# Implementation plan This plan supersedes the app-dependent experiment order in the reference Bible. Each gate produces evidence and an explicit GO, PAUSE or BLOCKED result. ## Current checkpoint — 2026-07-20 | Stage | Result | | --- | --- | | Stage 0 host/repository | GO — isolated Python 3.12 environment and private Git | | Gate 1 physical operation | GO — autonomous double-click start/stop verified | | Stage 1 BLE discovery | GO — repeatable advertisement and GATT profile | | Stage 2 BLE provisioning | GO — reviewed 99-byte profile, LAN association confirmed | | Stage 2 local connection matrix | GO for Bridge/direct-LAN; LAB-ONLY for Quick Connect — firmware review proved the K1 3.0.2 AP credential is firmware-constant and the SSID follows the `XGR-` identity/MAC rule. One prepared Mac completed AP-enable, AP-ready, CoreWLAN association and the normal control lifecycle. The successful host had been seeded from the reviewed firmware archive; a copied build on a clean host cannot acquire that material automatically. Automatic firmware download, iPhone extraction and hard-coding were rejected. Bridge remains the product path; Direct Connect remains physically pending | | Stage 3 application session | GO — MQTT 3.1.1 on confirmed K1 TCP 1883 | | Stage 4 artifacts/flows | GO — bounded capture, hashes and negative control | | Stage 5 point cloud | GO — raw-LZ4 protobuf, 1,140 live frames decoded | | Stage 5 pose | GO — 1,215 live frames decoded and motion-correlated | | Stage 5 modeling telemetry | GO (read-only) — bounded `ModelingReport` decoder and device-reported scan time/distance/speed are integrated before the preview queue | | Stage 5 camera | GO (live) — left/right RTSP/H.264 preview observed, read-only runtime adapter and physical UI acceptance completed | | Stage 6 live viewer | GO — React Control Station, embedded self-hosted Rerun cloud/trajectory, plugin-owned spatial controls and live device/Mac metrics | | Stage 7 observation archive | GO (point/pose/telemetry contract) — durable catalog, recovery, capture-clock-bounded RRD preparation, archived metric time series, saved-session timeline and atomic playback are implemented | | Stage 7 recorded cameras | GO — the accepted physical session sealed real fMP4 camera data, produced two digest-bound recorded sources and passed manifest/init/segment range admission alongside a 107 MB RRD; browser QA opened real frames from the 170 MB camera epoch at multiple shared-timeline positions | | Plugin isolation | GO (laboratory control plane) — vendor backend/frontend and optional scene controls are plugin-owned; manifest/runtime descriptor parity, versioned handshake, lifecycle health and transport correlation fail closed while execution remains in-process | | K1 application control | GO (physical staged cycle) — after fixing the PCAP-proven `sint64` time field, one explicit UI launch completed all 14 canonical operations on one control session, reached live `SCANNING + project + init_ready`, displayed real points, then one explicit STOP returned K1 to unbound `READY`. No retry or fallback command was sent. Native-project reuse through LixelGO/USB remains an independent verification | | Stage 8 product storage | PAUSE — retention, replication, encryption, capacity monitoring and long-run browser/WASM stress remain deployment gates | USB project copying remains optional ground truth rather than a blocker for the now-verified network path. Owner-operated LixelGO traffic verifies the MQTT start/stop mapping and RTSP camera transport. The exact start/stop protobuf encoder, response correlator and device-status state machine now run in the separately gated interactive publisher; the legacy shadow publisher remains disabled. Retained PCAP plus client static analysis prove that OpenAPI is one private application-level value in this LixelGO build, not a per-scanner credential. The selected BLE peripheral returns its own LAN IP; the initial unbound `DeviceInfoRequest` then returns vendor ID, serial, model, activation and version facts for that live K1. Mission Core's shadow bootstrap reproduces the 10 observed pre-START requests byte-for-byte across four response windows and five observed publish groups: initial DeviceInfo binding; then the ordered ModelingStatus/RTK and time-sync/DeviceInfo/RTK groups without an invented response barrier between them; NTRIP read; then cloud/RTK/DeviceInfo reads. The first ModelingStatus request has no mandatory synchronous response barrier; readiness is separately attested through live DeviceStatus. A fixed macOS Keychain loader now reads the exact 36-byte authority without environment/file/browser fallback. The response orchestrator correlates required headers/results and rejects