docs(perception): record external worker acceptance

README.md

@@ -103,6 +103,21 @@ uv run k1link doctor
 uv run pytest
 ```
 
+## External perception worker
+
+Mission Core can now package one sealed camera epoch as an immutable,
+content-addressed `missioncore.compute-job/v1` without changing the source
+evidence. The first recorded TEST007 job was digest-verified and decoded on a
+separate Windows RTX 4090 worker, executed through pinned Triton/YOLOX-S and
+returned a content-addressed `missioncore.compute-result/v1` containing 56
+session-timestamped generic COCO detections. The worker never connects to K1 and
+does not own the raw archive.
+
+This is a recorded execution acceptance, not a navigation stack: Rerun result
+projection, large-epoch performance, direct worker transport and bounded live
+fan-out remain open. See [ADR 0014](docs/adr/0014-bounded-external-perception-worker.md)
+and the [external worker contract](docs/10_EXTERNAL_PERCEPTION_WORKER.md).
+
 ## Mission Core Control Station and visualization adapters
 
 The browser application is the universal Mission Core Control Station rather than

docs/01_IMPLEMENTATION_PLAN.md

@@ -298,6 +298,37 @@ 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.
+
+The accepted result is perception evidence only. Rerun projection, 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;

docs/10_EXTERNAL_PERCEPTION_WORKER.md

@@ -0,0 +1,85 @@
+# External perception worker contract
+
+## Boundary
+
+```text
+K1
+  -> XGRIDS device plugin / Mission Core Edge
+  -> authoritative raw session on Mac
+  -> bounded missioncore.compute-job/v1
+  -> replaceable GPU worker / Triton
+  -> content-addressed missioncore.compute-result/v1
+  -> future optional Rerun derived layer
+```
+
+The GPU worker cannot discover, provision, start or stop a K1. It receives an
+observation package and returns derived observations. The Mac archive is never
+rewritten when a job is prepared or a result is received.
+
+## Recorded camera job v1
+
+`k1link compute prepare-camera-job` accepts a sealed cataloged session, one
+camera source and one physical codec epoch. It performs the canonical camera
+checks before and after copying into a private staging directory:
+
+- summary schema/source/epoch and per-segment durability contract;
+- exact ordered JSONL index and its summary SHA-256;
+- init SHA-256, every segment length/SHA-256 and complete stream SHA-256;
+- continuous ISO-BMFF decode timeline and finite media duration;
+- source session-time mapping through the captured host clock origin.
+
+Publication atomically renames this layout below ignored storage:
+
+```text
+<jobs>/<job-id>/
+├── job.json
+└── input/camera/<source-id>/epoch-N/
+    ├── summary.json
+    ├── index.jsonl
+    ├── init.mp4
+    └── segments/*.m4s
+```
+
+`input_sha256` is SHA-256 over canonical JSON for the complete path-free input
+descriptor. `job_id` includes its first 96 bits and validators require the full
+digest, so a prefix collision fails closed.
+
+## Recorded result v1
+
+The accepted worker profile verifies every transferred file, reconstructs the
+stream outside the job, decodes frames and preserves strictly increasing
+best-effort timestamps in session seconds. Its result identity is canonical JSON
+covering:
+
+- job and full input generation;
+- pipeline ID/version;
+- model ID/version/weight SHA-256;
+- score/NMS thresholds, tensor shape, color order and letterbox policy.
+
+The identity's full SHA-256 names an immutable `result-<sha256>` directory.
+The result manifest binds the detection artifact's length and SHA-256. A repeat
+with the same identity validates the existing directory and returns it without
+calling Triton.
+
+## Accepted model profile
+
+- YOLOX-S official ONNX from upstream release `0.1.1rc0`;
+- Apache-2.0;
+- SHA-256
+  `c5c2d13e59ae883e6af3b45daea64af4833a4951c92d116ec270d9ddbe998063`;
+- FP32 input `[1,3,640,640]`, BGR, bilinear top-left letterbox, pad 114;
+- FP32 output `[1,8400,85]`, official grid/stride decode, COCO-80;
+- Triton 2.70.0 / ONNX Runtime GPU backend in the pinned 26.06 image.
+
+This generic model is useful for proving the contract and timing path. It is not
+accepted for obstacle avoidance, free-space estimation or safety decisions.
+
+## Open gates
+
+1. Import the accepted result as an optional session-timestamped Rerun layer.
+2. Qualify TEST007's 206-second epoch and record throughput/resource telemetry.
+3. Replace SSH/SCP with a reviewed authenticated worker transport.
+4. Add bounded live queues, sampling/drop policy and acquisition-isolation
+   tests.
+5. Introduce tracking, segmentation/free-space, calibration and point-cloud
+   models as separate versioned pipelines.

docs/adr/0014-bounded-external-perception-worker.md

@@ -0,0 +1,73 @@
+# ADR 0014: Bounded external perception worker
+
+## Status
+
+Accepted for recorded laboratory execution on 2026-07-19. Live transport and
+behavior integration are not accepted.
+
+## Context
+
+Mission Core has a canonical K1 control/archive path and a second workstation
+with an RTX 4090. Sending the entire K1 archive to an ad hoc model environment,
+letting the GPU host connect to K1 or installing another monolithic Mission Core
+would erase the product boundary already established around the device plugin.
+
+The first useful hypothesis is narrower: can one sealed observation be sent to
+a replaceable executor, processed by a standard model server and returned as a
+versioned derived observation with enough identity to reproduce and reject
+stale results?
