docs(perception): accept integrated TGS graph load gate

2026-08-27 12:01:27 +03:00
parent f153d671d3
commit b70dc4298a
3 changed files with 153 additions and 8 deletions
@@ -572,6 +572,16 @@ Exit:

 ### M4.9 — recorded-realtime release candidate and cutover

+The first complete integrated source-paced load gate passed on `2026-08-27`:
+the frozen RF-DETR world-state graph and unchanged CPU TGS candidate accounted
+for all `4,489` frames at `11.860865 FPS`, with joined output p95/p99 of
+`46.825886/63.94485 ms` and zero capacity drops. The immutable identities and
+resource envelope are recorded in
+[`experiments/perception/M49_TGS_INTEGRATED_GRAPH_ACCEPTANCE_2026-08-27.md`](../experiments/perception/M49_TGS_INTEGRATED_GRAPH_ACCEPTANCE_2026-08-27.md).
+This closes only the one-loop integrated load gate. Visual traversability,
+uncapped parity, repeated-loop soak, restart/recovery, navigation, actuation
+and production remain open.
+
 Execute on Worker 006 with the frozen RAVNOVES00 source:

 1. one full `1.0×` 448.723-second replay;
@@ -344,6 +344,17 @@ This accepts the CPU source-paced shadow and retains the candidate. It does not
 accept visual traversability quality, integrated graph performance, navigation
 or actuation.

+The product-integration load gate then passed on `2026-08-27`. The unchanged
+TGS candidate ran concurrently with the frozen RF-DETR world-state graph for
+all `4,489` frames at a requested `12 Hz`: every graph frame was delivered,
+every TGS frame was accounted and capacity drops remained zero. The joined
+output sustained `11.860865 FPS`, measured `46.825886/63.94485 ms` at p95/p99,
+and the CPU-only TGS stage measured `1.735/2.14019 ms` at p95/p99 with a
+`22.6 MiB` memory maximum. See
+[`experiments/perception/M49_TGS_INTEGRATED_GRAPH_ACCEPTANCE_2026-08-27.md`](../experiments/perception/M49_TGS_INTEGRATED_GRAPH_ACCEPTANCE_2026-08-27.md).
+Integrated runtime capacity is accepted; visual traversability, physical free
+space, navigation, actuation and production remain unaccepted.
+
 ### T4 — Candidate C occupancy/ESDF probe

 - Run nvblox separately with RF-DETR disabled for the isolated probe.
@@ -361,15 +372,16 @@ or actuation.

 ## Immediate next action

-Review the published full gravity-aligned TGS timeline in camera, metric 3D and
-costmap space. Preserve
+Keep the accepted integrated graph frozen and review the published full
+gravity-aligned TGS timeline in camera, metric 3D and costmap space. Preserve
 `GROUND_SUPPORT`, `NONGROUND_OCCUPIED`, `UNKNOWN_REJECTED` and `UNOBSERVED` as
 separate products; do not use AOS or infer free cells from absent
-republication. Attach the unchanged candidate to the realtime world-state graph
-in non-authoritative shadow mode and measure complete-graph FPS, latency,
-resource use and queue drops. If occupied evidence carpets the route, soft
-traversable vegetation or usable gaps, reject TGS without tuning it against the
-review evidence. Candidate A T2 replay remains blocked by the failed unmodified
-T1 gate. No LOW-STEP tuning, new object model, camera resize, fisheye
+republication. Integrated FPS, latency, resources and queue behavior are now
+accepted for the one-loop source-paced gate. If occupied evidence carpets the
+route, soft traversable vegetation or usable gaps, reject TGS without tuning it
+against the review evidence. Close and induce-test the Worker inference-runtime
+restart guard, then proceed to repeated-loop soak only if visual review also passes.
+Candidate A T2 replay remains blocked by the failed unmodified T1 gate. No
+LOW-STEP tuning, new object model, camera resize, fisheye
 rectification, manual dataset or parallel heavy Worker job is authorized by
 this decision.
@@ -0,0 +1,123 @@
+# M4.9T5 integrated TGS + reference graph shadow — 2026-08-27
+
+Status: **complete source-paced integrated shadow accepted**. Integrated
+runtime capacity is accepted for the frozen RAVNOVES00 graph. Visual
+traversability quality, physical free space, navigation, actuation and
+production remain unaccepted.
+
+## Decision
+
+The unchanged gravity-aligned TGS candidate and the frozen RF-DETR
+world-state graph completed the same `4,489`-frame RAVNOVES00 timeline from a
+shared `12 Hz` start barrier on Worker 006. Every graph frame was delivered,
+every TGS frame was accounted, all queues remained bounded and no capacity
+drop or supersession occurred.
+
+The integrated graph sustained `11.860865 FPS`, compared with the frozen
+reference result at `11.79902 FPS`. TGS added `1.735 ms` at p95 and used at
+most `22.6 MiB` of container memory; it did not add a GPU model. The exact
+per-frame join of graph and TGS outputs measured `46.825886 ms` at p95 and
+`63.94485 ms` at p99. This accepts the runtime integration gate without
+changing any navigation or safety authority.
+
+## Frozen identity
+
+| Item | Identity |
+| --- | --- |
+| Integrated result | `m49-tgs-integrated-graph-shadow-75e48fd8acf0246be16613e087fcf26fe909f7834718d217c74785a57f494b24` |
+| Worker run | `Worker 006 / ravnoves00-integrated-12fps-003` |
+| Mission Core revision used by Worker | `f153d671d316ea7b2200c0edcb80206d896ffcd5` |
+| Deterministic Worker artifact | `1818ea57887db819e9d986713b050bd15eae1a43ee4af70c5b8b5d39f4cd8522` |
+| Worker wheel | `b9414a1f2f1998543aeb63824d3cdfe55bbc6e4eecc3146d65962ebc7da86b8b` |
+| TGS native binary | `2c3b4c7e6b66d19d2e32ee3ef867fa8c14073be29441bcd19fe565d9632c6538` |
+| Integrated profile | `b61e018b2d04eec58802e2d4186ce7a3dd3a15b254db106b57b609e903eeef80` |
+| Frozen graph config | `1db6f6fa0561d819505a5e4f62fe256dab6fe4d9da6b4f1a9a5073bfcf772c90` |
+| Frozen TGS profile | `c2e07010aaee78259d36c057962d6bfb885349251ff7356d867e5813e632881c` |
+| Linked accepted TGS result | `m49-tgs-full-shadow-ef98de7db7596d48e8c8c0549ce68e6704ee03e87c8c4bcf1e3e748b7ccb032e` |
+
+The release was built from `git archive` of the declared revision. Uncommitted
+GAUSS and PlayCanvas files were neither packaged nor changed by the run.
+
+## Full-run result
+
+| Measurement | Result |
+| --- | ---: |
+| Graph admitted / delivered | `4,489 / 4,489` |
+| TGS timeline accounted | `4,489 / 4,489` |
+| LiDAR available | `3,928` |
+| Delivery ratio | `1.0` |
+| Capacity drops / supersessions / failures | `0 / 0 / 0` |
+| Effective world-state rate | `11.860865 FPS` |
+| Frozen reference rate | `11.79902 FPS` |
+| World-state completion p50 / p95 / p99 / max | `28.437492 / 46.696395 / 63.94485 / 170.202343 ms` |
+| Joined graph+TGS output p50 / p95 / p99 / max | `28.57367 / 46.825886 / 63.94485 / 170.202343 ms` |
+| TGS stage p50 / p95 / p99 / max | `1.186 / 1.735 / 2.14019 / 6.151 ms` |
+| TGS completion p50 / p95 / p99 / max | `17.605461 / 31.771832 / 42.957350 / 70.851977 ms` |
+| Queue high-water marks | `2` for every bounded stage |
+| Worker wrapper wall time | `544.887447 s` |
+
+All nineteen formal identity, accounting, throughput, latency, telemetry,
+capacity and false-authority checks passed.
+
+## Resource envelope
+
+| Resource | Result |
+| --- | ---: |
+| GPU utilization mean / p95 / max | `43.114973 / 53 / 57 %` |
+| GPU memory mean / p95 / max | `9,737.144 / 9,744 / 9,751 MiB` |
+| GPU power p95 / max | `149.342 / 153.81 W` |
+| TGS CPU mean / p95 / max | `3.474793 / 5.27 / 6.04 %` |
+| TGS memory mean / p95 / max | `20.571818 / 21.84 / 22.6 MiB` |
+| Graph container memory max | `1,769.472 MiB` |
+| Triton container memory max | `582.9 MiB` |
+
+The measured GPU envelope belongs to the already accepted reference graph.
+TGS is a CPU-only stage and does not request CUDA, Vulkan or a second model.
+
+## LAB publication boundary
+
+No third visual viewer is created for this gate. The TGS cells, source points,
+camera frames and semantic archive are byte-identical to the already published
+`m49-tgs-full-shadow` LAB. This run proves concurrent runtime behavior, so its
+new evidence is the immutable timing, queue and resource report above. Visual
+quality review continues in the existing full-shadow LAB.
+
+## Runtime recovery finding
+
+Preflight found the canonical `ndc-mission-core-triton` container stopped after
+the Docker runtime restarted. Docker reported exit code `255`, no OOM event and
+Compose configured `restart: "no"`. The accepted run began only after the exact
+same container was started and healthy; its container ID remained
+`4232fb040062a8384809e73612baa92b343ddcb60f00c23e7091eb4909959223`
+throughout the run.
+
+This is a deployment-lifecycle defect, not a perception overload result. The
+container must use a durable restart policy and an unhealthy-state watchdog;
+changing the TGS or detector graph cannot fix it.
+
+The live container was subsequently changed in place to
+`restart: unless-stopped`. Docker confirmed the service remained `running` and
+`healthy` with the same container ID, so applying the mitigation caused no
+inference restart. This protects the current container from process and Docker
+daemon restarts. The versioned Worker Compose source still declares
+`restart: "no"`; closing that source-level drift and inducing a controlled
+recovery remain a separate deployment gate.
+
+## Sealed compact evidence
+
+| File | SHA-256 |
+| --- | --- |
+| `result.json` | `bfc881aa943d46582a248b6024a413c3c4b3a6959fa4a674f6ec33d658fdcbdf` |
+| `graph-result.json` | `784d27b908137843f6adf08d267576a74b0676f5f98226923399ae59f538b78a` |
+| `tgs-result.json` | `c0245744892b66c7f24af3cd7ea9c8cd8d543114fffa70e32d8d5fef81054716` |
+| `worker-summary.json` | `c449246d14fb927b6490bcf657950b354f101fd3eb4b296bc0f6dd8543aafe10` |
+
+## Next gate
+
+Keep this integrated graph frozen. Complete the independent visual
+traversability review in the existing TGS LAB, concentrating on vegetation,
+compact obstacles, false occupied carpets, usable gaps and missing-LiDAR
+behavior. In parallel, close the Worker inference-runtime restart policy and
+verify an induced stop/recovery without changing result identity. Only after
+both gates pass may the candidate advance to repeated-loop soak and recovery
+acceptance; navigation and actuation remain off.