identity/profile drift. The legacy shadow publisher remains a separate write-disabled type that cannot call its injected sink. Its dormant coordinator is wired into the XGRIDS facade only for redacted planning/state. Separately, plugin v0.5.0 installs the reviewed physical-acceptance transport behind explicit operator actions and one continuous background session owner. Each operation key is consumed before publish, automatic reconnect and application retry are forbidden, and any unknown outcome poisons the transport. Public state contains only bounded lifecycle, counter and failure metadata; authority and device identity remain private. The one-time `k1link authority provision` command delegates secret entry to the macOS Keychain TTY prompt and validates through the production loader without receiving the value in argv/environment/file/browser state. During the first physical attempt that CLI loader did not receive macOS approval, so a temporary process-local Security.framework lab adapter supplied the already reviewed value. The transport emitted bootstrap ordinals 1–6 and received all five required responses through the old third window, but topic-only routing misassociated the unbound and bound RTK responses and the orchestrator raised `ApplicationBootstrapError`. No START was emitted; live DeviceStatus stayed READY and the local point/pose counts remained zero. The local capture was sealed and the temporary Keychain item was deleted. The second physical attempt then correlated all ten bootstrap requests and one START success response. It nevertheless did not reproduce the full LixelGO lifecycle: all eleven publishes completed in 365 ms, the process did not retain the control session through initialization, and it omitted the immediate and post-initialization status reads present in both retained LixelGO START captures. K1 reported `SCAN_STARTING` for about 24 seconds, emitted system error `0x32040133` twice (the recovered `ALGORITHM_ERROR` namespace/value), returned to `OTHER_STATUS`, produced zero point/pose frames and showed steady red. The operator restored steady green through a normal power cycle. The legacy `run_bootstrap()` path now rejects before emission. The offline acceptance path maps requests 1–6 to connection, request 7 to explicit scan-workspace entry, requests 8–10 to explicit project-prompt opening and request 11 to the operator's START confirmation. Captured human delays are not replayed or treated as minimum timeouts. After the immediate operation-12 read, operations 13–14 wait for the same bound K1 to report `SCANNING`, a bound project and `init_ready=true`. The same transport is then serviced until an explicit STOP request; post-STOP ownership waits for live unbound `READY`. System-error and status reports are decoded into redacted safety state. The next staged physical attempt, after correcting the protobuf `sint64` time field, completed all 14 canonical operations, reached live scanning with real point/pose data and then completed one explicit STOP. K1 returned to unbound `READY`; no retry or fallback device command was sent. The durable local session is catalogued as ready/replayable, its RRD and two camera sources pass digest-bound range admission, and physical double-click remains the fallback. Reuse of the vendor-native project is still verified separately through LixelGO or USB. The Stage 6 live path uses a bounded raw-first bridge: loss in the visualization queue cannot discard MQTT evidence. The queue holds four preview messages; `ModelingReport` is consumed by a plugin-injected observer before that queue and does not compete with point/pose frames. A project name is mandatory at preparation, NFKC-normalized/trimmed on both sides, rejected for control or surrogate characters or more than 96 characters, and retained only as display metadata. Stage 7 adds an independent durable observation-session lifecycle. New native captures durably publish the clock origin before camera production, retain a transport-scoped provisional envelope and atomically switch the capture-summary pointer to a content-addressed session envelope after all producers stop. Completed or recovered captures are prepared once by a bounded backend worker into digest-bound RRD/cache-v9 and camera-manifest generations; a browser replay request never performs conversion. The RRD materializes real origin/end rows plus archived device-reported distance, speed and scan-time series. The saved scene, controller and timeline remain hidden until the complete RRD range and every declared camera pass admission. Sensor-to-display latency and camera/LiDAR sensor-clock alignment remain separate correlation tests. The generic host clears a prior manual or archived spatial source only