, write_json
+
+RUN_ID = re.compile(r"^airun-[a-f0-9]{32}$")
+TERMINAL = {"completed", "stopped", "failed"}
+WORKER_LEASE_SECONDS = 20
+
+
+class RunStore:
+    def __init__(self, worlds: WorldStore, queue: ObservatoryRecordedJobQueue | None = None):
+        self.worlds = worlds
+        self.queue = queue
+        self.root = worlds.root.parent / "runs"
+        self.root.mkdir(mode=0o700, exist_ok=True)
+        self.lock = threading.RLock()
+        self.worker: dict | None = None
+        self.seen = 0.0
+        self.active: str | None = None
+        self.token_path = self.root.parent / "worker.token"
+        if not self.token_path.exists():
+            try:
+                fd = os.open(self.token_path, os.O_CREAT | os.O_EXCL | os.O_WRONLY, 0o600)
+                with os.fdopen(fd, "w") as stream:
+                    stream.write(secrets.token_urlsafe(48))
+            except FileExistsError:
+                pass
+        metadata = self.token_path.lstat()
+        if not stat.S_ISREG(metadata.st_mode) or metadata.st_mode & 0o077:
+            raise ValueError("AI polygon worker token must be a private regular file")
+        with os.fdopen(os.open(self.token_path, os.O_RDONLY | os.O_NOFOLLOW), "r") as stream:
+            self.token = stream.read(513).strip()
+        if not 32 <= len(self.token) <= 512:
+            raise ValueError("AI polygon worker token is invalid")
+        # Keep realtime ownership uncertain until the same Worker reconciles it.
+        # A Core restart never proves that the remote GPU/process stopped.
+        for row in self.list():
+            if row["state"] not in TERMINAL:
+                if row.get("clock") == "realtime":
+                    if self.active is not None:
+                        raise RuntimeError("Multiple unreconciled simulation owners")
+                    self.active = row["run_id"]
+                    row.update(state="disconnected", message="Ожидаем состояние Worker.")
+                else:
+                    row.update(state="failed", message="Связь с симуляцией прервана перезапуском.")
+                self._save(row)
+
+    def directory(self, run_id: str) -> Path:
+        if not RUN_ID.fullmatch(run_id):
+            raise FileNotFoundError(run_id)
+        path = self.root / run_id
+        if not path.is_dir() or path.is_symlink():
+            raise FileNotFoundError(run_id)
+        return path
+
+    def get(self, run_id: str) -> dict:
+        return json.loads((self.directory(run_id) / "run.json").read_text())
+
+    def list(self) -> list[dict]:
+        return sorted(
+            [
+                self.get(p.name)
+                for p in self.root.iterdir()
+                if RUN_ID.fullmatch(p.name) and p.is_dir() and not p.is_symlink()
+            ],
+            key=lambda item: item["created_at"],
+            reverse=True,
+        )
+
+    def _save(self, doc: dict) -> None:
+        write_json(self.directory(doc["run_id"]) / "run.json", doc)
+
+    def _expire(self) -> None:
+        if self.worker is not None and time.monotonic() - self.seen > WORKER_LEASE_SECONDS:
+            if self.active:
+                row = self.get(self.active)
+                if row.get("clock") == "realtime":
+                    row.update(
+                        state="disconnected",
+                        message="Связь с Worker потеряна. Состояние уточняется.",
+                    )
+                else:
+                    row.update(state="failed", message="Worker потерял связь. Прогон остановлен.")
+                    self.active = None
+                self._save(row)
+            self.worker = None
+
+    def status(self) -> dict:
+        with self.lock:
+            self._expire()
+            return {
+                "available": self.worker is not None,
+                "worker": self.worker,
+                "active_run": self.get(self.active) if self.active else None,
+            }
+
+    def register(self, hello: WorkerHello) -> dict:
+        with self.lock:
+            self._expire()
+            if self.worker and self.worker["instance_id"] != hello.instance_id:
+                raise RuntimeError("Другой Worker уже подключён.")
+            if self.active and self.get(self.active)["worker"] != hello.model_dump():
+                raise RuntimeError("Нельзя менять модели во время прогона.")
+            self.worker = hello.model_dump()
+            self.seen = time.monotonic()
+            return {"registered": True}
+
+    def heartbeat(self, instance_id: str) -> dict:
+        with self.lock:
+            self._require_worker(instance_id)
+            self.seen = time.monotonic()
+            return {"alive": True}
+
+    def _require_worker(self, instance_id: str) -> None:
+        self._expire()
+        if self.worker is None or self.worker["instance_id"] != instance_id:
+            raise RuntimeError("Сессия Worker истекла; требуется переподключение.")
+
+    def start(self, request: RunCreate, request_id: str) -> dict:
+        if not re.fullmatch(r"[a-zA-Z0-9_-]{8,100}", request_id):
+            raise ValueError("Требуется уникальный идентификатор запуска.")
+        with self.lock:
+            self._expire()
+            for old in self.list():
+                if old["request_id"] == request_id:
+                    if old["request"] != request.model_dump():
+                        raise RuntimeError("Идентификатор запуска уже использован.")
+                    return old
+            if self.worker is None:
+                raise RuntimeError("Подключите Worker с симуляцией и моделями inference.")
+            if request.clock not in self.worker.get("execution_modes", ["lockstep"]):
+                raise RuntimeError("Worker не поддерживает этот режим симуляции.")
+            if request.clock == "realtime" and not self.worker.get("stream"):
+                raise RuntimeError("Видеопоток Worker не настроен.")
+            if self.queue is None:
+                raise RuntimeError("Контроль занятости Worker недоступен.")
+            if self.active is not None:
+                raise RuntimeError("Сначала завершите текущий прогон.")
+            world = self.worlds.get(request.world_id)
+            storage = world.get("storage", {})
+            if storage.get("kind") == "worker" and (
+                storage["worker_id"] != self.worker["worker_id"] or request.clock != "realtime"
+            ):
+                raise RuntimeError("Локация доступна только на подготовленном Worker.")
+            if world["status"] != "available" or not world["settings"]["prepared"]:
+                raise RuntimeError("Сначала проверьте масштаб, грунт и старт ровера.")
+            run_id = f"airun-{uuid4().hex}"
+            if (
+                request.clock == "lockstep"
+                and shutil.disk_usage(self.root).free < request.max_steps * 1024**2 + 512 * 1024**2
+            ):
+                raise ValueError("Недостаточно места для кадров прогона.")
+            (self.root / run_id).mkdir(mode=0o700)
+            if request.clock == "lockstep":
+                (self.root / run_id / "frames").mkdir(mode=0o700)
+            row = {
+                "schema_version": "missioncore.ai-polygon-run/v1",
+                "run_id": run_id,
+                "request_id": request_id,
+                "request": request.model_dump(),
+                "created_at": utc_now_iso(),
+                "world": world,
+                "worker": dict(self.worker),
+                "state": "starting",
+                "control": "pause" if request.start_paused else "play",
+                "control_sequence": 0,
+                "camera": "follow",
+                "telemetry": None,
+                "step_budget": 0,
+                "samples": 0,
+                "applied_steps": 0,
+                "last_applied": None,
+                "last_sample": None,
+                "message": None,
+                "phase": "world",
+                "authority": "virtual-only",
+                "clock": request.clock,
+                "step_ns": 100_000_000,
+            }
+            self._save(row)
+            try:
+                self.queue.reserve_simulation(run_id)
+            except RuntimeError:
+                row.update(state="failed", message="Worker занят другой задачей.")
+                self._save(row)
+                raise
+            self.active = run_id
+            return row
+
+    def control(self, run_id: str, command: str) -> dict:
+        with self.lock:
+            self._expire()
+            row = self.get(run_id)
+            if row["state"] in TERMINAL:
+                return row
+            if command not in {"pause", "play", "step", "stop"}:
+                raise ValueError("Неизвестная команда симуляции.")
+            if command == "step" and row.get("clock") == "realtime":
+                raise RuntimeError("Realtime не допускает пошаговое продвижение времени.")
+            if command == "step" and row["state"] != "paused":
+                raise RuntimeError("Один шаг доступен только на паузе.")
+            if row["control"] == "stop":
+                return row
+            if row["control"] == command:
+                return row
+            row["control_sequence"] = row.get("control_sequence", 0) + 1
+            row["control"] = "pause" if command == "step" else command
+            row["step_budget"] = 1 if command == "step" else 0
+            # Paused is acknowledged by the Worker, not inferred from this request.
+            if command == "stop":
+                row["state"] = "stopping"
+            self._save(row)
+            return row
+
+    def poll(self, instance_id: str, run_id: str | None) -> dict:
+        with self.lock:
+            self._require_worker(instance_id)
+            self.seen = time.monotonic()
+            if self.active is None:
+                return {"run": None, "action": "idle"}
+            row = self.get(self.active)
+            if run_id is None:
+                return {"run": row, "action": "load"}
+            if run_id != self.active:
+                raise RuntimeError("Прогон Worker не совпадает с активным.")
+            action = row["control"]
+            if row.get("clock") == "realtime":
+                # Only actual local snapshots can acknowledge pause/play.
+                return {"run": row, "action": action}
+            if action == "pause":
+                if row["step_budget"]:
+                    row["step_budget"] = 0
+                    action = "step"
+                else:
+                    row["state"] = "paused"
+            elif action == "play" and row["samples"]:
+                row["state"] = "running"
+            self._save(row)
+            return {"run": row, "action": action}
+
+    def progress(self, run_id: str, instance_id: str, phase: str) -> dict:
+        with self.lock:
+            self._require_worker(instance_id)
+            if run_id != self.active:
+                raise RuntimeError("Прогон уже завершён.")
+            row = self.get(run_id)
+            if row["state"] == "starting":
+                row["phase"] = phase
+                self._save(row)
+            return {"control": row["control"]}
+
+    def view(self, run_id: str, camera: str) -> dict:
+        with self.lock:
+            row = self.get(run_id)
+            if run_id != self.active or row.get("clock") != "realtime":
+                raise RuntimeError("Симуляция не запущена.")
+            if camera not in {"follow", "overview", "camera"}:
+                raise ValueError("Неизвестная камера.")
+            row["camera"] = camera
+            self._save(row)
+            return row
+
+    def snapshot(self, run_id: str, instance_id: str, snapshot: RealtimeSnapshot) -> dict:
+        with self.lock:
+            self._require_worker(instance_id)
+            row = self.get(run_id)
+            if self.active != run_id or row.get("clock") != "realtime":
+                raise RuntimeError("Прогон не принадлежит realtime Worker.")
+            if row["worker"]["instance_id"] != instance_id:
+                raise RuntimeError("Прогон принадлежит другому Worker.")
+            old = row.get("telemetry")
+            if old and snapshot.sequence <= old["sequence"]:
+                return {
+                    "recorded": old["sequence"]
+                }  # Same snapshot may be retried after reconnect.
+            if snapshot.control_sequence > row.get("control_sequence", 0):
+                raise ValueError("Worker подтвердил неизвестную команду.")
+            if old and snapshot.simulation_time_ns < old["simulation_time_ns"]:
+                raise ValueError("Время Worker не может идти назад.")
+            if abs(snapshot.applied_speed_mps) > row["world"]["settings"]["max_speed_mps"]:
+                raise ValueError("Команда превышает скорость прогона.")
+            row.update(
+                telemetry=snapshot.model_dump(),
+                phase=snapshot.phase,
+                samples=snapshot.inference_count,
+                message=None,
+            )
+            if row["control"] != "stop":
+                row["state"] = snapshot.state
+            row["telemetry_received_at"] = utc_now_iso()
+            self._save(row)
+            self.seen = time.monotonic()
+            return {"recorded": snapshot.sequence}
+
+    def sample(self, run_id: str, instance_id: str, sample: RunSample) -> dict:
+        with self.lock:
+            self._require_worker(instance_id)
+            if run_id != self.active:
+                raise RuntimeError("Прогон уже завершён.")
+            row = self.get(run_id)
+            if row.get("clock") != "lockstep":
+                raise RuntimeError("Realtime кадры хранятся только на Worker.")
+            if row["control"] == "stop":
+                raise RuntimeError("Получена команда остановки.")
+            if sample.sequence != row["samples"] or sample.sequence >= row["request"]["max_steps"]:
+                raise RuntimeError("Нарушена последовательность кадров.")
+            if row["samples"] != row["applied_steps"]:
+                raise RuntimeError("Предыдущий шаг физики не подтверждён.")
+            if sample.simulation_time_ns != sample.sequence * row["step_ns"]:
+                raise ValueError("Время кадра не соответствует шагу симуляции.")
+            if abs(sample.decision.speed_mps) > row["world"]["settings"]["max_speed_mps"]:
+                raise ValueError("Команда превышает скорость прогона.")
+            try:
+                image = base64.b64decode(sample.image_jpeg_base64, validate=True)
+                if not 4 <= len(image) <= 1024**2:
+                    raise ValueError()
+                with Image.open(io.BytesIO(image)) as decoded:
+                    if decoded.format != "JPEG" or decoded.size != (800, 600):
+                        raise ValueError()
+                    decoded.verify()
+            except Exception as exc:
+                raise ValueError("Неверный кадр камеры.") from exc
+            directory = self.directory(run_id)
+            if shutil.disk_usage(directory).free < len(image) + 512 * 1024**2:
+                raise ValueError("Недостаточно места для кадра прогона.")
