P23-03: 生活の観察

This commit is contained in:
Agent Control Runner
2026-10-15 07:59:43 +09:00
parent 860136f709
commit 03100beabe
13 changed files with 1386 additions and 101 deletions
@@ -0,0 +1,94 @@
# P23-03 worker status: submitted static work, implementation incomplete
Base commit: `860136f7099f5e3aee99da5798a627a77f2d8eea`.
This submission preserves the previous dirty work and adds genuine static
implementation. It is **not acceptance-ready**. No external Runtime capability
gap or out-of-scope defect has been established. Remaining implementation is
listed separately from pending Root evidence; no acceptance credit is claimed.
## Inspection and changes in this repair
- Re-read the recorded direction, Phase 0 map, previous status and renderer diff.
The single frame builder supplies drawing and the deep-copy public report.
No renderer redesign or additional renderer was introduced. Existing renderer
and component test bytes were retained.
- Added `tests/fixtures/p23-03/layout_comparison.gd`. Both layouts initialize the
public Runtime with seed 806042, reset budget and clock, identical production
defaults and the same recorded 48-person arrival script. Commands build the
existing elevator and corridors; B adds only the declared second elevator.
Preparation time derives from registered definition build times.
- Predeclared metric: sum of destination-arrival minutes for the entire cohort.
The checker rejects incomplete cohorts, changed actual arrival payloads,
additional layout changes, changed initial state, and non-strict improvement.
No observed improvement is claimed: the fixture has not run.
- Extended diagnostic_stream with hash-addressed typed event/input records.
The fixture records actual command results, domain movement events and
per-minute guest/path/EV/facility projections, including service history,
queues and density. Whole aggregate state/report pairs are sparse (30-minute
intervals); the existing P08 typed encoder is reused without truncation.
- Extended independent checking of room service/rescue occupants and business
enum fields, queue allocation/overflow, shared transforms, traffic, logistics
report selection/order and cabin geometry. Added Rect2 and vector-key decode.
The checker still refuses whole-unit acceptance.
## Exact remaining implementation (not external blockers)
1. Finish rest destination selection from typed path costs and event/defense
locks. Current room check validates eligibility and reported rest references;
it does not independently prove which rest destination Runtime selected.
2. Finish every actor/passenger mode and geometry, including cleaning,
rest/rescue/service slots, visual track history, badge text and screenshot
rect/pixel provenance. Cabin endpoints and rectangles are checked, but
unpaused prior-frame interpolation origin is not independently established.
3. Complete A/B definition/source/production pins and command/event/path
cross-validation, invoke the fixture from the Root collector, and add
representative real operational scenarios for all six categories and BACK.
The added experiment covers the fixed-demand FRONT transport comparison;
it does not establish actual service purchasing or all-six-category coverage.
4. Replace run_p23_life_observation.py readiness stub with the actual Root
collector using P08 owned-process/profile/source/engine/ordinary-audio and
CUA request/ACK helpers. It remains unchanged; it launches nothing.
Formal save, distinct-PID normal resume, both-layout FRONT/BACK/pan/zoom,
complete CUA receipts, all-run cleanup/protection and failure history remain
to be wired and independently checked.
5. Add missing meaningful negatives, especially normal OS/distinct-PID resume,
visual provenance and the remaining rest/motion branches. Existing 12 local
tests are not a complete seven-condition test matrix.
Steps (2) and (3) of the repair direction are therefore still incomplete;
step (4) is not implemented. This is a partial static submission, not a claim
that the requested repair is finished. Do not begin acceptance on an assumption
that all worker deliverables are complete.
## Checks actually run in this repair
- `python3 -B tools/test_p23_life_observation_evidence.py`: 12 test methods,
zero failures. Synthetic Python examples validate rejection logic only.
- AST syntax: all three P23 Python tools parsed.
- Marker JSON: old rows and metadata unchanged; exactly one appended row;
its SHA matches the unchanged GDScript component test.
- `git diff --check`: passed.
- No Godot parsing/runtime, native, OS, Runner or full regression execution.
- No git writes, subagents, integration, new assets or production changes.
suite-policy.json was not changed; no new suite was registered.
## Source SHA256 at local checks
| File | SHA256 |
| --- | --- |
| scenes/main/world_preview.gd | add770377c190a5ccdde95379539de9753f7dabc703c1d58452ba590c6eda2f0 |
| scripts/ui/room_presentation.gd | 55340b7b623532e861be1d21225635fd9169e2e40928597c1cd0458b99524dd9 |
| tests/ui/test_p23_life_observation.gd | b9afa66ec73d3a5cadf86f9a71f0d13f5b085c3cc9a534dfe29b2bed0e2e76a0 |
| tests/fixtures/p23-03/diagnostic_stream.gd | ff178a094fc310b0640d923a15058c6f6fcf859551c14f6428a5709dcd24260e |
| tests/fixtures/p23-03/layout_comparison.gd | ddf00268b8e32e3c5c078ebb57ada321382ecc7c5fbd87efad7e9dfbe768df6b |
| tools/run_p23_life_observation.py | 53c2f0638d4a861cf4fcb664504c25df33ae5478207867654f1c8cdf58fa23c7 |
| tools/check_p23_life_observation_evidence.py | 318b253a08248a753b88902e382edc5f01cbce6f1ed8092359698b69d3bb0503 |
| tools/test_p23_life_observation_evidence.py | 0f20d3de20b44f46f8738c0d022fcc1179656014e85fe31754f00b87ec2c288a |
## Root evidence: pending
After genuine completion and freezing: both-layout native and actual normal OS
observation, full CUA trace, formal save and distinct-PID resume, mandatory exact
version full regression and independent Opus review. These absent Root results
are **pending**, not blocked. The fixed 30-package / 88-unit denominator and all
P08/P22/P25/Windows/held-gesture/performance exclusions remain unchanged.