refactor(ui): reuse canonical graph for M4.8 regression

2026-08-25 00:08:33 +03:00
parent 669bb64423
commit 1084984da2
9 changed files with 322 additions and 261 deletions
@@ -160,6 +160,29 @@ history of rejected approaches belong in Ops.

 Primary visual evidence is hosted in one reusable viewer frame.

+### Frozen Milestone 4 perception instruments
+
+Milestone 4 admits exactly two operator instruments inside the shared LAB page
+template:
+
+1. the M4.8 assisted-correction instrument for editing frozen Worker 006
+   proposals and adding exact-frame misses;
+2. the M4.7 Canonical Reference Graph instrument for synchronized
+   `VIDEO/CAMERA/3D/PLAN/SEMANTICS`, Worker proposals, current LiDAR, bounded
+   local SLAM, rolling occupancy and metric-obstacle inspection.
+
+A later M4.8 regression or quality run supplies typed case navigation, manual
+anchor overlays and result copy to one of these instruments. It does not create
+an experiment-named player, viewer, Rerun configuration, timeline, splitter,
+spatial renderer or loading grammar. A manual exact-frame annotation is shown
+only on the annotated frame unless a separately evidenced and reviewable
+tracking result exists; ordinary model proposals and graph layers continue on
+their own full recorded timeline.
+
+A third perception instrument is a product-surface change. It requires explicit
+product-owner agreement before implementation and cannot be introduced by a LAB
+adapter, result component or experiment configuration.
+
 The viewer frame must:

 - provide the canonical expand/restore `IconButton`;
@@ -1384,12 +1384,66 @@ clearance and passability remain functions of admitted LiDAR/local occupancy,
 uncertainty and the configured vehicle footprint. Physical live, navigation,
 commands, actuation and collision-safety authority remain false.

-The regression reuses the held M4.8 recorded viewer and its CAMERA/3D/PLAN,
-single media clock, timeline and bounded spatial cache. Each assisted anchor and
-the frozen Worker objects are shown only on their exact source frame; the UI does
-not drift a manually drawn rectangle across subsequent frames or fabricate a
-track. A later Worker candidate must publish another immutable M4.8R run against
-the frozen seed, leaving this baseline available for before/after comparison.
+The regression reuses the accepted M4.7 Canonical Reference Graph instrument,
+not an experiment-local M4.8 viewer. It therefore retains the complete
+VIDEO/CAMERA/3D/PLAN/SEMANTICS graph, one media clock, Worker proposals, current
+LiDAR, bounded local SLAM, rolling occupancy and metric obstacles. Each assisted
+anchor is an additional exact-frame overlay and is shown only on its source
+frame; the UI does not drift a manually drawn rectangle across subsequent
+frames or fabricate a track. A later Worker candidate must publish another
+immutable M4.8R run against the frozen seed, leaving this baseline available for
+before/after comparison.
+
+### 2026-08-25 — perception interface freeze and dataset/training boundary
+
+Mission Core now has two and only two admitted Milestone 4 perception
+instruments:
+
+1. the M4.8 assisted-correction instrument used to accept, edit, delete and add
+   class-free exact-frame boxes over frozen Worker 006 proposals;
+2. the M4.7 Canonical Reference Graph instrument used to inspect detector,
+   camera, semantic, LiDAR, local-SLAM, rolling and threat evidence on one full
+   recorded timeline.
+
+`M48SmallStaticRegressionVisual` is removed. M4.8R results bind to the immutable
+M4.7 LAB identity stored in their source pack and project only typed regression
+anchors and navigation into the second instrument. Creating another player or
+viewer for an experiment is prohibited unless the product owner first approves
+a third reusable instrument. This is enforced by the M4.8R frontend contract
+test as well as the product UI canon.
+
+Worker 006 remains an inference-serving boundary. Its frozen `5,236` M4.8 boxes
+are model output, not an online-learning state. `0/14` in M4.8R1 means that none
+of the fourteen operator-added miss anchors reached IoU `0.50` against those
+frozen boxes; it does not mean that Worker 006 emitted no detections.
+
+The assisted-correction instrument is a valid annotation front end but is not,
+by itself, a train-ready dataset system. A released camera-obstacle dataset must
+add the following immutable control plane around its corrections:
+
+- exact source frame, raw/corrected optical representation, calibration,
+  preprocessing and frozen model identities;
+- explicit annotation completeness and negative-frame state rather than
+  treating absence of a box as background truth;
+- revisioned reviewer and adjudication identity, with candidate-assisted data
+  kept distinct from independent truth;
+- route/time-block train, development and sealed validation partitions to
+  prevent adjacent-frame leakage;
+- one versioned task taxonomy. The first task is class-free
+  `obstacle-presence`; `critical_corridor_obstacle` remains a planning/clearance
+  attribute and does not become a detector-class-specific collider rule;
+- a content-addressed internal manifest plus COCO detection export and hashes.
+  A lossy YOLO text folder is not the source of truth.
+
+The already deployed CVAT `v2.70.0` remains the offline review/export control
+plane behind this contract; it is not another Mission Core evidence viewer.
+Mission Core owns the two product instruments and immutable LAB publication.
+Training runs in a separate bounded Worker job/container, initially reusing the
+current YOLOX/PyTorch family, and may not mutate the serving Triton repository
+or the frozen Worker 006 process in place. A successful job seals weights,
+configuration, dataset release, metrics and ONNX identity; only then is a
+candidate loaded through the existing Triton seam and measured by a new
+append-only M4.8R run plus an independent validation gate.

 ## Implementation order