+            filename = f"{sample.sequence:06d}.jpg"
+            (directory / "frames" / filename).write_bytes(image)
+            recorded = sample.model_dump(exclude={"image_jpeg_base64"})
+            recorded.update(image_sha256=hashlib.sha256(image).hexdigest(), image=filename)
+            with (directory / "decisions.jsonl").open("a", encoding="utf-8") as stream:
+                stream.write(json.dumps(recorded, allow_nan=False) + "\n")
+                stream.flush()
+                os.fsync(stream.fileno())
+            row.update(samples=row["samples"] + 1, last_sample=recorded, phase="running")
+            if row["state"] == "starting":
+                row["state"] = "running"
+            self._save(row)
+            return {"recorded": sample.sequence}
+
+    def applied(self, run_id: str, instance_id: str, receipt: RunApplied) -> dict:
+        with self.lock:
+            self._require_worker(instance_id)
+            if self.active != run_id:
+                raise RuntimeError("Прогон уже завершён.")
+            row = self.get(run_id)
+            if row.get("clock") != "lockstep":
+                raise RuntimeError("Realtime не использует подтверждения физических шагов.")
+            if receipt.sequence != row["applied_steps"] or row["samples"] != receipt.sequence + 1:
+                raise RuntimeError("Шаг не соответствует решению модели.")
+            if receipt.simulation_time_ns != (receipt.sequence + 1) * row["step_ns"]:
+                raise ValueError("Неверное время завершения шага.")
+            recorded = receipt.model_dump()
+            with (self.directory(run_id) / "motion.jsonl").open("a", encoding="utf-8") as stream:
+                stream.write(json.dumps(recorded, allow_nan=False) + "\n")
+                stream.flush()
+                os.fsync(stream.fileno())
+            row.update(applied_steps=row["applied_steps"] + 1, last_applied=recorded)
+            self._save(row)
+            return {"applied": receipt.sequence}
+
+    def finish(self, run_id: str, instance_id: str, outcome: str, message: str) -> dict:
+        with self.lock:
+            self._require_worker(instance_id)
+            row = self.get(run_id)
+            if row["worker"]["instance_id"] != instance_id:
+                raise RuntimeError("Прогон принадлежит другому Worker.")
+            if row["state"] in TERMINAL:
+                if self.queue is not None:
+                    self.queue.release_simulation(run_id)
+                return row
+            if run_id != self.active:
+                raise RuntimeError("Прогон уже завершён.")
+            if (
+                row.get("clock") == "lockstep"
+                and outcome == "completed"
+                and (
+                    row["samples"] != row["request"]["max_steps"]
+                    or row["applied_steps"] != row["samples"]
+                )
+            ):
+                raise RuntimeError("Прогон не достиг заданного числа шагов.")
+            row.update(state=outcome, message=message or None)
+            self._save(row)
+            if self.queue is not None:
+                self.queue.release_simulation(run_id)
+            self.active = None
+            return row
diff --git a/src/k1link/simulation/ai_polygon/terrain_contract.py b/src/k1link/simulation/ai_polygon/terrain_contract.py
new file mode 100644
index 0000000..7748caa
--- /dev/null
+++ b/src/k1link/simulation/ai_polygon/terrain_contract.py
@@ -0,0 +1,24 @@
+"""Collision identity is independent of episode start, heading and camera."""
+
+
+def terrain_matches(manifest, world, generator_sha256=None):
+    if manifest.get("source_sha256") != world["sha256"]:
+        return False
+    if generator_sha256 and manifest.get("generator_sha256") != generator_sha256:
+        return False
+    old, new = manifest["settings"], world["settings"]
+    if any(old[key] != new[key] for key in ("meters_per_unit", "rotation_degrees")):
+        return False
+    if manifest.get("generator") == "paired-source":
+        # A supplied full-scene collider is not the generated 30 m tile below.
+        # Source/calibration and collider hashes still bind this asset; actual
+        # support and full-body clearance are checked at the new start by Isaac.
+        return manifest.get("collider_sha256") == world.get("collider_sha256") and bool(
+            manifest.get("collider_sha256")
+        )
+    # The versioned preparer captures a 30x30 m tile and a 12 m vertical band.
+    # Admit starts only within its interior; physical support is checked later.
+    return (
+        all(abs(old["spawn_xy"][i] - new["spawn_xy"][i]) <= 12 for i in (0, 1))
+        and abs(old["ground_z"] - new["ground_z"]) <= 2
+    )
diff --git a/src/k1link/simulation/ai_polygon/worlds.py b/src/k1link/simulation/ai_polygon/worlds.py
new file mode 100644
index 0000000..c56f81e
--- /dev/null
+++ b/src/k1link/simulation/ai_polygon/worlds.py
@@ -0,0 +1,256 @@
+"""Durable, resumable Gaussian asset admission, independent of XGRIDS sources."""
+
+import hashlib
+import json
+import os
+import re
+import shutil
+import threading
+from pathlib import Path
+from uuid import uuid4
+
+import numpy as np
+
+from k1link.artifacts import utc_now_iso
+from k1link.simulation.ai_polygon.contracts import WorkerWorldCreate, WorldCreate, WorldSettings
+
+CHUNK_BYTES = 4 * 1024**2
+WORLD_ID = re.compile(r"^aiworld-[a-f0-9]{32}$")
+SOURCES = [
+    {
+        "name": "Forest Scan",
+        "author": "draftmode",
+        "license": "CC BY 4.0",
+        "source_url": "https://superspl.at/scene/259c0051",
+        "description": "Лесная тропа и папоротники",
+    },
+    {
+        "name": "Bamboo Trail",
+        "author": "luckysplat",
+        "license": "CC BY 4.0",
+        "source_url": "https://superspl.at/scene/dd49e9a8",
+        "description": "Тропа в бамбуковой роще",
+    },
+]
+
+
+def write_json(path: Path, value: object) -> None:
+    temporary = path.with_name(f".{path.name}-{uuid4().hex}")
+    try:
+        with temporary.open("x", encoding="utf-8") as stream:
+            os.chmod(temporary, 0o600)
+            json.dump(value, stream, ensure_ascii=False, allow_nan=False)
+            stream.flush()
+            os.fsync(stream.fileno())
+        temporary.replace(path)
+    finally:
+        temporary.unlink(missing_ok=True)
+
+
+def inspect_gaussian_ply(path: Path) -> int:
+    """Admit only standard scalar binary 3DGS; do not label meshes as splats."""
+    with path.open("rb") as stream:
+        header = bytearray()
+        while len(header) < 65536:
+            line = stream.readline(1024)
+            header.extend(line)
+            if line.rstrip() == b"end_header":
+                break
+            if not line:
+                raise ValueError("В PLY отсутствует заголовок Gaussian-сцены.")
+        else:
+            raise ValueError("Заголовок PLY превышает допустимый размер.")
+    try:
+        lines = bytes(header).decode("ascii").splitlines()
+    except UnicodeDecodeError as exc:
+        raise ValueError("Некорректный заголовок PLY.") from exc
+    if lines[:2] != ["ply", "format binary_little_endian 1.0"]:
+        raise ValueError("Экспортируйте стандартный Gaussian PLY (binary little-endian).")
+    elements = [line for line in lines if line.startswith("element ")]
+    if len(elements) != 1 or not elements[0].startswith("element vertex "):
+        raise ValueError("Нужен Gaussian PLY с одним элементом vertex, без меша.")
+    count = int(elements[0].split()[-1])
+    properties = [line.split() for line in lines if line.startswith("property ")]
+    names = [prop[-1] for prop in properties]
+    required = {
+        "x",
+        "y",
+        "z",
+        "opacity",
+        "f_dc_0",
+        "f_dc_1",
+        "f_dc_2",
+        "scale_0",
+        "scale_1",
+        "scale_2",
+        "rot_0",
+        "rot_1",
+        "rot_2",
+        "rot_3",
+    }
+    if (
+        not 1 <= count <= 20_000_000
+        or not required.issubset(names)
+        or len(names) != len(set(names))
+        or not 14 <= len(properties) <= 128
+        or any(len(prop) != 3 or prop[1] not in {"float", "float32"} for prop in properties)
+    ):
+        raise ValueError("PLY не содержит поддерживаемые Gaussian-атрибуты.")
+    if path.stat().st_size != len(header) + count * len(properties) * 4:
+        raise ValueError("Размер PLY не соответствует его заголовку.")
+    with path.open("rb") as stream:
+        stream.seek(len(header))
+        for chunk in iter(lambda: stream.read(CHUNK_BYTES), b""):
+            if not np.isfinite(np.frombuffer(chunk, dtype="<f4")).all():
+                raise ValueError("PLY содержит нечисловые или бесконечные атрибуты.")
+    return count
+
+
+class WorldStore:
+    def __init__(self, root: Path):
+        self.root = root / "ai-polygon" / "worlds"
+        self.root.mkdir(parents=True, exist_ok=True, mode=0o700)
+        self.lock = threading.RLock()
+
+    def directory(self, world_id: str) -> Path:
+        if not WORLD_ID.fullmatch(world_id):
+            raise FileNotFoundError(world_id)
+        path = self.root / world_id
+        if path.is_symlink() or not path.is_dir():
+            raise FileNotFoundError(world_id)
+        return path
+
+    def get(self, world_id: str) -> dict:
+        with self.lock:
+            path = self.directory(world_id)
+            document = json.loads((path / "world.json").read_text())
+            if document["status"] == "uploading":
+                source = path / "source.part"
+                if not source.exists():
+                    source = path / "source.ply"
+                document["uploaded_bytes"] = source.stat().st_size
+            return document
+
+    def list(self) -> list[dict]:
+        return sorted(
+            [
+                self.get(p.name)
+                for p in self.root.iterdir()
+                if WORLD_ID.fullmatch(p.name) and p.is_dir() and not p.is_symlink()
+            ],
+            key=lambda item: item["created_at"],
+            reverse=True,
+        )
+
+    def create(self, request: WorldCreate) -> dict:
+        with self.lock:
+            if shutil.disk_usage(self.root).free < request.byte_length + 512 * 1024**2:
+                raise ValueError("Недостаточно места для исходника сцены.")
+            world_id = f"aiworld-{uuid4().hex}"
+            path = self.root / world_id
+            path.mkdir(mode=0o700)
+            (path / "source.part").touch(mode=0o600)
+            document = {
+                "schema_version": "missioncore.ai-polygon-world/v1",
+                "world_id": world_id,
+                **request.model_dump(mode="json"),
+                "status": "uploading",
+                "uploaded_bytes": 0,
+                "sha256": None,
+                "splat_count": None,
+                "created_at": utc_now_iso(),
+                "settings": WorldSettings().model_dump(mode="json"),
+            }
+            write_json(path / "world.json", document)
+            return document
+
+    def register_worker_asset(self, request: WorkerWorldCreate, worker_id: str) -> dict:
+        """Keep only an authenticated asset manifest on the operator machine."""
+        with self.lock:
+            payload = request.model_dump(mode="json")
+            storage = {"kind": "worker", "worker_id": worker_id}
+            for existing in self.list():
+                if existing.get("storage") == storage and existing["sha256"] == request.sha256:
+                    if any(existing.get(k) != v for k, v in payload.items() if k != "settings"):
+                        raise RuntimeError("Манифест сохранённой локации изменился.")
+                    return existing
+            world_id = f"aiworld-{uuid4().hex}"
+            path = self.root / world_id
+            path.mkdir(mode=0o700)
+            document = {
+                "schema_version": "missioncore.ai-polygon-world/v1",
+                "world_id": world_id,
+                **payload,
+                "storage": storage,
+                "status": "available",
+                "uploaded_bytes": 0,
+                "created_at": utc_now_iso(),
+            }
+            write_json(path / "world.json", document)
+            return document
+
+    def append(self, world_id: str, offset: int, payload: bytes) -> dict:
+        with self.lock:
+            doc = self.get(world_id)
+            if doc["status"] != "uploading" or offset != doc["uploaded_bytes"]:
+                raise RuntimeError("Позиция загрузки изменилась. Возобновите передачу.")
+            if not 0 < len(payload) <= CHUNK_BYTES or offset + len(payload) > doc["byte_length"]:
+                raise ValueError("Размер фрагмента загрузки недопустим.")
+            if shutil.disk_usage(self.root).free < len(payload) + 512 * 1024**2:
+                raise ValueError("Недостаточно свободного места.")
+            part = self.directory(world_id) / "source.part"
+            with part.open("r+b") as stream:
+                stream.seek(offset)
+                stream.write(payload)
+                stream.flush()
+                os.fsync(stream.fileno())
+            return self.get(world_id)
+
+    def complete(self, world_id: str) -> dict:
+        with self.lock:
+            doc = self.get(world_id)
+            if doc["status"] == "available":
+                return doc
+            if doc["uploaded_bytes"] != doc["byte_length"]:
+                raise RuntimeError("Загрузка сцены ещё не завершена.")
