feat(perception): complete E32 full replay

2026-07-27 12:49:42 +03:00
parent 10c306f162
commit 32257fc2b4
7 changed files with 2940 additions and 16 deletions
@@ -645,19 +645,26 @@ with a complete `missioncore.laboratory-method/v1` manifest.
 The current E30 AI-assisted engineering generation is
 `e30-engineering-generation-62a4fea10dea9b77f69ceac1af5bf0e4928d9c7716083c22258a03670fe5bd4f`.
 It covers all `486` selected items: `403` confirmed, `81` corrected and `2`
-retained as `insufficient-evidence` human exceptions. The earlier five-case
-generation and its human draft remain immutable history. Frames 1213, 162 and
-1823 are now automatic engineering outcomes; only geometry frames 2622 and
-4147 require a human decision. It found no systematic camera↔LiDAR
-registration failure in the review set. The dominant actionable signals are
-detector errors (`110`, including barrier false positives and 21 missed
-class-bearing objects), sparse occupied support (`98`), held-track time
-freshness (`92`) and self points (`10`). This is diagnostic engineering
-evidence, not human ground truth, detection accuracy or safety acceptance. The
-exception budget is 1% only for diagnostic residual work and never overrides a
-repeated-cause or high-impact blocker. The current queue is `2 / 486` (`0.41%`);
-E31 remains blocked until those two decisions freeze the E30 minimum correction
-set.
+initially routed human exceptions. The immutable human generation
+`e30-review-generation-7982a882558d0be690b4c7092e328c080bfcbf52478a220452be7e887a588250`
+closes both: geometry `2622:4` is background/noise and `4147:3` is a real
+occupied object. The earlier five-case generation and drafts remain immutable
+history. E30 found no systematic camera↔LiDAR registration failure in the
+review set. The dominant actionable signals remain detector errors (`110`,
+including barrier false positives and 21 missed class-bearing objects), sparse
+occupied support (`98`), held-track time freshness (`92`) and self points
+(`10`). This is diagnostic engineering evidence, not human ground truth,
+detection accuracy or safety acceptance.
+
+E31 accepted immutable source-scoped profile
+`e31-source-qualification-b2460a5eb143688c7eea6821b2277e13aea79868abe81d83f7e78548c119159a`.
+E32 then produced
+`e32-track-geometry-a14ca0e7fb3850ca0dfa3c41634e1b490a2d58ab74d101afc6d6921fbdb0e6fd`:
+all 4,489 E29 frames reproduce exactly, all source/object/point claims close,
+the self and exact correction set is applied without a generic geometry mask,
+and exclusive PointSlab ownership is enforced. Conflict count remains `38`;
+the result is accepted only as the source-scoped diagnostic/shadow input for
+E33 and is not promoted as a detector-accuracy or staleness improvement.

 Execution is strictly sequential through E33: E30 determines what E31 is
 allowed to change; E31 determines the E32 profile; E32 determines the E33
@@ -171,9 +171,31 @@ RAVNOVES00 diagnostic binding and does not claim that host arrival is hardware
 firing time. The factory KB4 identity is exact, but a measured
 calibration-target residual and physical body/mount dimensions are
 unavailable. The accepted profile is therefore source-session scoped and
-cannot transfer to another mount. A6/E32 full replay is the next critical-path
-implementation and must bind every published frame to both `TrackGeometry v1`
-and the accepted E31 profile.
+cannot transfer to another mount.
+
+A6/E32 is complete in immutable result
+`e32-track-geometry-a14ca0e7fb3850ca0dfa3c41634e1b490a2d58ab74d101afc6d6921fbdb0e6fd`.
+It reproduces all 4,489 E29 frames exactly before correction and binds every
+published frame to `TrackGeometry v1` plus the accepted E31 profile. All
+20,513 semantic observations, 21,321 geometry clusters and 2,125,813 qualified
+point claims close without hidden loss. The replay removes 385 source-scoped
+self observations and three evidenced geometry clusters, withholds 1,558
+unqualified ranges and resolves 4,461 overlapping point claims. Of 709
+arbitrated semantic observations, 563 retain `agree` and 146 conservatively
+become `unknown`. The 38 E29 conflicts remain 38; A6 is accepted as a
+diagnostic/shadow contract, not as detector-accuracy improvement. A7/E33
+recorded-source-paced worker execution is now the critical path.
+
+- [x] Reproduce all 4,489 immutable E29 frames with the exact frozen profile
+  before applying E31/E30 changes.
+- [x] Apply only the admitted semantic self-mask, two exact geometry
+  corrections and the complete A3 human exception set.
+- [x] Enforce exclusive PointSlab ownership and journal every arbitration,
+  excluded claim and conservative state transition.
+- [x] Compare E29/E32 by status, class, range, 60-second scene interval and
+  cause while preserving source availability and unknown/free-space policy.
+- [x] Add a digest-bound compact binary PointSlab encoding and strict
+  TrackGeometryFrame reconstruction/validation.

 ### A3 residual and human-exception policy

@@ -0,0 +1,92 @@
+# ADR 0025 — E32 full replay and exclusive point ownership
+
+Date: 2026-07-27
+
+Status: accepted for source-scoped diagnostic/shadow use
+
+## Context
+
+A5 defined `TrackGeometry v1` and `PointSlab`, but E29 did not retain the exact
+frame-local point indices that produced each semantic and geometry result.
+E32 therefore had to recover those indices without changing the frozen E29
+thresholds, apply only the accepted E31/E30 corrections and prove that the
+translated result had no hidden source loss.
+
+The accepted E31 profile is limited to the immutable RAVNOVES00 source session.
+It admits one normalized person self-mask, rejects a generic geometry mask and
+binds two exact geometry self-corrections. The A3 human generation additionally
+rejects geometry cluster `2622:4` as background/noise and retains `4147:3` as a
+real occupied object.
+
+## Decision
+
+1. E32 recomputes E29 support indices using the exact E29 profile and immutable
+   camera, LiDAR, pose and local-surface inputs.
+2. Every recomputed frame must equal the stored E29 frame before any E31/E30
+   correction is applied. A difference in any observation, cluster, metric,
+   policy or frame binding fails the replay closed.
+3. E32 applies only:
+   - the admitted E31 person self-mask by normalized bbox centre;
+   - the two exact E31 geometry correction item locators;
+   - the complete immutable A3 human geometry dispositions.
+4. The rejected generic E31 geometry mask is never applied. E29 thresholds are
+   never retuned and excluded points are never reclassified as free space.
+5. `PointSlab` grants one owner to each published source point. When two E29
+   camera observations claim the same point, E32 uses a deterministic
+   non-threshold policy: smallest bbox, then higher detector score, lower track
+   id and lower observation ordinal. A track that retains points stays `agree`;
+   a track that loses all exclusive points becomes explicit `unknown`.
+6. Camera-only and conflict observations cannot publish an E29 range derived
+   from support that failed the E29 qualification threshold. Their semantic
+   observation remains, but `metric_basis=unavailable` and `range_m=null`.
+7. A held observation is published only when an earlier current observation
+   for the same track exists in the replay. A held observation without such
+   provenance is explicitly journalled and omitted instead of inventing
+   `held_from_frame_index`.
+8. Compact storage uses deterministic `.npy` arrays for frame offsets,
+   frame-local source indices, map-frame float32 points and local owner
+   indices. The JSONL frame record retains the complete TrackGeometry table and
+   PointSlab reference. The public reader reconstructs and validates the exact
+   `TrackGeometryFrame`.
+9. Every excluded or altered E29 product is written to
+   `e29-e32-changes.jsonl` with its source locator, before/after state, reason,
+   decision identity when applicable and affected claim count.
+10. E32 remains diagnostic. It grants no command, navigation, traversability,
+    free-space or safety authority and does not modify the persistent map.
+
+## Accepted result
+
+The immutable result is
+`e32-track-geometry-a14ca0e7fb3850ca0dfa3c41634e1b490a2d58ab74d101afc6d6921fbdb0e6fd`.
+
+- 4,489 / 4,489 E29 frames reproduce exactly.
+- Source availability remains 3,928 available and 561 unavailable frames.
+- 20,513 semantic observations close as 20,119 published, 385 source-scoped
+  self-mask exclusions and 9 held observations without prior-current
+  provenance.
+- 21,321 geometry clusters close as 21,318 published, two exact E31
+  self-corrections and one A3 human background/noise exclusion. The A3
+  object-present cluster remains published.
+- 2,125,813 E29 qualified point claims close as 2,119,302 published rows,
+  2,050 explicitly excluded claims and 4,461 overlapping claims removed by
+  ownership arbitration.
+- 709 semantic observations required point-ownership arbitration: 563 retained
+  `agree`; 146 became `unknown`.
+- 1,558 ranges backed only by unqualified semantic support are withheld.
+- E29/E32 conflict count remains 38 and no conflict observation changes. E32
+  is not presented as a detector-accuracy improvement.
+- `agree` changes from 6,341 to 6,195, `single-source-camera` from 13,246 to
+  12,861, `unknown` from 888 to 1,025 and geometry-only from 21,321 to 21,318.
+
+## Consequences
+
+- A6 is complete as a reproducible, source-scoped TrackGeometry input for the
+  A7/E33 recorded-source-paced worker shadow.
+- E33 receives a closed ownership and source-accounting contract instead of
+  ambiguous E29 point claims.
+- The unchanged 38 conflicts and retained held observations remain perception
+  limitations. Runtime qualification must not describe them as corrected.
+- The 146 conservative `agree → unknown` transitions are an intentional
+  consequence of exclusive ownership, not threshold degradation.
+- A second source or changed mount must pass a new qualification; this E32
+  result cannot be transferred by assumption.