docs(perception): record M4.8R2 occupancy gate

This commit is contained in:
DCCONSTRUCTIONS
2026-08-26 11:52:09 +03:00
parent 6290c75210
commit 5983883ffe
3 changed files with 47 additions and 0 deletions
@@ -912,6 +912,17 @@ authority.
Exit: local collision-space quality and deadline gates pass in replay and
shadow. This still does not authorize control.
M4.8R2 now supplies the bounded pre-shadow qualification for the first item.
On eleven operator-assisted static passage anchors, the accepted current/rolling
graph covers `6/11`, while the additive sealed low-step signal covers `10/11`;
the critical `08 m` subset is `9/9`, four canonical engineering anchors remain
matched and false-free count is zero. The item remains unchecked because this
is not independent truth and the candidate is not yet part of the Worker
pipeline. Its next gate is a full `4,489`-frame occupied-only Worker 006 shadow
with FPS, latency, occupancy growth, component-count and capacity-drop bounds.
Dynamic decay, Nav2 publication and planner-authoritative free space remain
separate later work.
### L6 — alternative odometry/mapping
This stage starts only when a source supplies unregistered sensor scans. Add
@@ -179,6 +179,13 @@ 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.
M4.8R2 is the canonical static-occupancy example of this rule. Its report and
eleven typed anchors bind to the accepted M4.7 LAB identity, use the same
`VIDEO/CAMERA/3D/PLAN/SEMANTICS` timeline and add no renderer or transport. A
qualified anchor means the additive occupied-only candidate has LiDAR support
inside that exact reviewed camera extent; it does not turn the extent into a 3D
collider or grant clearance, free-space, navigation or safety authority.
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.
@@ -1475,6 +1475,35 @@ candidate acceptance, production acceptance, navigation, safety, commands and
actuation authority remain false. The next action is operator adjudication in
the existing M4.8 review instrument.
### 2026-08-26 — M4.8R2 conservative static-occupancy qualification
M4.8R2 moves the small-static problem out of semantic detector tuning and into
the existing LiDAR occupancy boundary. RF-DETR, native fisheye input and its
risk taxonomy are unchanged. The deterministic qualification consumes the
accepted M4.7 current/rolling graph plus the frozen M4.8R1 operator-assisted
static anchors. It evaluates one already sealed local-surface signal,
`point_step_candidate`, only as additive occupied-or-unknown evidence. It never
clears a cell, claims free space, starts detector inference or uses the GPU.
Immutable result
`m48-static-occupancy-qualification-568024554cff011332ff19ca4739f70555a6232c0607dec2a71be6db408ea69a`
contains `11` static passage anchors. The accepted graph qualifies `6/11`; the
additive low-step candidate qualifies `10/11`. In the critical `08 m` band the
change is `6/9 → 9/9`; in the `812 m` approach band it is `0/2 → 1/2`. All
four existing canonical engineering anchors remain matched and the result
contains zero false-free claims. One approach anchor remains explicitly
`unresolved-unknown-never-free`.
This is candidate-visible development evidence, not independent truth or a
production cutover. The result accepts only a bounded next action: integrate
the low-step contribution as an occupied-only Worker shadow and replay all
`4,489` frames on Worker 006 while measuring effective FPS, stage latency,
occupancy/component growth, capacity drops and the same static anchors. The LAB
reuses the M4.7 `VIDEO/CAMERA/3D/PLAN/SEMANTICS` viewer and projects the eleven
anchors onto its immutable timeline; no third perception instrument exists.
Navigation, commands, actuation, physical clearance and collision-safety
authority remain false.
## Implementation order
The implementation sequence is intentionally strict: