Files
dxdesigner-part-replacement/tools/random_packaged_acceptance.md
T

99 lines
5.2 KiB
Markdown

Frozen random acceptance
========================
`accept_random_packaged_ui.py` runs the actual EXE through external Windows UIA.
It imports no application service or source UI. Database access is SQLite
`mode=ro`; fixture files are generated separately before launching the EXE.
The admitted final candidate, original hashes and intended roots are in
`outputs/completion-validation-20261111/frozen-random-final-plan.json`. That file
is a plan, not evidence of an executed test. Wait for the exclusive CAD/UI slot.
Prepare the final fixture once, from the repository root (Python):
```python
import json, sys
from pathlib import Path
sys.path.insert(0, 'tools')
from accept_random_cad import prepare, workbook, fingerprint, write
p = json.loads(Path('outputs/completion-validation-20261111/frozen-random-final-plan.json').read_text(encoding='utf-8'))
assert fingerprint(Path(p['source']).parent) == p['source_hashes']
root = Path(p['root'])
m = prepare(root, Path(p['source']), json.loads(Path(p['graph']).read_text(encoding='utf-8')),
p['target_parts'], count=p['count'], cells=p['cells'])
# No seed argument: draw from system randomness after tools stabilize.
# Existing independent baseline belongs to the unchanged original source.
m['baseline'] = p['graph']
write(root / 'manifest.json', m)
workbook(root, m)
```
With the CAD/UI slot granted, launch/import/save/confirm/execute/pause/resume:
```powershell
python -X utf8 tools/accept_random_packaged_ui.py --exe outputs/completion-validation-20261111/app/DxPartReplace/DxPartReplace.exe --expected-sha256 0fb92b15ab57fa8765853ae63d0a7ef098280c21b40ec0e3768a726009e88015 --root outputs/completion-validation-20261111/frozen-random-final20 --stage execute --run-native
```
Then run independent native ASCII/Keyin verification, using the same slot:
```powershell
python -X utf8 tools/accept_random_cad.py --root outputs/completion-validation-20261111/frozen-random-final20 --verify
```
The driver leaves its exact app process available for review. It never kills an
app after a timeout. Continue only with that root's recorded `--pid`; attachment
requires its existing process/EXE/hash binding. `--stage import` creates tasks
and imports source drafts without executing a queue. Current candidate import
persists `pending_bom` drafts immediately and defers CAD checks to execution.
Evidence: timestamped `packaged-ui-report-*.json` (including failures), EXE log,
`ui-session.json`, `packaged-recording/*/frame-*.png`, actual queue/resource
journals, desktop observation and independent report. A passed import report
does not prove execution, pause/resume, or independent schematic correctness.
Supplemental controls implemented, awaiting an exclusive real CAD/UI slot:
* Valid Cell: generator's `--supplemental valid_cell --count 1`; driver clicks
the real rule dialog and loads/selects an authentic library Cell.
* Force: generator's `--supplemental force_stop --count 1`; driver observes the
worker stage, then click `forceStopButton`, confirm `强制结束当前操作`, and
perform `recoverQuickRunButton` / `核验并恢复`. Never edit worker progress.
* Resource: standalone fixture plus `--resource-hold`; test-owned external
desktop reference only; after completed saved
business, release that actor and click `cleanupBatchResources`. Preserve
failure/success history and prove saved results were not executed again.
* PON: prepare a separate pristine PON family/graph with authentic compatible
`--part-prefix` and target parts. Normal directory/import/execution controls
reuse this driver; no ordinary-family outcomes imply PON acceptance.
Review record
-------------
Normal completion requires queue phase `finished`, no running entries, settled
supervisor journals, positive desktop/process release evidence, and a disabled
UI stop control. Execute may legitimately be disabled after success. Independent
schematic verification remains mandatory.
Pause is clicked only while a real entry runs after at least one completed entry.
Resume waits for pending/paused entries and settled resources, uses the actual
UI confirmation, and asserts unchanged queue filename, ordered task IDs and plan
hashes. Previously saved UID event counts must remain unchanged.
Blocked/uncertain entries stop the normal driver. It never clears uncertainty,
edits the DB, fabricates progress, retries failed work, or kills an app on timeout.
Only the explicit force supplemental clicks force/recovery. Resource cleanup
compares saved business-result hashes and saved UID counts before/after, without
clicking Execute again. Missing real fault triggers fail acceptance.
The existing data-root option and process-local TEMP/TMP isolate the real EXE.
Windows process image and expected EXE SHA256 establish identity. Attachment
requires the isolated root's recorded PID/EXE/hash binding. Only UIA changes app
state; original source fingerprints are checked before/after. Independent native
observation checks CAD exit, desktop disappearance, foreground/input isolation
and responsiveness. There is no source Qt window, service shortcut, DB mutation
or simulated worker.
Verified scope: clean import-only on the old candidate. Supplemental and final
candidate execution await actual runs; their implementation is not proof of CAD
behavior.