+            directory = self.directory(world_id)
+            source = directory / "source.part"
+            # Recover an interrupted atomic publication without discarding the source.
+            if not source.exists():
+                source = directory / "source.ply"
+            count = inspect_gaussian_ply(source)
+            digest = hashlib.sha256()
+            with source.open("rb") as stream:
+                for chunk in iter(lambda: stream.read(CHUNK_BYTES), b""):
+                    digest.update(chunk)
+            source.replace(directory / "source.ply")
+            doc.update(status="available", sha256=digest.hexdigest(), splat_count=count)
+            write_json(directory / "world.json", doc)
+            return doc
+
+    def prefix_hashes(self, world_id: str) -> dict:
+        """Verify a resumed file against every byte already admitted, in bounded chunks."""
+        with self.lock:
+            doc = self.get(world_id)
+            if doc.get("storage", {}).get("kind") == "worker":
+                raise RuntimeError("Исходник локации хранится на Worker.")
+            directory = self.directory(world_id)
+            source = directory / "source.part"
+            if not source.exists():
+                source = directory / "source.ply"
+            chunks = []
+            with source.open("rb") as stream:
+                for chunk in iter(lambda: stream.read(CHUNK_BYTES), b""):
+                    chunks.append(
+                        {"byte_length": len(chunk), "sha256": hashlib.sha256(chunk).hexdigest()}
+                    )
+            return {"uploaded_bytes": doc["uploaded_bytes"], "chunks": chunks}
+
+    def configure(self, world_id: str, settings: WorldSettings) -> dict:
+        with self.lock:
+            doc = self.get(world_id)
+            if doc["status"] != "available":
+                raise RuntimeError("Сначала завершите импорт сцены.")
+            doc["settings"] = settings.model_dump(mode="json")
+            write_json(self.directory(world_id) / "world.json", doc)
+            return doc
diff --git a/src/k1link/web/ai_polygon_api.py b/src/k1link/web/ai_polygon_api.py
new file mode 100644
index 0000000..b8581aa
--- /dev/null
+++ b/src/k1link/web/ai_polygon_api.py
@@ -0,0 +1,231 @@
+"""Control Station and authenticated simulation-worker ports for AI polygon."""
+
+import secrets
+from pathlib import Path
+from typing import Literal
+
+from fastapi import APIRouter, Header, HTTPException, Request, Response
+from fastapi.responses import FileResponse
+
+from k1link.observatory.recorded_jobs import ObservatoryRecordedJobQueue
+from k1link.simulation.ai_polygon.composition import compose, registry
+from k1link.simulation.ai_polygon.contracts import (
+    RealtimeSnapshot,
+    RunApplied,
+    RunCreate,
+    RunProgress,
+    RunSample,
+    ViewControl,
+    WorkerHello,
+    WorkerPoll,
+    WorkerResult,
+    WorkerWorldCreate,
+    WorldCreate,
+    WorldSettings,
+)
+from k1link.simulation.ai_polygon.runs import RunStore
+from k1link.simulation.ai_polygon.worlds import CHUNK_BYTES, SOURCES, WorldStore
+
+
+def build_ai_polygon_router(
+    data_dir: Path, queue: ObservatoryRecordedJobQueue | None = None
+) -> APIRouter:
+    worlds = WorldStore(data_dir)
+    runs = RunStore(worlds, queue)
+    router = APIRouter(prefix="/api/v1/ai-polygon", tags=["ai-polygon"])
+    adapters = Path(__file__).resolve().parents[3] / "simulation/ai-polygon"
+
+    def invoke(fn, *args):
+        try:
+            return fn(*args)
+        except FileNotFoundError as exc:
+            raise HTTPException(404, "Локация или прогон не найдены.") from exc
+        except ValueError as exc:
+            raise HTTPException(400, str(exc)) from exc
+        except RuntimeError as exc:
+            raise HTTPException(409, str(exc)) from exc
+
+    def authenticate(authorization: str | None) -> None:
+        expected = f"Bearer {runs.token}"
+        if not authorization or not secrets.compare_digest(authorization, expected):
+            raise HTTPException(401, "Worker authentication required")
+
+    @router.get("/catalog")
+    def catalog():
+        return {
+            "schema_version": "missioncore.ai-polygon-catalog/v1",
+            "sources": SOURCES,
+            "worlds": worlds.list(),
+            "runtime": runs.status(),
+            "runs": runs.list()[:30],
+        }
+
+    @router.post("/worlds", status_code=201)
+    def create_world(body: WorldCreate):
+        return invoke(worlds.create, body)
+
+    @router.get("/worlds/{world_id}")
+    def get_world(world_id: str):
+        return invoke(worlds.get, world_id)
+
+    @router.get("/worlds/{world_id}/upload-prefix")
+    def upload_prefix(world_id: str):
+        return invoke(worlds.prefix_hashes, world_id)
+
+    @router.patch("/worlds/{world_id}/source")
+    async def upload(
+        world_id: str, request: Request, upload_offset: int = Header(alias="Upload-Offset", ge=0)
+    ):
+        data = bytearray()
+        async for block in request.stream():
+            if len(data) + len(block) > CHUNK_BYTES:
+                raise HTTPException(413, "Фрагмент загрузки превышает 4 МБ.")
+            data.extend(block)
+        return invoke(worlds.append, world_id, upload_offset, bytes(data))
+
+    @router.post("/worlds/{world_id}/complete")
+    def complete(world_id: str):
+        return invoke(worlds.complete, world_id)
+
+    @router.put("/worlds/{world_id}/settings")
+    def settings(world_id: str, body: WorldSettings):
+        return invoke(worlds.configure, world_id, body)
+
+    @router.get("/worlds/{world_id}/source.ply")
+    def source(world_id: str):
+        row = invoke(worlds.get, world_id)
+        if row["status"] != "available":
+            raise HTTPException(409, "Импорт не завершён.")
+        if row.get("storage", {}).get("kind") == "worker":
+            raise HTTPException(409, "Исходник локации хранится на Worker.")
+        return FileResponse(
+            worlds.directory(world_id) / "source.ply",
+            media_type="application/octet-stream",
+            headers={
+                "ETag": f'"{row["sha256"]}"',
+                "Cache-Control": "private, max-age=31536000, immutable",
+            },
+        )
+
+    @router.post("/runs", status_code=201)
+    def start(body: RunCreate, idempotency_key: str = Header(alias="Idempotency-Key")):
+        if body.clock == "realtime":
+            graph = invoke(compose, adapters, body.composition)
+            body = body.model_copy(update={"composition": graph.selection_document()})
+        return invoke(runs.start, body, idempotency_key)
+
+    @router.get("/ai-modules")
+    def modules():
+        installed = invoke(registry, adapters)
+        graph = invoke(compose, adapters)
+        return {
+            "catalog": {**installed.catalog(), "authority": "virtual-only"},
+            "selection": graph.selection_document(),
+            "composition_sha256": graph.sha256,
+        }
+
+    @router.get("/runs/{run_id}")
+    def get_run(run_id: str):
+        runs.status()
+        return invoke(runs.get, run_id)
+
+    @router.post("/runs/{run_id}/{command}")
+    def control(run_id: str, command: Literal["pause", "play", "step", "stop"]):
+        return invoke(runs.control, run_id, command)
+
+    @router.get("/runs/{run_id}/frames/{sequence}.jpg")
+    def frame(run_id: str, sequence: int):
+        row = invoke(runs.get, run_id)
+        if not 0 <= sequence < row["samples"]:
+            raise HTTPException(404, "Кадр не найден.")
+        return FileResponse(
+            runs.directory(run_id) / "frames" / f"{sequence:06d}.jpg",
+            media_type="image/jpeg",
+            headers={"Cache-Control": "private, max-age=31536000, immutable"},
+        )
+
+    @router.put("/runs/{run_id}/view")
+    def view(run_id: str, body: ViewControl):
+        return invoke(runs.view, run_id, body.camera)
+
+    @router.get("/runs/{run_id}/decisions")
+    def decisions(run_id: str):
+        directory = invoke(runs.directory, run_id)
+        path = directory / "decisions.jsonl"
+        if not path.exists():
+            return Response("", media_type="application/x-ndjson")
+        return FileResponse(path, media_type="application/x-ndjson", filename="decisions.jsonl")
+
+    @router.post("/worker/register")
+    def register(body: WorkerHello, authorization: str | None = Header(default=None)):
+        authenticate(authorization)
+        return invoke(runs.register, body)
+
+    @router.post("/worker/worlds", status_code=201)
+    def register_worker_world(
+        body: WorkerWorldCreate,
+        authorization: str | None = Header(default=None),
+        instance_id: str = Header(alias="Worker-Instance"),
+    ):
+        authenticate(authorization)
+        with runs.lock:
+            invoke(runs.heartbeat, instance_id)
+            worker = runs.status()["worker"]
+            return invoke(worlds.register_worker_asset, body, worker["worker_id"])
+
+    @router.post("/worker/heartbeat")
+    def heartbeat(body: WorkerPoll, authorization: str | None = Header(default=None)):
+        authenticate(authorization)
+        return invoke(runs.heartbeat, body.instance_id)
+
+    @router.post("/worker/poll")
+    def poll(body: WorkerPoll, authorization: str | None = Header(default=None)):
+        authenticate(authorization)
+        return invoke(runs.poll, body.instance_id, body.run_id)
+
+    @router.post("/worker/runs/{run_id}/progress")
+    def progress(
+        run_id: str,
+        body: RunProgress,
+        instance_id: str = Header(alias="Worker-Instance"),
+        authorization: str | None = Header(default=None),
+    ):
+        authenticate(authorization)
+        return invoke(runs.progress, run_id, instance_id, body.phase)
+
+    @router.post("/worker/runs/{run_id}/samples")
+    def sample(
+        run_id: str,
+        body: RunSample,
+        instance_id: str = Header(alias="Worker-Instance"),
+        authorization: str | None = Header(default=None),
+    ):
+        authenticate(authorization)
+        return invoke(runs.sample, run_id, instance_id, body)
+
+    @router.post("/worker/runs/{run_id}/snapshot")
+    def snapshot(
+        run_id: str,
+        body: RealtimeSnapshot,
+        instance_id: str = Header(alias="Worker-Instance"),
+        authorization: str | None = Header(default=None),
+    ):
+        authenticate(authorization)
+        return invoke(runs.snapshot, run_id, instance_id, body)
+
+    @router.post("/worker/runs/{run_id}/finish")
+    def finish(run_id: str, body: WorkerResult, authorization: str | None = Header(default=None)):
+        authenticate(authorization)
+        return invoke(runs.finish, run_id, body.instance_id, body.outcome, body.message)
+
+    @router.post("/worker/runs/{run_id}/applied")
+    def applied(
+        run_id: str,
+        body: RunApplied,
+        instance_id: str = Header(alias="Worker-Instance"),
+        authorization: str | None = Header(default=None),
+    ):
+        authenticate(authorization)
+        return invoke(runs.applied, run_id, instance_id, body)
+
+    return router
diff --git a/src/k1link/web/app.py b/src/k1link/web/app.py
index 0fbc3ed..b4eb7d3 100644
--- a/src/k1link/web/app.py
+++ b/src/k1link/web/app.py
@@ -238,6 +238,7 @@ from k1link.missions.registration_runs import RegistrationRuns
 from k1link.web.mission_registration_api import build_mission_registration_router
 from k1link.web.mission_planner_api import build_mission_planner_router
 from k1link.web.simulation_projects_api import build_simulation_projects_router
+from k1link.web.ai_polygon_api import build_ai_polygon_router
 from k1link.web.simulation_world_provider_api import build_simulation_world_provider_router
 from k1link.web.system_telemetry_api import build_system_telemetry_router
 from k1link.web.vegetation_shadow_lab_api import (
@@ -1996,6 +1997,10 @@ app.include_router(
     )
 )
 frontend_dist = REPOSITORY_ROOT / "apps" / "control-station" / "dist"
+try:
+    app.include_router(build_ai_polygon_router(session_store.data_dir, OBSERVATORY_RECORDED_JOB_QUEUE))
+except (OSError, ValueError):
+    logging.getLogger(__name__).exception("AI polygon could not load its private runtime state")
 app.include_router(
     build_viewer_diagnostics_router(
         expected_ui_build_id=lambda: frontend_build_id(frontend_dist),
diff --git a/tests/test_ai_polygon.py b/tests/test_ai_polygon.py
new file mode 100644
index 0000000..57731c6
--- /dev/null
+++ b/tests/test_ai_polygon.py
@@ -0,0 +1,398 @@
+"""Synthetic protocol checks; these do not qualify rendering or AI model quality."""