after a plugin-owned live/file-replay start succeeds; a failed start preserves the current scene. Once an acquisition is nonterminal, a fail-closed guard blocks saved-session switching, persisted replay reattach and manual source input/apply/reset until that acquisition has been finalized. The successful automatic-source action is an internal start-result boundary and intentionally does not masquerade as an operator source switch. The K1 scene-level stop control is plugin-owned. A plugin-commanded acquisition sends one canonical STOP and seals locally after protocol-reported standby; an operator-manual acquisition still stops and seals only local reception and reports scanner state as unknown. ## Stage 0 — repository and host baseline Deliverables: - local Python 3.12 `.venv` and lock-file; - `k1link doctor` with no side effects; - immutable reference documents and checksums; - safety policy, artifact policy, session manifest schema; - no Homebrew changes and no K1 interaction. Acceptance: - `uv sync --group dev` succeeds; - `uv run k1link doctor` identifies a local `.venv`; - tests, Ruff and mypy pass; - repository contains no secrets or real captures. ## Gate 1 — physical and USB baseline Code needed: none initially. Actions: 1. Record masked device identity, LED boot sequence, battery and visible state. 2. Determine whether the unit is activated; solid-yellow or an activation prompt is treated as a distinct blocker. 3. Run one short autonomous button scan: start, keep still for at least 20 seconds, move through a simple scene, stop. 4. Enter USB mode using the documented physical shortcut if available. 5. Copy a project tree and selected metadata/logs into an ignored session path. 6. Hash every copied file and create a redacted manifest. GO: the unit boots normally and records a usable autonomous project. PAUSE: USB access is unavailable but the scan succeeds; continue BLE while documenting the missing ground truth. STOP: serious fault, overheating, activation lock, or destructive/ambiguous device state. ## Stage 1 — BLE non-mutating discovery and metadata toolkit Minimal dependencies: `bleak` plus the existing CLI stack. Add them only in the repository-local environment. Commands: ```text k1link ble scan k1link ble gatt-dump k1link ble listen ``` Requirements: - preserve advertisement snapshots, manufacturer/service data and RSSI; - describe macOS scanning as active discovery: CoreBluetooth does not support passive scan mode, although discovery does not change K1 configuration; - identify by observed name + macOS UUID + advertisement fingerprint; - inspect standby, scanning and USB states separately; - start with standard Device Information/Battery services; - custom characteristic reads are opt-in because proprietary reads may have side effects; - notifications are subscribed one characteristic at a time and logged to timestamped JSONL; - no arbitrary characteristic or provisioning writes in this stage. Enabling a notification normally updates the standard CCCD descriptor; this temporary, reversible protocol write is disclosed and handled separately from device configuration. GO: reproducible K1 advertisement and useful GATT map. PAUSE: permission or pairing/authentication is required. ## Gate 2 — existing network behavior Do not change Deco settings first. Actions: 1. Save the Mac interface and route table. 2. Observe router client list and Mac ARP/neighbor state with K1 off. 3. Repeat after K1 boot, during standby, scanning and after one power cycle. 4. Check for new Wi-Fi SSIDs exposed by K1 without attempting to join unknown networks automatically. 5. Confirm any candidate by disappearance/reappearance with K1 power state. GO-A: K1 joins a remembered LAN. Provisioning leaves the immediate critical path and work moves to the application-session gate. GO-B: K1 exposes its own AP. Document it as a separate topology and determine whether Mac can join without modifying router configuration. BLOCKED: neither association nor AP exists; provisioning becomes Stage 2. ## Stage 2 — no-app provisioning research This stage answers whether SSID/PSK can be sent safely from Mac. A generic BLE transport does not define the vendor payload. Evidence sources allowed before a write: - GATT UUIDs, properties and descriptors; - standard service identification; - read/notify state transitions correlated with network state; - owned K1 project logs and metadata; - public official firmware/documentation static evidence if separately reviewed; - deterministic structure inferred without sending device mutations. Before enabling any write profile, document: - exact service/characteristic and required security mode; - packet framing, encoding, ordering and length rules; - SSID/PSK redaction