Files
ARCADIA-Studio/harness/experiments/peer-union-e34-native-20260924/README.md
T

49 lines
6.2 KiB
Markdown

# e34 native Windows evidence checkpoint — 2026-09-24
Recorded native work against source `e34ba27e26ede3ab55f276acc250d92b2e0c3204` passed the scoped Scenario, geometry and OA relink checks below. The archive retains the failed pointer G2 attempt and all recovery evidence. It contains **no executable or DLL** and does not itself establish release readiness.
`evidence.zip` contains 1,264 raw members, 45,897,130 uncompressed bytes. Archive size: 23,699,323 bytes; SHA256 `3aea888bca130542e731de191bb79676ab00a222f37f9884b366cf22818ffa1a`. `manifest.json` pins every member and records its original local path. Original logs, snapshots and prior preparation documents are byte-preserved; historical PREPARED/UNRUN wording inside those documents is historical, not this checkpoint's outcome.
| Recorded evidence | Result |
| --- | --- |
| Source-to-model checkpoints | 33 PASS, one retained G2 FAIL across 34 receipts |
| Native transition checks | Five PASS: baseline, Cancel, relink, fresh reopen, unchanged reapply/save |
| Transition oracle runtime admission | 17 expected outcomes: four positive, 13 negative, with raw rejection boundaries |
| Native captures | All 245 PNG pins checked; 38 selected keyframes independently inspected |
| Source binding | All 760 captured build inputs match current frozen source and Git blobs at e34 |
The portable verifier completed with exit 0 and empty stderr, including the 760 Git-blob checks. Its exact command, unchanged artifact pins and raw stdout/stderr are in `verification.json` and the accompanying logs. This verifies recorded evidence and saved-model equality; it does not replay native actions or rerun an evaluator.
## Results and boundaries
Scenario evidence covers the original Text Tab/Undo/Redo/save, saved-view name and occurrence-value Undo/Redo, same-process S2 Text roundtrip→S2L→Undo, pending inspector draft across Connections/Properties, incompatible-field rejection and Revert, independent occurrence selection across saved views, and fresh-process reopen. The S2L checkpoint changes only caller lane x200→220; Undo restores exact predicted S2 source bytes, including the Unicode/mixed-line-ending preservation prediction.
Geometry evidence covers ordinary attachment G1, bend insertion G2, bend move G3, semantic-port position G4, relevant Undo/Redo, and fresh-process reopen with complete-model equality. The first pointer G2 trial failed at (526.25,309.06247) versus the unchanged expected (526,310); Undo restored exact G1. A separately predicted Properties Actions → Add bend trial then reached the original G2. Pending-draft arrow probes were blocked. After Revert, a clean arrow actually panned the camera 60 pixels with unchanged saved model.
Transition begins with an already recreated OA root; it is **not new native deletion/recreation coverage**. Reviewed Cancel preserves the original 13-file bundle. Reviewed relink retains target name/description and creates trace `d278274e-8153-44e7-8b38-a48f5ad73c3c` in a new `traceability.arcadia` plus an entry-file include. Fresh reopen and unchanged reapply/save retain all 14 resulting files byte-for-byte and reuse the exact first identity receipt. Its SHA256 is `fa58f8f015529a9a1f44428b44df13c4edb35a1875c0550ee3e92464ff6ccd90`.
The original 40-action process was absent after interruption and was **not recorded as a normal close**. Later normal-close observations are preserved separately. Transition's first close observation has `closedNormally:false` despite its misleading label; the subsequent observation has no Arcadia windows and `closedNormally:true`. Final closure is separately observed. No GUI exit code is inferred from the launch tool.
Remaining limits: shorter Scenario window exercised Revert, not Apply; dirty-State middle-button pan and Escape during a held drag remain unobserved; unchanged source alone does not prove absence of an extra no-op history entry. This is not comprehensive accessibility, warning-multiset, performance, or all-platform acceptance. See archived `review/REVIEW.md` and `review/TRANSITION-REVIEW.md` for the independent review and precise capture references.
## Recheck and inspect
With Python 3 standard library, run from this directory:
```text
python -B verify.py
python -B verify.py --repo PATH_TO_ARCADIA_CLONE
```
The first command reads the ZIP directly without extracting it. The optional repository command additionally checks exact 760 source blobs using Git; the clone must contain e34. Neither command executes archived code or native binaries. For visual inspection, extract the ZIP with a normal ZIP reader and open `native/captures/initial/` or `native/captures/resumed/`; their action JSONs associate every frame with its recorded action.
Archive layout: `native/` holds captures, run receipts/full source snapshots and build provenance; `preparation/` holds fixed model/source expectations, comparator source and recorded comparator controls; `transition-admission/` and `runtime-controls/` hold the reviewed oracle adaptation and 17 actual controls; `review/` holds independent audits and the local package producer. The verifier checks pins, source-model comparisons including the exact G2 failure, source restoration groups, transition bundle/identity retention, action boundaries, raw runtime-control outcomes and optional Git source binding. It cannot independently prove GUI gestures from saved bytes.
Historical source-runner synthetic control directories are omitted, with their summary retained. The prior test-gate archive is already published at `be7470afca4894c265d61800f9b94d31444e7234:harness/experiments/peer-union-e34-windows-20260923` and is not duplicated. Build identities remain in this package, including the GUI SHA256 `d8f60c910c6316878538689a5ae0cc6aaf173735dcf7f5aeef75026ce86f4931` and the required `libunwind.dll` launch-recovery pin. A runnable prerelease is a separate deliverable and must include that runtime dependency.
This version-controlled checkpoint preserves the reviewed archive and verifier
bytes from the local proposal. The parent read the verifier and independent
review before publication. `.gitattributes` preserves the exact raw Windows
logs; their original local paths are provenance, not cross-machine handoff paths.
The source commit and repository-relative paths above are the portable references.