+
+import base64
+import hashlib
+import io
+import json
+
+import numpy as np
+import pytest
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+from PIL import Image
+from pydantic import ValidationError
+from test_observatory_recorded_jobs import _definitions
+
+from k1link.observatory.recorded_jobs import ObservatoryRecordedJobQueue
+from k1link.simulation.ai_polygon.contracts import (
+    Decision,
+    RunApplied,
+    RunCreate,
+    RunSample,
+    WorkerHello,
+    WorkerWorldCreate,
+    WorldCreate,
+    WorldSettings,
+)
+from k1link.simulation.ai_polygon.policy import RoadPolicy
+from k1link.simulation.ai_polygon.runs import RunStore
+from k1link.simulation.ai_polygon.worlds import WorldStore, inspect_gaussian_ply
+from k1link.web.ai_polygon_api import build_ai_polygon_router
+
+
+def ply(value=0.0):
+    names = [
+        "x",
+        "y",
+        "z",
+        "opacity",
+        "f_dc_0",
+        "f_dc_1",
+        "f_dc_2",
+        "scale_0",
+        "scale_1",
+        "scale_2",
+        "rot_0",
+        "rot_1",
+        "rot_2",
+        "rot_3",
+    ]
+    header = "ply\nformat binary_little_endian 1.0\nelement vertex 1\n"
+    header += "".join(f"property float {name}\n" for name in names) + "end_header\n"
+    return header.encode() + np.full(14, value, dtype="<f4").tobytes()
+
+
+def make_world(worlds):
+    data = ply()
+    world = worlds.create(
+        WorldCreate(
+            name="Synthetic protocol fixture",
+            filename="fixture.PLY",
+            byte_length=len(data),
+            author="test",
+            license="CC0",
+        )
+    )
+    worlds.append(world["world_id"], 0, data)
+    worlds.complete(world["world_id"])
+    return worlds.configure(world["world_id"], WorldSettings(prepared=True))
+
+
+@pytest.fixture
+def runs(tmp_path):
+    queue = ObservatoryRecordedJobQueue(tmp_path, definitions=_definitions())
+    worlds = WorldStore(tmp_path)
+    store = RunStore(worlds, queue)
+    hello = WorkerHello(
+        worker_id="worker-006",
+        instance_id="a" * 32,
+        runtime="isaac-sim-6.1",
+        model_ids=["ddrnet-goose-pytorch-reference", "rf_detr_large"],
+        profile_sha256="b" * 64,
+        runtime_sources={key: "c" * 64 for key in ("worker", "scene", "models", "robot")},
+    )
+    store.register(hello)
+    return store, hello, make_world(worlds)
+
+
+def sample(sequence=0, **changes):
+    stream = io.BytesIO()
+    Image.new("RGB", (800, 600)).save(stream, "JPEG")
+    row = dict(
+        sequence=sequence,
+        simulation_time_ns=sequence * 100_000_000,
+        inference_ms=5,
+        pose_xy=(0, 0),
+        decision=Decision(
+            speed_mps=0, yaw_rate_rps=0, reason="no-road", road_fraction=0, obstacle_count=0
+        ),
+        image_jpeg_base64=base64.b64encode(stream.getvalue()).decode(),
+    )
+    return RunSample(**{**row, **changes})
+
+
+def test_resume_and_atomic_publication_recovery(tmp_path):
+    worlds = WorldStore(tmp_path)
+    data = ply()
+    row = worlds.create(
+        WorldCreate(
+            name=" Test ", filename="scan.ply", byte_length=len(data), author="Test", license="CC0"
+        )
+    )
+    key = row["world_id"]
+    worlds.append(key, 0, data[:20])
+    assert worlds.prefix_hashes(key)["chunks"] == [
+        {"byte_length": 20, "sha256": hashlib.sha256(data[:20]).hexdigest()}
+    ]
+    with pytest.raises(RuntimeError):
+        worlds.append(key, 0, data[:20])
+    with pytest.raises(RuntimeError):
+        worlds.complete(key)
+    worlds.append(key, 20, data[20:])
+    directory = worlds.directory(key)
+    (directory / "source.part").replace(directory / "source.ply")
+    recovered = WorldStore(tmp_path).complete(key)
+    assert recovered["status"] == "available"
+    assert recovered["sha256"] == hashlib.sha256(data).hexdigest()
+    assert recovered["name"] == "Test"
+    assert worlds.complete(key) == recovered
+
+
+@pytest.mark.parametrize(
+    "data", [b"not ply", ply()[:-1], ply() + b"xxxx", ply(float("nan")), ply(float("inf"))]
+)
+def test_invalid_ply_rejected(tmp_path, data):
+    path = tmp_path / "bad.ply"
+    path.write_bytes(data)
+    with pytest.raises(ValueError):
+        inspect_gaussian_ply(path)
+
+
+@pytest.mark.parametrize(
+    "settings",
+    [{"spawn_xy": [float("nan"), 0]}, {"rotation_degrees": [0, 0, 900]}, {"max_speed_mps": 2}],
+)
+def test_nonfinite_or_out_of_bounds_settings_rejected(settings):
+    with pytest.raises(ValidationError):
+        WorldSettings(**settings)
+
+
+def test_start_snapshots_and_idempotency(runs):
+    store, hello, world = runs
+    request = RunCreate(world_id=world["world_id"], max_steps=2)
+    row = store.start(request, "request-001")
+    assert store.start(request, "request-001") == row
+    store.worlds.configure(world["world_id"], WorldSettings(max_speed_mps=0.5, prepared=True))
+    assert store.get(row["run_id"])["world"]["settings"]["max_speed_mps"] == 0.3
+    with pytest.raises(RuntimeError):
+        store.start(RunCreate(world_id=world["world_id"], max_steps=3), "request-001")
+    with pytest.raises(RuntimeError):
+        store.start(request, "request-002")
+    assert store.poll(hello.instance_id, None)["action"] == "load"
+
+
+def test_sequence_clock_pause_step_stop_and_resource_release(runs):
+    store, hello, world = runs
+    key = store.start(RunCreate(world_id=world["world_id"], max_steps=2), "request-001")["run_id"]
+    with pytest.raises(RuntimeError):
+        store.control(key, "step")
+    store.poll(hello.instance_id, key)
+    store.sample(key, hello.instance_id, sample())
+    with pytest.raises(RuntimeError):
+        store.sample(key, hello.instance_id, sample())
+    store.applied(
+        key,
+        hello.instance_id,
+        RunApplied(sequence=0, simulation_time_ns=100_000_000, physics_steps=6, pose_xy=(0, 0)),
+    )
+    with pytest.raises(ValueError):
+        store.sample(key, hello.instance_id, sample(1, simulation_time_ns=999))
+    store.control(key, "pause")
+    assert store.get(key)["state"] == "running"
+    assert store.poll(hello.instance_id, key)["run"]["state"] == "paused"
+    store.control(key, "step")
+    assert store.poll(hello.instance_id, key)["action"] == "step"
+    assert store.poll(hello.instance_id, key)["action"] == "pause"
+    store.control(key, "stop")
+    with pytest.raises(RuntimeError):
+        store.sample(key, hello.instance_id, sample(1))
+    store.finish(key, hello.instance_id, "stopped", "")
+    store.queue.reserve_simulation("airun-" + "b" * 32)
+
+
+def test_expired_worker_stops_run_but_does_not_claim_gpu_released(runs):
+    store, hello, world = runs
+    key = store.start(RunCreate(world_id=world["world_id"]), "request-001")["run_id"]
+    store.seen -= 21
+    assert not store.status()["available"]
+    assert store.get(key)["state"] == "failed"
+    with pytest.raises(RuntimeError):
+        store.sample(key, hello.instance_id, sample())
+    with pytest.raises(RuntimeError):
+        store.queue.reserve_simulation("airun-" + "b" * 32)
+    store.register(hello)
+    assert store.finish(key, hello.instance_id, "failed", "")["state"] == "failed"
+    store.queue.reserve_simulation("airun-" + "b" * 32)
+
+
+def test_restart_fences_worker_and_preserves_journal(runs):
+    store, hello, world = runs
+    key = store.start(RunCreate(world_id=world["world_id"], max_steps=2), "request-001")["run_id"]
+    store.poll(hello.instance_id, key)
+    store.sample(key, hello.instance_id, sample())
+    restarted = RunStore(store.worlds, store.queue)
+    assert restarted.get(key)["state"] == "failed"
+    with pytest.raises(RuntimeError):
+        restarted.sample(key, hello.instance_id, sample(1))
+    saved = json.loads((store.directory(key) / "decisions.jsonl").read_text())
+    assert (
+        saved["image_sha256"]
+        == hashlib.sha256((store.directory(key) / "frames/000000.jpg").read_bytes()).hexdigest()
+    )
+    assert "image_jpeg_base64" not in saved
+
+
+def test_completion_requires_expected_steps(runs):
+    store, hello, world = runs
+    key = store.start(RunCreate(world_id=world["world_id"], max_steps=1), "request-001")["run_id"]
+    with pytest.raises(RuntimeError):
+        store.finish(key, hello.instance_id, "completed", "")
+    store.poll(hello.instance_id, key)
+    store.sample(key, hello.instance_id, sample())
+    with pytest.raises(RuntimeError):
+        store.finish(key, hello.instance_id, "completed", "")
+    store.applied(
+        key,
+        hello.instance_id,
+        RunApplied(sequence=0, simulation_time_ns=100_000_000, physics_steps=6, pose_xy=(0, 0)),
+    )
+    assert store.finish(key, hello.instance_id, "completed", "")["state"] == "completed"
+
+
+def test_api_auth_and_bounded_upload(tmp_path):
+    app = FastAPI()
+    app.include_router(build_ai_polygon_router(tmp_path))
+    with TestClient(app) as client:
+        root = "/api/v1/ai-polygon"
+        assert client.get(root + "/catalog").json()["runtime"]["available"] is False
+        assert client.post(root + "/worker/poll", json={"instance_id": "a" * 32}).status_code == 401
+        assert client.post(root + "/worlds", json={"filename": "../bad.ply"}).status_code == 422
+        world = make_world(WorldStore(tmp_path))
+        response = client.get(root + f"/worlds/{world['world_id']}/source.ply")
+        assert response.status_code == 200 and response.content == ply()
+        assert (
+            client.post(
+                root + "/runs",
+                headers={"Idempotency-Key": "test-run-001"},
+                json={"world_id": world["world_id"]},
+            ).status_code
+            == 409
+        )
+        assert not any((tmp_path / "ai-polygon/runs").iterdir())
+
+
+def test_camera_policy_brakes_and_waits_before_resume():
+    policy = RoadPolicy(0.3)
+    road = np.ones((512, 512), dtype=bool)
+    assert policy.decide(road, []).speed_mps == 0
+    policy.decide(road, [])
+    assert policy.decide(road, []).speed_mps == 0.3
+    obstacle = policy.decide(road, [(0.4, 0.2, 0.6, 0.8)])
+    assert obstacle.reason == "obstacle" and obstacle.speed_mps == 0
+    assert policy.decide(road, []).speed_mps == 0
+    assert policy.decide(np.zeros_like(road), []).reason == "no-road"
+    with pytest.raises(ValueError):
+        policy.decide(road, [(float("nan"), 0, 1, 1)])
+
+
+def worker_asset():
+    return WorkerWorldCreate(
+        name="Synthetic paired scene",
+        filename="fixture.ply",
+        byte_length=len(ply()),
+        author="test",
+        license="CC0",
+        sha256="a" * 64,
+        collider_sha256="b" * 64,
+        splat_count=1,
+        settings=WorldSettings(prepared=True),
+    )
+
+
+def test_worker_asset_is_manifest_only_and_retry_preserves_settings(tmp_path):
+    worlds = WorldStore(tmp_path)
+    request = worker_asset()
+    row = worlds.register_worker_asset(request, "worker-006")
+    assert {p.name for p in worlds.directory(row["world_id"]).iterdir()} == {"world.json"}
+    changed = worlds.configure(row["world_id"], WorldSettings(spawn_xy=(3, 4), prepared=True))
+    assert worlds.register_worker_asset(request, "worker-006") == changed
+    with pytest.raises(RuntimeError):
+        worlds.register_worker_asset(
+            request.model_copy(update={"collider_sha256": "c" * 64}), "worker-006"
+        )
+    with pytest.raises(RuntimeError):
+        worlds.prefix_hashes(row["world_id"])
+
+
+def test_worker_asset_api_requires_current_authenticated_instance(tmp_path):
+    app = FastAPI()
+    app.include_router(build_ai_polygon_router(tmp_path))
+    base = "/api/v1/ai-polygon"
+    store = RunStore(WorldStore(tmp_path))
+    hello = WorkerHello(
+        worker_id="worker-006",
+        instance_id="a" * 32,
+        runtime="isaac-sim-6.1",
+        model_ids=["test"],
+        runtime_sources={k: "c" * 64 for k in ("worker", "scene", "models", "robot")},
+        profile_sha256="b" * 64,
+    )
+    auth = {"Authorization": "Bearer " + store.token, "Worker-Instance": hello.instance_id}
+    body = worker_asset().model_dump(mode="json")
+    with TestClient(app) as client:
+        assert (
+            client.post(
+                base + "/worker/worlds", json=body, headers={"Worker-Instance": hello.instance_id}
+            ).status_code
+            == 401
+        )
+        assert client.post(base + "/worker/worlds", json=body, headers=auth).status_code == 409
+        assert (
+            client.post(
+                base + "/worker/register", json=hello.model_dump(), headers=auth
+            ).status_code
+            == 200
+        )
+        response = client.post(base + "/worker/worlds", json=body, headers=auth)
+        assert response.status_code == 201
+        row = response.json()
+        assert row["storage"] == {"kind": "worker", "worker_id": "worker-006"}
+        assert client.get(base + f"/worlds/{row['world_id']}/source.ply").status_code == 409
+        assert (
+            client.post(
+                base + "/worker/worlds", json=body, headers={**auth, "Worker-Instance": "d" * 32}
+            ).status_code
+            == 409
+        )
+
+
+def test_camera_policy_turns_towards_connected_road():
+    policy = RoadPolicy(0.3)
+    road = np.zeros((512, 512), dtype=bool)
+    road[:, 100:260] = True
+    policy.decide(road, [])
+    policy.decide(road, [])
+    decision = policy.decide(road, [])
+    assert decision.yaw_rate_rps > 0 and 0 < decision.speed_mps < 0.3
+
+
+def test_progress_does_not_claim_camera_ready(runs):
+    store, hello, world = runs
+    run = store.start(RunCreate(world_id=world["world_id"]), "progress-case-001")
+    store.progress(run["run_id"], hello.instance_id, "scene")
+    assert store.get(run["run_id"])["phase"] == "scene"
+    assert store.get(run["run_id"])["state"] == "starting"
+    store.control(run["run_id"], "stop")
+    assert store.progress(run["run_id"], hello.instance_id, "models")["control"] == "stop"
+    assert store.get(run["run_id"])["state"] == "stopping"
+
+
+def test_worker_stop_race_accepts_already_exited_process(monkeypatch):
+    import importlib.util
+    from pathlib import Path
+
+    runtime = Path(__file__).resolve().parents[1] / "simulation/ai-polygon"
+    monkeypatch.syspath_prepend(str(runtime))
+    spec = importlib.util.spec_from_file_location("polygon_worker_stop_test", runtime / "worker.py")
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    calls = []
+
+    class Child:
+        pid = 123
+
+        def poll(self):
+            return None
+
+        def wait(self, timeout):
+            calls.append(("wait", timeout))
+            return 0
+
+    def taskkill(*args, **kwargs):
+        # Windows reports a non-zero taskkill when the process exits in the race.
