P23-03: 生活の観察

This commit is contained in:
Agent Control Runner
2026-10-16 19:18:11 +09:00
parent 26f60095bf
commit 059c2f743b
6 changed files with 285 additions and 4 deletions
@@ -3707,7 +3707,7 @@
"template": "P23 life observation: <checks> checks, <failures> failures",
"required": true,
"fixed_check_count_used": false,
"source_sha256_at_schema_audit": "5c52db4dd70f9702c096e918d849703d14aec1acb88bb2458ca01049da71ce3c",
"source_sha256_at_schema_audit": "750254af7d4b8b00508e25d1ef744837d3a7d14ddebe12398e8014bf5e8c68e8",
"authority": {
"path": "docs/spec/acceptance/P23-03.json",
"scope": "component and read-only observation only; native unexecuted; whole unit blocked"
@@ -0,0 +1,69 @@
# Repair36 worker submission
Scope: world_preview observation labels, independent checker, Python/GD controls.
No Runtime, service, data, art, policy, original negative, or acceptance changes.
Input authority: registered BUILD `audit-support/badge-repair36-exact-inputs/EXACT_INPUTS.json`, source 26f60095; original paths are provenance identifiers.
The resolver places fixed cabin rectangles/badges first, then business, supply,
landing, aggregation in array order. A positive-area collision advances to the
first blocking rectangle's right edge and repeats. With finite rectangles of
positive width this always terminates on the anchor row; the specified upper
and lower fallback rows are therefore unreachable. No margin, pixel threshold,
clipping, truncation, member alteration, or new viewport rule is introduced.
Original anchors and ordered blocker identities are retained in deep-copy reports.
Traffic obstacles use the resolved positions; actor placement/traffic subtraction
and cabin/transit guards retain their existing implementations.
The checker separately projects anchors back into the original report shape for
existing raw-state queue/business/aggregation recomputation. It independently
replays label geometry and validates reported positions and blocker identities.
All label pairs must have zero positive-area intersection. Both production image
review consumers require explicit indexed `judgments`; the legacy `rows` helper
interface remains available for unchanged historical unit controls only. Compact
bits are never accepted for new image-review evidence.
## Static all40 comparison
Inputs: all40 original manifests plus their unchanged original requests in
`p23-normal-os-26f-r17`, read only. This is a counterfactual static projection,
not a new rendered screenshot or native/OS acceptance.
Original manifest label pairs: exactly one intersection, B/writer-request-0021,
rows 5 (queue) and 26 (aggregation). Resolved labels: zero intersections in all
40 frames, including fixed cabin labels. Thirty frames have no movement.
The following is the complete movement list in local canvas coordinates:
| Frame | Label | Original | Resolved | displaced_by |
| --- | --- | --- | --- | --- |
| A/reader-request-0003 | landing/0 | 563,468 | 755,468 | cabin/0/rect |
| A/writer-request-0014 | landing/0 | 563,468 | 845,468 | cabin/0/badge |
| A/writer-request-0021 | landing/0 | 563,468 | 875,468 | cabin/0/badge |
| A/writer-request-0025 | landing/0 | 563,468 | 855,468 | cabin/0/badge |
| A/writer-request-0030 | landing/0 | 563,468 | 755,468 | cabin/0/rect |
| B/reader-request-0003 | landing/1 | 819,468 | 1111,468 | cabin/0/badge, cabin/1/badge |
| B/writer-request-0014 | landing/0 | 563,468 | 1101,468 | cabin/0/badge, cabin/1/badge |
| B/writer-request-0014 | landing/1 | 819,468 | 1217,468 | cabin/1/badge, landing/0 |
| B/writer-request-0021 | landing/0 | 563,468 | 755,468 | cabin/0/rect |
| B/writer-request-0021 | landing/1 | 819,468 | 1011,468 | cabin/1/rect |
| B/writer-request-0021 | aggregation/0 | 563,372 | 825,372 | landing/2, cabin/0/badge |
| B/writer-request-0025 | landing/1 | 819,468 | 1111,468 | cabin/0/badge, cabin/1/badge |
| B/writer-request-0030 | landing/1 | 819,468 | 1111,468 | cabin/0/badge, cabin/1/badge |
The affected queue anchor remains local [563,372], corresponding to original
manifest screenshot [642,469]. Its five members remain hidden and its complete
label remains unchanged. The historical reviewer index24/text26 disagreement
and all prior negatives remain untouched; no staff defect or acceptance credit
is inferred.
## Validation boundary
Worker runs Python syntax/local semantic tests only. GD controls are added but
not executed here. Root still must replay currentbad, run localized/native,
new normal OS A/B with indexed independent image review, then one full regression
and formal review/integration/ledger in the recorded order. No formal acceptance,
original88 credit, Windows, held-gesture, performance, or asset credit is claimed.
Local result: Python AST syntax PASS; 146 Python tests PASS (142 retained plus
four focused controls). All 142 pre-existing test method ASTs match HEAD exactly.
`git diff --check` PASS. Only the existing P23 GD marker source SHA was refreshed;
no marker requirements or policies changed. No Godot/native/OS/full was run.