feat: add local surface review triage

docs/13_LIDAR_WORKER_PRODUCT_AND_ROADMAP.md

@@ -2,9 +2,9 @@
 
 Date: 2026-07-25
 Status: accepted architecture plan; L0/L1 implemented; L2 diagnostic A/B
-complete; full GOOSE and RELLIS qualification complete; L2.6b K1 replay
-local-surface temporal qualification implemented; operator review and live
-shadow next
+complete; full GOOSE and RELLIS qualification complete; L2.6c K1 replay
+local-surface temporal qualification and operator triage implemented;
+residual explainability and live shadow next
 Scope: passively received real-time K1 point/pose evidence, immutable replay and
 future live shadow processing
 Explicitly out of scope: K1 firmware modification, a new onboard exporter, new
@@ -411,6 +411,12 @@ Dataset expansion is no longer the next gate.
   layer with source, confidence, freshness and conflict fields.
 - [x] Qualify next-frame prediction and temporal stability with the evaluated
   frame excluded from prediction input.
+- [x] Publish a deterministic, read-only operator queue for temporal
+  transitions and heavy prediction tails, with source-frame navigation and no
+  safety authority.
+- [ ] Preserve the prior prediction plane and point/cell-aligned residual
+  evidence so a selected tail can be explained spatially instead of only by an
+  aggregate p95.
 - [ ] Complete the remaining qualification report with per-frame latency,
   point age, obstacle preservation and memory growth.
 - [ ] Replay the same profiles through a bounded latest-wins live-shadow queue;
@@ -418,8 +424,8 @@ Dataset expansion is no longer the next gate.
 
 The implemented `missioncore.k1-local-surface/v1` derivative is reproducible
 through `experiments/perception/run_k1_local_surface.py` and is exposed
-read-only through `GET /api/v1/lidar/local-surfaces` plus the bound frame
-and timeline endpoints. **Парк → Диагностика LiDAR** reuses the five
+read-only through `GET /api/v1/lidar/local-surfaces` plus the bound frame,
+timeline and review endpoints. **Парк → Диагностика LiDAR** reuses the five
 RAVNOVES00 scene selectors and also exposes a clickable timeline over the
 complete recording. The selected source frame shows observed surface, observed
 occupied-above-surface, negative outlier, unclassified evidence and yellow
@@ -451,6 +457,26 @@ which proves that the review layer is active but also that it is broad and
 still requires independent review. Free space remains unavailable; none of
 these metrics grants navigation, command or safety authority.
 
+The `missioncore.k1-local-surface-review/v1` triage contract selects a frame
+when its prediction p95 residual is at least `0.45 m`, prediction inlier
+fraction falls below `85%`, or one of the content-bound temporal thresholds is
+crossed. It groups observations separated by no more than two frames into one
+episode and ranks threshold exceedance without changing the source artifact.
+RAVNOVES00 produced `37` review frames in `21` episodes: `18` prediction-tail
+frames, `14` inlier drops and `12` height transitions. Four frames have a
+normalized attention score of at least `2.0`.
+
+The four highest-priority frames are concentrated in two episodes. Source
+frame `1195` crosses height, slope and roughness thresholds together; the new
+surface regime remains on `1196`, so the evidence is a sustained transition,
+not merely a one-frame numerical spike. Source frames `1253–1254` reach
+`1.047–1.083 m` p95 residual and `68.0–63.3%` inliers while the fitted global
+surface remains stable. This separates a local prediction-tail problem from a
+global plane-transition problem. The recording has no independent ground truth
+for either episode, so the UI calls them review evidence rather than algorithm
+failures. The next replay slice must retain and display the prior-plane
+cell/point residuals before changing fit thresholds or entering live shadow.
+
 Exit: one immutable K1 session yields both a persistent reconstruction and a
 bounded local world state without hard-coded terrain height or scanner-side
 changes.
@@ -536,8 +562,9 @@ enough for the current decision. The first K1 local-surface replay slice now
 covers all available `RAVNOVES00` samples and is visible in the operator
 interface. Leave-current-frame-out temporal qualification now covers `525`
 samples: the median surface error is stable, while the p95 tail remains too
-large for a free-space claim. The highest-value immediate work is review of the
-`12` temporal jumps and worst prediction tails, then a bounded live-shadow
-queue and separate dynamic-observation layer. Nvblox, raw-scan detectors and
-alternative SLAM remain optional later gates because the current report
-contract does not carry their required ray/timing semantics.
+large for a free-space claim. Deterministic triage has reduced the first manual
+inspection set to four high-priority frames in two episodes. The highest-value
+immediate work is prior-plane residual explainability for those episodes, then
+a bounded live-shadow queue and separate dynamic-observation layer. Nvblox,
+raw-scan detectors and alternative SLAM remain optional later gates because the
+current report contract does not carry their required ray/timing semantics.

docs/14_LIDAR_DATASET_GATEWAY.md

@@ -51,6 +51,9 @@ Implemented now:
   explicit pose-binding age and conservative observed occupied/unknown policy;
 - leave-current-frame-out next-frame qualification over `525` samples,
   temporal jump evidence and unverified local-discontinuity candidates;
+- deterministic `missioncore.k1-local-surface-review/v1` replay triage:
+  `37` attention frames grouped into `21` episodes, with four
+  high-priority frames and direct source-frame navigation;
 - a provider-neutral read-only local-surface view in
   **Парк → Диагностика LiDAR**, synchronized to the five existing
   RAVNOVES00 scene selectors and a clickable complete-recording timeline;
@@ -86,6 +89,8 @@ Not implemented:
 - no RELLIS ROS bag admission, continuous synchronized playback or production
   promotion;
 - no ray-cleared free-space or planner-authoritative rolling occupancy map.
+- no point/cell-aligned prior-plane residual overlay yet; aggregate tail
+  evidence is not enough to identify its physical cause.
 
 ## Product surface boundary