+        assert kwargs["check"] is False
+        calls.append(("taskkill", 255))
+
+    monkeypatch.setattr(module.subprocess, "run", taskkill)
+    module.terminate_episode(Child())
+    assert calls == [("taskkill", 255), ("wait", 30)]
diff --git a/tests/test_ai_polygon_mission.py b/tests/test_ai_polygon_mission.py
new file mode 100644
index 0000000..880fbcf
--- /dev/null
+++ b/tests/test_ai_polygon_mission.py
@@ -0,0 +1,141 @@
+import numpy as np
+import pytest
+
+from k1link.simulation.ai_polygon.contracts import WorldSettings
+from k1link.simulation.ai_polygon.mission_policy import WaypointMission, inclination
+
+
+def pose(x=0, y=0):
+    return [x, y, 0.37, 0, 0, 0, 1]
+
+
+def goal(target, prior, excluded):
+    return [2, 0.7 * len(excluded), 0]
+
+
+def test_chassis_jitter_cannot_hide_stall_and_retries_are_bounded():
+    mission = WaypointMission([[8, 0]])
+    states = []
+    for t in np.arange(0, 34, 0.2):
+        _, intent = mission.update(pose(0.01 * np.sin(t * 10)), float(t), goal)
+        states.append(intent["state"])
+    assert "replanning" in states
+    assert states[-1] == "stuck"
+    assert mission.update(pose(1), 35, goal)[1]["state"] == "stuck"
+
+
+def test_route_cursor_and_completion_survive_pause_without_restarting_task():
+    mission = WaypointMission([[1, 0], [3, 0]])
+    mission.update(pose(), 0, goal)
+    assert mission.update(pose(1), 6, goal)[1]["waypoint"] == 1
+    mission.resume()
+    assert mission.update(pose(1), 6, goal)[1]["waypoint"] == 1
+    assert mission.update(pose(3), 18, goal)[1]["state"] == "goal-reached"
+    mission.resume()
+    assert mission.update(pose(3), 18, goal)[1]["state"] == "goal-reached"
+
+
+def test_overturned_and_excessive_tilt_are_latched_before_goal_success():
+    mission = WaypointMission([[0, 0]])
+    overturned = [0, 0, 0.37, 1, 0, 0, 0]
+    assert inclination(overturned) == pytest.approx(180)
+    assert mission.update(overturned, 0, goal)[1]["state"] == "unstable"
+    mission.resume()
+    assert mission.update(pose(), 1, goal)[1]["state"] == "unstable"
+
+
+def test_missing_surface_never_becomes_a_drive_permission_or_unbounded_wait():
+    mission = WaypointMission([[8, 0]])
+    for t in range(34):
+        target, intent = mission.update(pose(), t, lambda *_: None)
+        assert target is None
+    assert intent["state"] == "stuck"
+
+
+def test_route_contract_bounds_and_finiteness():
+    from pydantic import ValidationError
+
+    for route in ([[float("nan"), 0]], [[0, 0]] * 33, [[0, 10001]]):
+        with pytest.raises(ValidationError):
+            WorldSettings(route_xy=route)
+
+
+def test_recovery_uses_observed_goal_and_does_not_count_retreat_as_progress():
+    mission = WaypointMission([[8, 0]])
+
+    def retreat(prior):
+        return prior or [-0.65, 0, 0]
+
+    mission.update(pose(), 0, goal, retreat)
+    target, intent = mission.update(pose(), 8, goal, retreat)
+    assert target == [-0.65, 0, 0] and intent["state"] == "reversing"
+    assert mission.update(pose(-0.2), 10, goal, retreat)[1]["state"] == "reversing"
+    assert mission.update(pose(-0.4), 12, goal, retreat)[1]["state"] == "replanning"
+    mission.update(pose(), 16, goal, retreat)
+    assert mission.attempts == 1
+    # Moving back to where we started cannot create a fresh three-attempt budget.
+    assert mission.update(pose(), 20, goal, retreat)[1]["recovery_attempt"] == 2
+
+
+def test_recovery_stops_immediately_on_lost_support_and_times_out_without_motion():
+    mission = WaypointMission([[8, 0]])
+
+    def retreat(prior):
+        return prior or [-0.65, 0, 0]
+
+    mission.update(pose(), 0, goal, retreat)
+    mission.update(pose(), 8, goal, retreat)
+    assert mission.update(pose(), 8.2, goal, lambda _: None)[0] is None
+    assert mission.recovery_goal is None
+    for seconds in (17, 23, 32, 38, 47):
+        _, intent = mission.update(pose(), seconds, goal, retreat)
+    assert intent["state"] == "stuck"
+
+
+def test_wrong_direction_never_replenishes_route_attempts():
+    mission = WaypointMission([[8, 0]])
+    for seconds in range(34):
+        _, intent = mission.update(pose(-seconds * 0.1), seconds, goal)
+    assert intent["state"] == "stuck"
+
+
+def test_signed_motion_contract_accepts_reverse_but_stays_bounded():
+    from pydantic import ValidationError
+
+    from k1link.simulation.ai_polygon.contracts import Decision
+
+    assert (
+        Decision(
+            speed_mps=-0.1, yaw_rate_rps=0, reason="replanning", road_fraction=0, obstacle_count=0
+        ).speed_mps
+        < 0
+    )
+    with pytest.raises(ValidationError):
+        Decision(speed_mps=-1.1, yaw_rate_rps=0, reason="road", road_fraction=1, obstacle_count=0)
+
+
+def test_slow_regulated_progress_is_not_mistaken_for_stall():
+    mission = WaypointMission([[8, 0]])
+    for seconds in range(60):
+        _, intent = mission.update(pose(seconds * 0.02), seconds, goal)
+        assert intent["state"] == "following"
+        assert intent["recovery_attempt"] == 0
+
+
+def test_rejected_observation_stops_but_does_not_forget_revalidated_goal():
+    mission = WaypointMission([[2, 0]])
+    observed = [2, 0, 0]
+    assert mission.update(pose(), 0, lambda *_: observed)[0] == observed
+    assert mission.update(pose(1), 1, lambda *_: None)[0] is None
+    assert mission.goal == observed
+    mission.resume()
+    calls = []
+
+    def revalidate(_target, prior, _excluded):
+        calls.append(prior)
+        return prior
+
+    assert mission.update(pose(1.3), 1.2, revalidate)[0] == observed
+    assert calls == [observed]
+    # Memory alone never authorizes a command when the new frame is rejected.
+    assert mission.update(pose(1.3), 1.4, lambda *_: None)[0] is None
diff --git a/tests/test_ai_polygon_motion.py b/tests/test_ai_polygon_motion.py
new file mode 100644
index 0000000..868dd45
--- /dev/null
+++ b/tests/test_ai_polygon_motion.py
@@ -0,0 +1,40 @@
+"""The actuator may soften acceleration but must never prolong a safety command."""
+
+import importlib.util
+from pathlib import Path
+
+import pytest
+
+spec = importlib.util.spec_from_file_location(
+    "motion_control", Path(__file__).parents[1] / "simulation/ai-polygon/motion_control.py"
+)
+module = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(module)
+
+
+def test_slow_perception_keeps_continuous_physics_command():
+    drive = module.DriveEnvelope()
+    rows = [drive.step(0.15, 0, 1 / 60)[0] for _ in range(180)]
+    assert rows[0] == pytest.approx(0.2 / 60)
+    assert all(0 <= b - a <= 0.2 / 60 + 1e-9 for a, b in zip(rows, rows[1:], strict=False))
+    assert rows[44:] == pytest.approx([0.15] * 136)
+
+
+def test_braking_and_deadman_bypass_ramp():
+    drive = module.DriveEnvelope()
+    for _ in range(60):
+        drive.step(0.15, 0, 1 / 60)
+    assert drive.step(0.03, 0, 1 / 60) == (0.03, 0)
+    assert drive.step(0.15, 0.2, 1 / 60, stop=True) == (0, 0)
+    assert drive.wheels == (0, 0)
+    assert drive.step(0.15, 0, 1 / 60)[0] == pytest.approx(0.2 / 60)
+
+
+def test_curvature_and_direction_change():
+    drive = module.DriveEnvelope()
+    for _ in range(60):
+        speed, yaw = drive.step(0.15, 0.2, 1 / 60)
+        assert yaw / speed == pytest.approx(0.2 / 0.15)
+    speed, yaw = drive.step(-0.15, -0.2, 1 / 60)
+    assert speed < 0 and yaw < 0
+    assert max(map(abs, drive.wheels)) <= 0.2 / 60 + 1e-9
diff --git a/tests/test_ai_polygon_navigation.py b/tests/test_ai_polygon_navigation.py
new file mode 100644
index 0000000..3ab5140
--- /dev/null
+++ b/tests/test_ai_polygon_navigation.py
@@ -0,0 +1,390 @@
+import importlib.util
+import subprocess
+import sys
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+from k1link.observatory.modular_composition import CompositionError
+from k1link.simulation.ai_polygon.composition import compose, registry
+from k1link.simulation.ai_polygon.terrain_contract import terrain_matches
+
+ROOT = Path(__file__).resolve().parents[1] / "simulation/ai-polygon"
+
+
+def module(name):
+    spec = importlib.util.spec_from_file_location(name, ROOT / (name + ".py"))
+    result = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(result)
+    return result
+
+
+def test_simulation_composition_has_causal_dependencies_and_separate_authority():
+    graph = compose(ROOT)
+    assert graph.as_dict()["execution"]["mode"] == "worker-local-simulation"
+    motion = graph.nodes[-1]
+    assert motion.module.group == "motion"
+    assert dict(motion.inputs)["segmentation.surface"] == "simulation-segformer-ade"
+    assert dict(motion.inputs)["detection.boxes"] == "simulation-rf-detr"
+    assert motion.module.state_policy == "causal-reset-at-source-start"
+    selection = graph.selection_document()
+    selection["selections"] = [r for r in selection["selections"] if r["group"] != "segmentation"]
+    with pytest.raises(CompositionError, match="segmentation.surface"):
+        compose(ROOT, selection)
+
+
+def test_surface_providers_are_interchangeable_in_the_shared_constructor():
+    selection = compose(ROOT).selection_document()
+    reference = next(m for m in registry(ROOT).modules if m.module_id == "simulation-ddrnet-goose")
+    for row in selection["selections"]:
+        if row["group"] == "segmentation":
+            row.update(module_id=reference.module_id, module_sha256=reference.sha256)
+    graph = compose(ROOT, selection)
+    assert dict(graph.nodes[-1].inputs)["segmentation.surface"] == reference.module_id
+
+
+def test_composition_rejects_stale_module_identity():
+    selection = compose(ROOT).selection_document()
+    selection["selections"][0]["module_sha256"] = "0" * 64
+    with pytest.raises(CompositionError, match="not installed"):
+        compose(ROOT, selection)
+
+
+def test_shared_constructor_import_needs_no_core_or_third_party_runtime():
+    # -S removes site-packages, as in the minimal Windows coordinator. Loading
+    # a contract must not load the POSIX-only artifact gateway through __init__.
+    source = str(ROOT.parents[1] / "src")
+    result = subprocess.run(
+        [
+            sys.executable,
+            "-S",
+            "-c",
+            f"import sys; sys.path.insert(0, {source!r}); "
+            "from k1link.observatory.modular_composition import ModuleRegistry; "
+            "from k1link.simulation.ai_polygon.composition import compose; "
+            "assert 'k1link.artifact_gateway' not in sys.modules",
+        ],
+        capture_output=True,
+        text=True,
+    )
+    assert result.returncode == 0, result.stderr
+
+
+def test_semantic_goal_uses_range_and_correct_square_camera_crop():
+    nav = module("navigation_client")
+    points = np.array(
+        [[x, y, 0] for x in np.linspace(1.5, 3, 20) for y in np.linspace(-0.3, 0.3, 9)],
+        dtype=np.float32,
+    )
+    pose = [0, 0, 0.27, 0, 0, 0, 1]
+    calibration = {
+        "origin": [0.38, 0, 0.8],
+        "rotation": np.eye(3).reshape(-1).tolist(),
+        "intrinsics": [800 * 24 / 36, 800 * 24 / 36, 400, 300],
+    }
+    leaves = np.ones((512, 512), dtype=bool)
+    goal = nav.visual_goal(leaves, points, pose, calibration)
+    assert goal is not None and 1.5 < goal[0] < 3 and abs(goal[1]) < 0.3
+    assert nav.visual_goal(np.zeros_like(leaves), points, pose, calibration) is None
+    assert nav.visual_goal(leaves, points + [0, 0, 2], pose, calibration) is None
+    # A previously valid goal cannot authorize motion through newly unknown RGB.
+    assert nav.visual_goal(np.zeros_like(leaves), points, pose, calibration, goal) is None
+    # An explicit-route waypoint entering the camera blind strip is retained,
+    # but losing all current visual surface support still forbids movement.
+    close = [0.65, 0, 0]
+    assert nav.visual_goal(leaves, points, pose, calibration, close, target=[0.65, 0]) == close
+    assert (
+        nav.visual_goal(np.zeros_like(leaves), points, pose, calibration, close, target=[0.65, 0])
+        is None
+    )
+
+
+def test_ground_placement_uses_actual_triangle_intersection():
+    terrain = module("terrain")
+    vertices = np.array([[0, 0, 0], [1, 0, 0.2], [0, 1, 0]], dtype=np.float32)
+    faces = np.array([[0, 1, 2]], dtype=np.int32)
+    assert terrain.ground_intersections(vertices, faces, 0.25, 0.25)[0] == pytest.approx(0.05)
+    assert len(terrain.ground_intersections(vertices, faces, 0.9, 0.9)) == 0
+
+
+def test_observed_route_goal_keeps_task_position_and_cannot_run_away_from_it():
+    choose = module("navigation_client").visual_goal
+    points = np.array(
+        [[x, y, 0] for x in np.arange(1.2, 3.1, 0.05) for y in np.arange(-0.5, 0.51, 0.05)]
+    )
+    pose = [0, 0, 0.37, 0, 0, 0, 1]
+    calibration = {
+        "origin": [0.38, 0, 0.8],
+        "rotation": np.eye(3).reshape(-1).tolist(),
+        "intrinsics": [800 * 24 / 36, 800 * 24 / 36, 400, 300],
+        "body_contact_height_m": 0.37,
+    }
+    surface = np.ones((512, 512), dtype=bool)
+    target = [2.03, 0.27]
+    assert choose(surface, points, pose, calibration, target=target) == pytest.approx([*target, 0])
+    # The close, already observed waypoint can enter the camera blind strip.
+    advanced = [1.5, 0, 0.37, 0, 0, 0, 1]
+    camera = {**calibration, "origin": [1.88, 0, 0.8]}
+    prior = [*target, 0]
+    assert choose(surface, points + [1.5, 0, 0], advanced, camera, prior, target) == prior
+    # The actual camera loses nearby ground beyond the old hardcoded 0.8 m.
+    advanced = [1.1, 0, 0.37, 0, 0, 0, 1]
+    camera = {**calibration, "origin": [1.48, 0, 0.8]}
+    assert choose(surface, points + [1.1, 0, 0], advanced, camera, prior, target) == prior
+    assert choose(np.zeros_like(surface), points, pose, calibration, target=target) is None
+    # Clear road ahead is not permission to drive away from a missed waypoint.
+    assert choose(surface, points, pose, calibration, target=[-1, 0]) is None
+    # A distant task may still use an observed local goal towards it.
+    far = choose(surface, points, pose, calibration, target=[8, 0])
+    assert far is not None and 1.2 <= far[0] <= 3.1
+
+
+def test_collision_identity_follows_geometry_and_tile_coverage_not_camera_or_start():
+    settings = dict(
+        meters_per_unit=1,
+        rotation_degrees=[-90, 0, 180],
+        spawn_xy=[0, 0],
+        ground_z=0,
+        camera_height_m=0.8,
+        max_speed_mps=0.15,
+    )
+    terrain = dict(source_sha256="a" * 64, generator_sha256="b" * 64, settings=settings)
+    world = dict(sha256="a" * 64, settings={**settings, "spawn_xy": [1, 1], "camera_height_m": 1})
+    assert terrain_matches(terrain, world, "b" * 64)
+    assert not terrain_matches(terrain, world, "c" * 64)
+    assert not terrain_matches(terrain, {**world, "sha256": "d" * 64})
+    for change in ({"spawn_xy": [20, 0]}, {"meters_per_unit": 2}, {"rotation_degrees": [0, 0, 0]}):
+        assert not terrain_matches(terrain, {**world, "settings": {**world["settings"], **change}})
+
+
+def test_paired_full_scene_does_not_inherit_generated_tile_bounds():
+    settings = dict(meters_per_unit=1, rotation_degrees=[90, 0, 0], spawn_xy=[0, 0], ground_z=5)
+    terrain = dict(
+        generator="paired-source",
+        source_sha256="a" * 64,
+        collider_sha256="b" * 64,
+        settings=settings,
+    )
+    world = dict(
+        sha256="a" * 64,
+        collider_sha256="b" * 64,
+        settings={**settings, "spawn_xy": [210, 30], "ground_z": 1.5},
+    )
+    assert terrain_matches(terrain, world)
+    assert not terrain_matches(terrain, {**world, "collider_sha256": "c" * 64})
+    assert not terrain_matches(
+        terrain, {**world, "settings": {**world["settings"], "meters_per_unit": 2}}
+    )
+    assert not terrain_matches({**terrain, "generator": "generated-tile"}, world)
+
+
+def test_square_footprint_fits_straight_corridor_and_rejects_corner_sweep():
+    check = module("navigation/footprint").swept_footprint_clear
+    pose = [0, 0, 0.27, 0, 0, 0, 1]
+    path = [[0, 0, 0], [0.5, 0, 0], [1, 0, 0]]
+    walls = np.array([[x, y, 0.5, 0.5] for x in np.arange(-1, 2, 0.1) for y in [-0.65, 0.65]])
+    assert check(path, walls, pose)
+    walls[:, 1] *= 0.45 / 0.65
+    assert not check(path, walls, pose)
+    # A diagonal turn sweeps a square corner into this obstacle, even though
+    # the chassis at its initial and final straight poses does not contain it.
+    assert not check([[0, 0, 0], [0.5, 0.5, 0]], [[0.7, 0, 0.5, 0.5]], pose)
+
+
+def test_command_monitor_covers_deadman_braking_distance_and_rotation():
+    check = module("navigation/footprint").command_footprint_clear
+    pose = [0, 0, 0.27, 0, 0, 0, 1]
+    assert check(0.15, 0, [[1.5, 0, 0.5, 0.5]], pose)
+    assert not check(0.15, 0, [[0.7, 0, 0.5, 0.5]], pose)
+    assert not check(0, 0.8, [[0.7, 0, 0.5, 0.5]], pose)
+
+
+def test_smooth_slope_is_distinct_from_a_step_or_vertical_terrain():
+    costs = module("navigation/terrain_costs").supported_slope_costs
+    xy = np.array([[x, y] for x in np.arange(-1, 1.01, 0.1) for y in np.arange(-1, 1.01, 0.1)])
+    slope = np.column_stack((xy, xy[:, 0] * np.tan(np.radians(20)), np.full(len(xy), 0.15)))
+    normalized, corrected = costs(slope)
+    assert corrected > len(slope) * 0.9
+    assert normalized[len(slope) // 2, 3] == 0
+    assert np.array_equal(normalized[:, :3], slope[:, :3])
+    # A 15 cm ledge across the initial footprint cannot become a traversable ramp.
+    step = slope.copy()
+    step[:, 2] = np.where(step[:, 0] >= 0, 0.15, 0)
+    assert costs(step)[1] == 0
+    cliff = slope.copy()
+    cliff[:, 2] = np.where(cliff[:, 0] >= 0, -0.4, 0)
+    assert costs(cliff)[1] == 0
+    steep = slope.copy()
+    steep[:, 2] = steep[:, 0] * np.tan(np.radians(35))
+    assert costs(steep)[1] == 0
+    assert costs(slope[np.abs(slope[:, 1]) < 0.01])[1] == 0  # Unobserved lateral support.
+
+
+def test_underbody_support_does_not_clear_future_terrain_walls_or_drops():
+    correct = module("navigation/terrain_costs").underbody_support_costs
+    terrain = np.array(
+        [
+            [-0.375, -0.28, 0.066, 0.103],  # Low return already under the chassis.
+            [0.46, 0, 0.066, 0.103],  # Inset excludes the leading edge.
+            [0.75, 0, 0.066, 0.103],  # Never change future terrain from body pose.
+            [0, 0, 0.12, 0.12],  # A real step within the footprint remains blocked.
+            [0, 0, 0.5, 0.5],
+            [0, 0, -0.4, 0.4],
+        ]
+    )
+    original = terrain.copy()
+    result, count = correct(terrain, [0, 0, 0.37, 0, 0, 0, 1])
+    assert count == 1 and result[0, 3] == pytest.approx(0.066)
+    assert np.array_equal(result[1:], original[1:])
+    assert np.array_equal(terrain, original)  # Never erase the causal raw map.
+    # Rotate both observations and the measured chassis; the result must agree.