+
+## Decision
+
+1. Mission Core Edge on the Mac remains the sole K1 command owner and
+   authoritative raw archive owner.
+2. The edge publishes a bounded `missioncore.compute-job/v1`. The first profile
+   contains one canonical camera codec epoch, its archive index/summary and a
+   path-free manifest; it never contains the full MQTT capture.
+3. The input generation SHA-256 covers session/source identity, media contract,
+   timeline, ordered file identities and byte counts. The job directory name is
+   derived from that generation and publication is atomic.
+4. Camera packaging uses the saved-session `RecordedMediaInspector` timing and
+   digest contract. It additionally verifies summary-level index and complete
+   stream digests before handoff.
+5. The Windows worker is an executor, not a second Mission Core. It revalidates
+   every staged file, decodes a derived copy and cannot talk to K1.
+6. NVIDIA Triton is the first replaceable inference runtime. YOLOX-S is the
+   first generic detector because its official repository/deployment path is
+   Apache-2.0 and an official ONNX artifact can be pinned by SHA-256. Model
+   weights stay outside Git.
+7. A result identity covers the full input generation, pipeline, model digest
+   and parameters. `missioncore.compute-result/v1` plus its detection artifact
+   is atomically published under `result-<identity-sha256>`. Exact repeats
+   validate and return the existing generation without inference.
+8. Detections are observations only. They cannot emit scanner, motion or mission
+   commands.
+9. SSH/SCP is permitted for the recorded lab bootstrap only. Direct Triton or a
+   later routed transport must receive its own address, authentication, bounded
+   queue and failure-policy review before live use.
+
+## Accepted evidence
+
+TEST007 left epoch 1 produced job
+`recorded-camera-fae25b92ecf645c2c9765dd4`. All 59 transferred files and
+4,567,569 package bytes were verified. The worker reconstructed the canonical
+4,549,458-byte H.264 stream with SHA-256
+`bceca577f762fdb79c4e8901e5c2a2330fd07d24dc091b1b703c5af4010d8a56`
+and decoded 56 frames over 5.497 seconds.
+
+Pinned YOLOX-S produced the accepted result
+`result-5484a72e81192b19f4e1da2a2dcfd10c876c277ff3e96c6d60cbc9d917b9f604`.
+It contains 56 generic `person` detections over 54 frames. Mean client-observed
+Triton request latency was 15.611 ms and p95 was 23.728 ms. The result is not a
+quality acceptance for navigation.
+
+## Consequences
+
+- Model/runtime changes can be compared against the same immutable input.
+- The worker can later move from the lab LAN to a routed host without changing
+  K1 ownership.
+- Storage temporarily duplicates the bounded encoded stream and preprocessed
+  frames on the worker; retention remains unimplemented.
+- Host-arrival camera timing does not become sensor calibration.
+- The next accepted vertical must project this result into the saved Rerun
+  timeline before live fan-out is introduced.

docs/lab/005_FIRST_RECORDED_PERCEPTION_20260719.redacted.md

@@ -0,0 +1,50 @@
+# Lab 005: First recorded external perception acceptance
+
+Date: 2026-07-19 (Europe/Moscow).
+
+## Scope
+
+One sealed TEST007 left-camera epoch was packaged by Mission Core Edge, copied
+to the Windows RTX 4090 worker, decoded and submitted to pinned YOLOX-S through
+Triton. No K1 command was sent and no raw source file was modified.
+
+## Evidence identities
+
+- session: `20260718T201659Z_viewer_live`;
+- job: `recorded-camera-fae25b92ecf645c2c9765dd4`;
+- input SHA-256:
+  `fae25b92ecf645c2c9765dd4b4963f23cf8c101f22fd5a2e14a104f0b447140d`;
+- 59 transferred files, 4,567,569 bytes;
+- canonical stream: 4,549,458 bytes, SHA-256
+  `bceca577f762fdb79c4e8901e5c2a2330fd07d24dc091b1b703c5af4010d8a56`;
+- decode: H.264 High, 800x600, 56 frames, 5.497 seconds;
+- model SHA-256:
+  `c5c2d13e59ae883e6af3b45daea64af4833a4951c92d116ec270d9ddbe998063`;
+- result:
+  `result-5484a72e81192b19f4e1da2a2dcfd10c876c277ff3e96c6d60cbc9d917b9f604`;
+- detections artifact: 23,802 bytes, SHA-256
+  `71b291616e87bbdd2b7a295b1e00dac2ff3e60c627e63b72a27ba802c32a3ae0`.
+
+## Observed result
+
+All 56 frames were processed. The generic detector returned 56 `person`
+detections across 54 frames, with confidence 0.250729–0.530741. Client-observed
+Triton request latency was mean 15.611 ms, p50 15.001 ms, p95 23.728 ms and max
+31.099 ms.
+
+Fragment boundaries contain repeated DTS values, but all 56 frames decode and
+the persisted best-effort timestamps are strictly increasing. Synchronization
+remains host-arrival best effort.
+
+The first runner attempt completed all inference requests but failed a
+post-inference EOF guard before publication. No partial result directory was
+published. After correction, the content-addressed result was published. An
+identical repeat left Triton's inference count unchanged at 168, proving reuse
+of the existing result.
+
+## Interpretation
+
+This accepts the external recorded compute boundary, not detector quality or a
+vehicle decision loop. The result has not yet been projected into Rerun. SSH/SCP
+is retained only as laboratory bootstrap evidence and is not the product data
+plane.