and secure input path; - checksum, nonce, token, certificate or signature fields; - commit/connect command and status response; - timeout, failure state, power-cycle behavior and rollback; - one explicit confirmation flag in addition to a named reviewed profile. GO: a deterministic provisioning profile exists with an explained expected state transition and safe rollback. HARD BLOCK: K1 knows no network, exposes only opaque vendor GATT, and no evidence source reveals the protocol. Random writes are not an alternative. ## Stage 3 — application session and targeted capture Prerequisite: confirmed K1 IP or direct AP topology. Implementation order: 1. Passive ARP/mDNS/SSDP and K1-filtered `tcpdump` baseline. 2. Small targeted TCP connect check against the confirmed IP. 3. Service detection only where the prior step supplies evidence. 4. Parse K1 logs for daemon names, ports and protocol strings. 5. Add protocol-specific connection attempts only for demonstrated endpoints. 6. Model `WIFI_ASSOCIATED_BUT_NO_DATA_SESSION` explicitly. `k1link net capture` must show the exact interface/filter, request privilege in the foreground, handle Ctrl+C, and never configure persistent `sudo` or ChmodBPF. GO: Mac becomes a K1 endpoint or receives a reproducible scan-correlated flow. BLOCKED: application session requires an unavailable mutual-auth token, certificate or undocumented handshake. ## Stage 4 — session artifacts and flow analysis Implement only after real traffic exists: - versioned session manifest; - UTC and monotonic timeline events; - SHA-256 inventory; - TCP reassembly and UDP flow grouping; - packet/byte rates, length histograms, entropy and first-byte fingerprints; - idle versus scanning correlation; - negative-control flows to reduce false positives. GO: repeatable flows can be separated and ranked without losing raw evidence. ## Stage 5 — decoders by evidence priority Priority: 1. device/scan status; 2. point-cloud preview; 3. pose/trajectory; 4. camera/panorama/video. Add dependencies per decoder extra, not to the base environment. Point-cloud and video signature probes operate on reassembled flows, not isolated packet payloads. A candidate becomes a decoder only after a controlled physical test confirms it. MVP GO: K1-to-Mac scan-correlated data can be captured reliably and at least one useful stream is decoded or structurally identified. ## Stage 8 — replaceable external perception worker Stages 6–7 established the raw-first live bridge and durable combined session archive described above. Stage 8 introduces a replaceable GPU executor without giving it K1 command authority or the authoritative archive. Accepted recorded gate, 2026-07-19: - the Mac validates one sealed camera epoch through the same digest and ISO-BMFF timing inspector used by saved-session replay; - `k1link compute prepare-camera-job` publishes an immutable `missioncore.compute-job/v1` below ignored runtime storage; - the first TEST007 job contains only one bounded 56-frame camera epoch, not the MQTT archive or another K1 control path; - Windows revalidates all 59 job files, reconstructs the exact canonical stream digest and decodes all 56 frames; - pinned Apache-2.0 YOLOX-S executes through pinned Triton 2.70.0 on the RTX 4090 and atomically publishes one content-addressed `missioncore.compute-result/v1`; - repeating the exact result identity returns the existing result without increasing Triton inference count; - Mission Core revalidates the returned generation, projects the exact camera frames and boxes onto `session_time`, and exposes the optional native Rerun view only after its complete RRF2 layer is accepted. The accepted result is perception evidence only. The larger TEST007 epoch, direct/routed worker transport and bounded live fan-out remain separate gates. SSH/SCP is allowed only as the recorded laboratory bootstrap. The worker never connects to K1. See [ADR 0014](adr/0014-bounded-external-perception-worker.md), the [worker contract](10_EXTERNAL_PERCEPTION_WORKER.md) and [Lab 005](lab/005_FIRST_RECORDED_PERCEPTION_20260719.redacted.md). ## Deferred work - ROS2/MCAP bridge; - onboard computer and drone integration; - automated scan-button electronics; - OpenWrt/monitor-mode infrastructure; - firmware or internal-Linux analysis; - browser-visible end-to-end shared-timeline playback of the accepted archive containing point cloud plus one selected K1 camera (backend preparation and digest-bound byte admission are already accepted); - OS-independent plugin/edge authority provider, PCAP-to-executor transcript regression and runtime ownership integration; - long-running large-session WebViewer/WASM memory telemetry; - production retention, replication, encryption and cross-platform packaging.