+    yaw = np.pi / 2
+    rotated = terrain.copy()
+    rotated[:, :2] = terrain[:, :2] @ np.array([[0, 1], [-1, 0]]) + [3, 4]
+    pose = [3, 4, 0.37, 0, 0, np.sin(yaw / 2), np.cos(yaw / 2)]
+    assert np.allclose(correct(rotated, pose)[0][:, 3], result[:, 3])
+    assert correct(terrain, [0, 0, 0.37, 0, np.sin(np.pi / 12), 0, np.cos(np.pi / 12)])[1] == 0
+
+
+def test_retreat_requires_observed_full_width_support_and_no_drop_or_step():
+    choose = module("navigation_client").recovery_goal
+    points = np.array(
+        [[x, y, 0.0] for x in np.arange(-1.5, -0.39, 0.05) for y in np.arange(-0.85, 0.86, 0.05)]
+    )
+    pose = [0, 0, 0.37, 0, 0, 0, 1]
+    calibration = {"body_contact_height_m": 0.37}
+    assert choose(points, pose, calibration) == pytest.approx([-0.65, 0, 0])
+    assert choose(points[points[:, 1] > -0.2], pose, calibration) is None
+    assert choose(points[points[:, 0] < -0.9], pose, calibration) is None
+    for height in (-0.4, 0.15):
+        discontinuous = points.copy()
+        discontinuous[points[:, 0] < -0.9, 2] = height
+        assert choose(discontinuous, pose, calibration) is None
+    slope = points.copy()
+    slope[:, 2] = slope[:, 0] * np.tan(np.radians(10))
+    assert choose(slope, pose, calibration) is not None
+    assert choose(points, pose, calibration, [-0.65, 0.3, 0]) is None
+
+
+def test_reverse_monitor_checks_behind_the_body():
+    check = module("navigation/footprint").command_footprint_clear
+    pose = [0, 0, 0.37, 0, 0, 0, 1]
+    assert check(-0.1, 0, [[-1.5, 0, 0.5, 0.5]], pose)
+    assert not check(-0.1, 0, [[-0.65, 0, 0.5, 0.5]], pose)
+
+
+def test_legacy_coordinate_migration_is_an_exact_rigid_rotation():
+    import json
+    import struct
+
+    migrate = module("navigation/migrate_terrain_coordinates")
+    positions = [[1.0, -2.0, -3.0], [2.0, -2.0, -3.0], [1.0, -1.0, -3.0]]
+    binary = struct.pack("<9f3I", *(v for p in positions for v in p), 0, 1, 2)
+    document = {
+        "nodes": [{"mesh": 0}],
+        "meshes": [{"primitives": [{"attributes": {"POSITION": 0}, "indices": 1}]}],
+        "accessors": [
+            {
+                "bufferView": 0,
+                "componentType": 5126,
+                "type": "VEC3",
+                "count": 3,
+                "min": [1, -2, -3],
+                "max": [2, -1, -3],
+            },
+            {"bufferView": 1, "componentType": 5125, "type": "SCALAR", "count": 3},
+        ],
+        "bufferViews": [{"byteOffset": 0, "byteLength": 36}, {"byteOffset": 36, "byteLength": 12}],
+    }
+    raw = json.dumps(document).encode()
+    raw += b" " * ((-len(raw)) % 4)
+    glb = (
+        struct.pack("<III", 0x46546C67, 2, 28 + len(raw) + len(binary))
+        + struct.pack("<II", len(raw), 0x4E4F534A)
+        + raw
+        + struct.pack("<II", len(binary), 0x004E4942)
+        + binary
+    )
+    corrected = migrate.rotate_glb(glb)
+    length = struct.unpack_from("<I", corrected, 12)[0]
+    result = np.array(struct.unpack_from("<9f", corrected, 28 + length)).reshape(-1, 3)
+    assert np.array_equal(result, np.array(positions) * [-1, 1, -1])
+    assert np.array_equal(
+        result[:, [0, 2, 1]] * [1, -1, 1], [[-1, -3, -2], [-2, -3, -2], [-1, -3, -1]]
+    )
+    assert struct.unpack_from("<3I", corrected, 28 + length + 36) == (0, 1, 2)
+
+
+def test_voxel_quantized_grade_does_not_become_a_wall_but_ledge_remains():
+    costs = module("navigation/terrain_costs").supported_slope_costs
+    xy = np.array([[x, y] for x in np.arange(-1, 1.01, 0.06) for y in np.arange(-1, 1.01, 0.06)])
+    height = np.round(xy[:, 0] * np.tan(np.radians(20)) / 0.06) * 0.06
+    surface = np.column_stack((xy, height, np.full(len(xy), 0.15)))
+    assert costs(surface)[1] > len(surface) * 0.8
+    for discontinuity in (0.12, 0.15, -0.4):
+        ledge = surface.copy()
+        ledge[:, 2] = np.where(xy[:, 0] >= 0, discontinuity, 0)
+        corrected, _ = costs(ledge)
+        near_edge = np.abs(xy[:, 0]) < 0.12
+        assert np.all(corrected[near_edge, 3] > 0.1)
+    stone = surface.copy()
+    stone[:, 2] = 0
+    stone[(abs(xy[:, 0]) < 0.12) & (abs(xy[:, 1]) < 0.12), 2] = 0.15
+    assert costs(stone)[1] == 0
+
+
+def test_grade_fit_cannot_bridge_an_unobserved_gap():
+    costs = module("navigation/terrain_costs").supported_slope_costs
+    points = np.array(
+        [
+            [x, y, 0 if x < 0 else 0.15, 0.15]
+            for x in [-0.3, -0.2, 0.2, 0.3]
+            for y in np.arange(-0.3, 0.31, 0.1)
+        ]
+    )
+    assert costs(points)[1] == 0
+
+
+def test_existing_reserve_overlap_only_allows_departure_not_approach_or_body_overlap():
+    check = module("navigation/footprint").command_footprint_clear
+    pose = [0, 0, 0.37, 0, 0, 0, 1]
+    behind = [[-0.53, 0, 0.5, 0.5]]
+    assert check(0.15, 0, behind, pose)
+    assert not check(-0.1, 0, behind, pose)
+    assert not check(0, 0.35, behind, pose)
+    assert not check(0.15, 0, [[-0.49, 0, 0.5, 0.5]], pose)
+    assert not check(0.15, 0, [[0.53, 0, 0.5, 0.5]], pose)
+    # A longer admitted camera age must also enlarge the collision envelope.
+    assert not check(0.15, 0, [[0.80, 0, 0.5, 0.5]], pose)
+
+
+def test_terrain_fit_cache_invalidates_when_a_new_obstacle_is_observed():
+    costs = module("navigation/terrain_costs")
+    normalize = costs.TerrainCostNormalizer()
+    xy = np.array([[x, y] for x in np.arange(-1, 1.01, 0.1) for y in np.arange(-1, 1.01, 0.1)])
+    grade = np.column_stack((xy, xy[:, 0] * np.tan(np.radians(20)), np.full(len(xy), 0.15)))
+    clear, count = normalize(grade)
+    assert count > len(grade) * 0.9
+    assert np.array_equal(normalize(grade)[0], clear)
+    changed = np.vstack((grade, [0.02, 0.02, 0.3, 0.3]))
+    cached, count = normalize(changed)
+    fresh, expected_count = costs.supported_slope_costs(changed)
+    assert np.array_equal(cached, fresh) and count == expected_count
+    assert cached[len(grade) // 2, 3] > 0.1
+    assert np.array_equal(normalize(grade)[0], clear)
+
+
+def test_velocity_regulation_keeps_the_selected_arc_and_its_braking_clearance():
+    monitor = module("navigation/footprint")
+    pose = [0, 0, 0.37, 0, 0, 0, 1]
+    hazard = [[0.834, -0.08, 0.15, 0.15]]
+    speed, yaw, scale = monitor.regulate_command(0.15, -0.245, hazard, pose)
+    assert 0.2 <= scale < 1
+    assert speed / yaw == pytest.approx(0.15 / -0.245)
+    assert monitor.command_footprint_clear(speed, yaw, hazard, pose)
+    assert monitor.regulate_command(0.15, 0, [[0.56, 0, 0.2, 0.2]], pose) == (0, 0, 0)
+    assert monitor.regulate_command(0.15, 0, [[0.49, 0, 0.2, 0.2]], pose) == (0, 0, 0)
+    reverse, _, scale = monitor.regulate_command(-0.1, 0, [[-0.68, 0, 0.2, 0.2]], pose)
+    assert 0.2 <= scale < 1 and reverse < 0
+    assert monitor.command_footprint_clear(reverse, 0, [[-0.68, 0, 0.2, 0.2]], pose)
diff --git a/tests/test_ai_polygon_realtime.py b/tests/test_ai_polygon_realtime.py
new file mode 100644
index 0000000..0d7ee02
--- /dev/null
+++ b/tests/test_ai_polygon_realtime.py
@@ -0,0 +1,303 @@
+"""Lifecycle and clock-isolation checks; GPU/stream performance is qualified on Worker."""
+
+import importlib.util
+import threading
+import time
+from pathlib import Path
+from types import SimpleNamespace
+
+import numpy as np
+import pytest
+from pydantic import ValidationError
+from test_ai_polygon import make_world, sample
+from test_observatory_recorded_jobs import _definitions
+
+from k1link.observatory.recorded_jobs import ObservatoryRecordedJobQueue
+from k1link.simulation.ai_polygon.contracts import (
+    RealtimeSnapshot,
+    RunCreate,
+    StreamEndpoint,
+    WorkerHello,
+)
+from k1link.simulation.ai_polygon.runs import RunStore
+from k1link.simulation.ai_polygon.worlds import WorldStore
+
+
+@pytest.fixture
+def realtime(tmp_path):
+    queue = ObservatoryRecordedJobQueue(tmp_path, definitions=_definitions())
+    store = RunStore(WorldStore(tmp_path), queue)
+    hello = WorkerHello(
+        worker_id="worker-006",
+        instance_id="a" * 32,
+        runtime="isaac-sim-6.1",
+        model_ids=["reference"],
+        profile_sha256="b" * 64,
+        runtime_sources={key: "c" * 64 for key in ("worker", "scene", "models", "robot")},
+        execution_modes=["realtime"],
+        stream=StreamEndpoint(server="100.80.1.2"),
+    )
+    store.register(hello)
+    world = make_world(store.worlds)
+    row = store.start(
+        RunCreate(world_id=world["world_id"], clock="realtime", start_paused=True),
+        "test-realtime-01",
+    )
+    return store, hello, row
+
+
+def snapshot(**changes):
+    fields = dict(
+        sequence=0,
+        control_sequence=0,
+        state="ready",
+        phase="running",
+        simulation_time_ns=0,
+        wall_elapsed_seconds=1,
+        physics_steps=0,
+        render_frames=30,
+        sensor_frames=0,
+        inference_count=0,
+        dropped_frames=0,
+        rtf=0,
+        render_fps=30,
+        sensor_fps=0,
+        ai_hz=0,
+        pose_xy=(0, 0),
+        pose_yaw=0,
+        speed_mps=0,
+        applied_speed_mps=0,
+        applied_yaw_rate_rps=0,
+        stop_reason="paused",
+        ai_ready=False,
+        stream_ready=True,
+        camera="follow",
+    )
+    return RealtimeSnapshot(**{**fields, **changes})
+
+
+def test_worker_owns_acknowledgement_and_no_frame_archive_on_core(realtime):
+    store, hello, row = realtime
+    key = row["run_id"]
+    assert not (store.directory(key) / "frames").exists()
+    assert store.poll(hello.instance_id, key)["run"]["state"] == "starting"
+    store.snapshot(key, hello.instance_id, snapshot())
+    command = store.control(key, "play")
+    assert command["state"] == "ready"
+    assert store.control(key, "play")["control_sequence"] == command["control_sequence"]
+    assert store.poll(hello.instance_id, key)["run"]["state"] == "ready"
+    store.snapshot(
+        key,
+        hello.instance_id,
+        snapshot(sequence=1, control_sequence=1, state="running", simulation_time_ns=66666667),
+    )
+    assert store.get(key)["state"] == "running"
+    with pytest.raises(RuntimeError):
+        store.sample(key, hello.instance_id, sample())
+    with pytest.raises(RuntimeError):
+        store.control(key, "step")
+
+
+def test_disconnect_restart_reconcile_never_releases_gpu_or_replays(realtime):
+    store, hello, row = realtime
+    key = row["run_id"]
+    store.snapshot(key, hello.instance_id, snapshot())
+    store.seen -= 21
+    assert store.status()["active_run"]["state"] == "disconnected"
+    with pytest.raises(RuntimeError):
+        store.queue.reserve_simulation("airun-" + "d" * 32)
+    recovered = RunStore(store.worlds, store.queue)
+    with pytest.raises(RuntimeError):
+        recovered.register(hello.model_copy(update={"instance_id": "d" * 32}))
+    recovered.register(hello)
+    assert recovered.poll(hello.instance_id, key)["action"] == "pause"
+    recovered.snapshot(key, hello.instance_id, snapshot(sequence=3))
+    assert recovered.get(key)["state"] == "ready"
+    recovered.control(key, "stop")
+    recovered.snapshot(key, hello.instance_id, snapshot(sequence=4))
+    assert recovered.get(key)["state"] == "stopping"
+    recovered.finish(key, hello.instance_id, "stopped", "")
+    recovered.queue.reserve_simulation("airun-" + "d" * 32)
+
+
+def test_snapshot_is_idempotent_and_rejects_unissued_ack(realtime):
+    store, hello, row = realtime
+    key = row["run_id"]
+    first = snapshot(sequence=5, simulation_time_ns=1000)
+    assert store.snapshot(key, hello.instance_id, first) == store.snapshot(
+        key, hello.instance_id, first
+    )
+    with pytest.raises(ValueError):
+        store.snapshot(key, hello.instance_id, snapshot(sequence=6, control_sequence=2))
+    with pytest.raises(ValueError):
+        store.snapshot(key, hello.instance_id, snapshot(sequence=6, simulation_time_ns=999))
+    with pytest.raises(ValidationError):
+        StreamEndpoint(server="8.8.8.8")
+
+
+def test_slow_ai_is_bounded_and_pause_fences_inflight_result(tmp_path):
+    path = Path(__file__).parents[1] / "simulation/ai-polygon/realtime_ai.py"
+    spec = importlib.util.spec_from_file_location("polygon_realtime_ai", path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    entered, release = threading.Event(), threading.Event()
+    seen = []
+
+    class Model:
+        def ready(self):
+            pass
+
+        def close(self):
+            pass
+
+        def infer(self, rgb):
+            seen.append(int(rgb[0, 0, 0]))
+            entered.set()
+            assert release.wait(2)
+            return None, []
+
+    decision = {
+        "speed_mps": 0.3,
+        "yaw_rate_rps": 0,
+        "reason": "road",
+        "road_fraction": 0.7,
+        "obstacle_count": 0,
+    }
+    ai = module.LatestInference(
+        Model,
+        SimpleNamespace(
+            reset=lambda: None, decide=lambda *_: SimpleNamespace(model_dump=lambda: decision)
+        ),
+        tmp_path,
+    )
+    ai.start()
+    try:
+        ai.enable(True)
+        ai.submit(np.zeros((2, 2, 3), dtype=np.uint8), 0, time.monotonic(), 0)
+        assert entered.wait(2)
+        for frame_id in range(1, 11):
+            ai.submit(
+                np.full((2, 2, 3), frame_id, dtype=np.uint8),
+                frame_id,
+                time.monotonic(),
+                frame_id * 1000,
+            )
+        assert ai.dropped == 9
+        assert ai.command(time.monotonic())[0] == 0
+        ai.enable(False)
+        release.set()
+        ai.close()
+        assert seen == [0]
+        assert ai.result is None
+    finally:
+        release.set()
+        ai.close()
+
+
+def test_expired_source_stops_even_if_inference_just_completed(tmp_path):
+    path = Path(__file__).parents[1] / "simulation/ai-polygon/realtime_ai.py"
+    spec = importlib.util.spec_from_file_location("polygon_command_deadline", path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    ai = module.LatestInference(None, None, tmp_path)
+    ai.enabled = True
+    ai.result = {
+        "captured_at": 0.0,
+        "completed_at": 1.0,
+        "decision": {"speed_mps": 0.3, "yaw_rate_rps": 0},
+    }
+    assert ai.command(1.1)[:3] == (0.0, 0.0, "stale-camera")
+    ai.result["captured_at"] = 1.0
+    assert ai.command(1.1)[:3] == (0.3, 0, "none")
+    ai.result["decision"]["speed_mps"] = -0.1
+    assert ai.command(1.1)[:3] == (-0.1, 0, "none")
+    assert ai.command(1.6)[:2] == (0, 0)
+    # The worker's 0.8 s camera budget still rejects old images, even when
+    # a result has just arrived, and never extends the command watchdog.
+    ai.result.update(captured_at=1.0, completed_at=1.7)
+    assert ai.command(1.79, frame_deadline=0.8)[:3] == (-0.1, 0, "none")
+    assert ai.command(1.81, frame_deadline=0.8)[:3] == (0.0, 0.0, "stale-camera")
+    ai.result.update(captured_at=1.7, completed_at=1.0)
+    assert ai.command(1.79, frame_deadline=0.8)[:2] == (0.0, 0.0)
+
+
+def test_reverse_telemetry_keeps_the_run_speed_limit(realtime):
+    store, hello, row = realtime
+    store.snapshot(row["run_id"], hello.instance_id, snapshot(applied_speed_mps=-0.1))
+    assert store.get(row["run_id"])["telemetry"]["applied_speed_mps"] == -0.1
+    with pytest.raises(ValueError, match="скорость"):
+        store.snapshot(
+            row["run_id"], hello.instance_id, snapshot(sequence=1, applied_speed_mps=-0.4)
+        )
+
+
+def test_resume_requires_fresh_clear_observations(tmp_path):
+    from k1link.simulation.ai_polygon.policy import RoadPolicy
+
+    path = Path(__file__).parents[1] / "simulation/ai-polygon/realtime_ai.py"
+    spec = importlib.util.spec_from_file_location("polygon_resume", path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+
+    class Model:
+        def ready(self):
+            pass
+
+        def close(self):
+            pass
+
+        def infer(self, rgb):
+            return np.ones((512, 512), dtype=bool), []
+
+    ai = module.LatestInference(Model, RoadPolicy(0.3), tmp_path)
+    ai.start()
+
+    def frame(number):
+        ai.submit(np.zeros((2, 2, 3), dtype=np.uint8), number, time.monotonic(), number)
+        deadline = time.monotonic() + 2
+        while ai.count < number and time.monotonic() < deadline:
+            time.sleep(0.001)
+        assert ai.count == number
+        return ai.command(time.monotonic())[0]
+
+    try:
+        ai.enable(True)
+        assert frame(1) == frame(2) == 0
+        assert frame(3) > 0
+        ai.enable(False)
+        ai.enable(True)
+        assert frame(4) == frame(5) == 0
+        assert frame(6) > 0
+    finally:
+        ai.close()
+
+
+def test_windows_snapshot_sharing_retry_is_bounded(tmp_path, monkeypatch):
+    path = Path(__file__).parents[1] / "simulation/ai-polygon/local_state.py"
+    spec = importlib.util.spec_from_file_location("polygon_ipc", path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    original = module.os.replace
+    attempts = []
+
+    def transient(source, target):
+        attempts.append(1)
+        if len(attempts) <= 2:
+            raise PermissionError("Windows sharing violation")
+        original(source, target)
+
+    monkeypatch.setattr(module.os, "replace", transient)
+    monkeypatch.setattr(module.time, "sleep", lambda _: None)
+    target = tmp_path / "snapshot.json"
+    module.write_json(target, {"sequence": 3})
+    assert module.read_json(target) == {"sequence": 3}
+    assert len(attempts) == 3
+    attempts.clear()
+
+    def permanent():
+        attempts.append(1)
+        raise PermissionError("Not a transient lock")
+
+    with pytest.raises(PermissionError):
+        module.sharing_retry(permanent)
+    assert len(attempts) == 8
diff --git a/tests/test_ai_polygon_spawn.py b/tests/test_ai_polygon_spawn.py
new file mode 100644
index 0000000..79e93ef
--- /dev/null
+++ b/tests/test_ai_polygon_spawn.py
@@ -0,0 +1,42 @@
+"""Prevent a reconstructed trunk/face from being admitted between ground rays."""
+
+import importlib.util
+from pathlib import Path
+
+import numpy as np
+
+SPEC = importlib.util.spec_from_file_location(
+    "spawn_clearance", Path(__file__).parents[1] / "simulation/ai-polygon/spawn_clearance.py"
+)
+MODULE = importlib.util.module_from_spec(SPEC)
+SPEC.loader.exec_module(MODULE)
+
+
+def count(points, xy=(0, 0), heading=0, plane=(0, 0, 0)):
+    return MODULE.obstructing_triangles(np.array(points), np.array([[0, 1, 2]]), xy, heading, plane)
+
+
+def test_narrow_trunk_between_support_rays_and_crossing_triangle():
+    assert count([[0.21, 0.21, 0], [0.23, 0.21, 0], [0.21, 0.21, 2]]) == 1
+    # All vertices outside the prism, but the face crosses through its centre.
+    assert count([[-2, 0, 0.3], [2, 0, 0.3], [0, 2, 0.3]]) == 1
+
+
+def test_ground_step_and_overhead_geometry_do_not_block_start():
+    for z in (0, 0.05, 0.10, 1.5):
+        assert count([[-2, -2, z], [2, -2, z], [0, 2, z]]) == 0
+    assert count([[-2, -2, 0.12], [2, -2, 0.12], [0, 2, 0.12]]) == 1
+
+
+def test_triangle_bounding_box_alone_does_not_reject_empty_corner():
+    assert count([[0.4, 2, 0.5], [2, 0.4, 0.5], [2, 2, 0.5]]) == 0
+
+
+def test_clearance_tracks_translation_heading_and_support_slope():
+    points = np.array([[0.6, 0, 0.3], [0.7, 0, 0.3], [0.6, 0.03, 0.8]])
+    assert count(points) == 0
+    assert count(points, heading=45) == 1
+    points[:, 2] += points[:, 0] * 0.2 + 4
+    points[:, :2] += [10, -3]
+    assert count(points, (10, -3), 45, (0.2, 0, 4)) == 1
+    assert count([[8, -5, 3.6], [12, -5, 4.4], [10, -1, 4]], (10, -3), 45, (0.2, 0, 4)) == 0
diff --git a/tests/test_ai_polygon_worker_assets.py b/tests/test_ai_polygon_worker_assets.py
new file mode 100644
index 0000000..ed6dfb7
--- /dev/null
+++ b/tests/test_ai_polygon_worker_assets.py
@@ -0,0 +1,74 @@
+"""Paired assets remain on Worker; hashes and manifests cross the operator link."""
+
+import hashlib
+import importlib
+import json
+import struct
+from pathlib import Path
+
+import numpy as np
+import pytest
+from test_ai_polygon import ply, worker_asset
+
+
+def triangle_glb():
+    binary = np.array([[0, 0, 0], [1, 0, 0], [0, 0, 1]], dtype="<f4").tobytes()
+    binary += np.array([0, 1, 2], dtype="<u4").tobytes()
+    doc = {
+        "buffers": [{"byteLength": 48}],
+        "bufferViews": [
+            {"buffer": 0, "byteLength": 36},
+            {"buffer": 0, "byteOffset": 36, "byteLength": 12},
+        ],
+        "accessors": [
+            {"bufferView": 0, "componentType": 5126, "type": "VEC3", "count": 3},
+            {"bufferView": 1, "componentType": 5125, "type": "SCALAR", "count": 3},
+        ],
+        "meshes": [{"primitives": [{"attributes": {"POSITION": 0}, "indices": 1}]}],
+    }
+    encoded = json.dumps(doc).encode()
+    encoded += b" " * (-len(encoded) % 4)
+    return (
+        struct.pack("<III", 0x46546C67, 2, 28 + len(encoded) + len(binary))
+        + struct.pack("<II", len(encoded), 0x4E4F534A)
+        + encoded
+        + struct.pack("<II", len(binary), 0x004E4942)
+        + binary
+    )
+
+
+def test_paired_import_checks_actual_files_and_never_fetches_from_operator(tmp_path, monkeypatch):
+    monkeypatch.syspath_prepend(str(Path(__file__).parents[1] / "simulation/ai-polygon"))
+    importer = importlib.import_module("register_paired_scene")
+    preparer = importlib.import_module("prepare_terrain")
+    client_module = importlib.import_module("core_client")
+    source, collider = tmp_path / "source.ply", tmp_path / "mesh.glb"
+    source.write_bytes(ply())
+    collider.write_bytes(triangle_glb())
+    descriptor = worker_asset().model_dump(mode="json")
+    descriptor.update(
+        sha256=hashlib.sha256(source.read_bytes()).hexdigest(),
+        collider_sha256=hashlib.sha256(collider.read_bytes()).hexdigest(),
+    )
+    root = tmp_path / "worker"
+    request = importer.stage(root, source, collider, descriptor)
+    world = {**request.model_dump(mode="json"), "storage": {"kind": "worker"}}
+    manifest_path = preparer.prepare_terrain(root, tmp_path, world)
+    assert json.loads(manifest_path.read_text())["collider_sha256"] == descriptor["collider_sha256"]
+    assert importer.stage(root, source, collider, descriptor) == request
+    token = tmp_path / "token"
+    token.write_text("test" * 12)
+    client = client_module.CoreClient("http://127.0.0.1:18080", token, "a" * 32)
+    monkeypatch.setattr(
+        client_module.http.client,
+        "HTTPConnection",
+        lambda *args, **kwargs: pytest.fail("Operator must not supply this asset"),
+    )
+    cached = root / "state/worlds" / (request.sha256 + ".ply")
+    client.download(world, cached)
+    with pytest.raises(RuntimeError, match="missing or changed"):
+        client.download(world, tmp_path / "missing.ply")
+    with pytest.raises(ValueError, match="match"):
+        preparer.prepare_terrain(root, tmp_path, {**world, "collider_sha256": "f" * 64})
+    with pytest.raises(ValueError, match="identity"):
+        importer.stage(root, source, collider, {**descriptor, "collider_sha256": "f" * 64})
diff --git a/tests/test_observatory_recorded_jobs.py b/tests/test_observatory_recorded_jobs.py
index 8ca046a..06cfbca 100644
--- a/tests/test_observatory_recorded_jobs.py
+++ b/tests/test_observatory_recorded_jobs.py
@@ -1604,3 +1604,47 @@ def test_queue_detects_mutated_immutable_identity(tmp_path: Path) -> None:
 
     with pytest.raises(ObservatoryRecordedQueueIntegrityError, match="identity"):
         queue.get(job.job_id)
+
+
+def test_simulation_reservation_serializes_recorded_claims_and_survives_restart(tmp_path):
+    queue = _queue(tmp_path)
+    job, _ = queue.submit(_intent(), enqueue=True)
+    owner = "airun-" + "a" * 32
+    queue.reserve_simulation(owner)
+    queue.reserve_simulation(owner)
+    restarted = _queue(tmp_path)
+    assert restarted.claim_next(claimant_id="worker-006", claim_request_id="sim-blocked") is None
+    assert restarted.get(job.job_id).state == "queued"
+    with pytest.raises(ObservatoryRecordedQueueConflictError):
+        restarted.release_simulation("airun-" + "b" * 32)
+    restarted.release_simulation(owner)
+    claim = restarted.claim_next(claimant_id="worker-006", claim_request_id="sim-released")
+    assert claim.job.job_id == job.job_id
+    with pytest.raises(ObservatoryRecordedQueueBusyError):
+        queue.reserve_simulation(owner)
+
+
+def test_simulation_and_recorded_claim_cannot_win_together(tmp_path):
+    queue = _queue(tmp_path)
+    queue.submit(_intent(), enqueue=True)
+    second = _queue(tmp_path)
+    barrier = Barrier(2)
+
+    def simulate():
+        barrier.wait()
+        try:
+            queue.reserve_simulation("airun-" + "a" * 32)
+            return True
+        except ObservatoryRecordedQueueBusyError:
+            return False
+
+    def recorded():
+        barrier.wait()
+        return (
+            second.claim_next(claimant_id="worker-006", claim_request_id="racing-claim") is not None
+        )
+
+    with ThreadPoolExecutor(max_workers=2) as pool:
+        simulation = pool.submit(simulate)
+        inference = pool.submit(recorded)
+        assert simulation.result() != inference.result()