From 4aee179385aa3fefab22aecca668f3058d41024c Mon Sep 17 00:00:00 2001 From: "sorokin.oleg.edu" Date: Thu, 24 Sep 2026 15:42:21 -0700 Subject: [PATCH] Admit bounded Windows packaged GUI and independent machine QA --- .../prerelease37-independent-qa-20260924.md | 80 +++++++++++++++++++ .../prerelease37-native-gui-20260924.md | 31 +++++++ .../prerelease37-parent-visual-20260924.md | 35 ++++++++ 3 files changed, 146 insertions(+) create mode 100644 harness/reviews/prerelease37-independent-qa-20260924.md create mode 100644 harness/reviews/prerelease37-native-gui-20260924.md create mode 100644 harness/reviews/prerelease37-parent-visual-20260924.md diff --git a/harness/reviews/prerelease37-independent-qa-20260924.md b/harness/reviews/prerelease37-independent-qa-20260924.md new file mode 100644 index 00000000..5e24d6a7 --- /dev/null +++ b/harness/reviews/prerelease37-independent-qa-20260924.md @@ -0,0 +1,80 @@ +# Independent release-37 QA delivery reconciliation + +2026-09-24. Reviewer `/root/mixed_index_source_review` (GPT-6 Astra). **No material contradiction or blocker found in this bounded evidence review.** The exact package, captured State presentation/picker/reopen workflow and acknowledged-write server restart evidence support the reported limited QA result. Release admission/publication remain parent-owned. + +This review fetched immutable delivery `684061ec0f486825886c077a038f06d17896392e`, inspected archived helper code before executing the two read-only portable evaluators, independently reconciled raw records, and viewed original full-window JPEGs. It did not launch the product, use local GUI automation, build/Cargo, modify source, post remotely or publish anything. Only local inspection artifacts were written. + +## Exact artifacts and reproducible inspection + +- Delivery branch: `codex/prerelease-package-qa-9e8r6t-20260924`; directory `harness/prerelease37-independent-9e8r6t-20260924` at the immutable commit above. +- Evidence ZIP: **3,425,348 bytes**, SHA-256 **`cb084be5f316f1b3be45bc5493fe0cbd6076052198c09f9dbac441b8f2588e63`**, 190 unique members. Checked every member's size/hash against manifest, case-insensitive uniqueness and safe relative paths. All five detailed outer reports equal their archived copies. +- Product source: **`715036e132677dd2bc2fb37ed316d93dc9987aec`**, not the documentation delivery commit. +- Exact local Studio ZIP: `.local/prerelease37-assets/ArcadiaStudio-0.2.0-preview.integrated.20260924-x86_64-pc-windows-gnullvm.zip`, **34,168,661 bytes**, SHA-256 **`5554e32533e3cb2c33fb4143a5a0902733d5a8d47577303cb2842d268e7bc76d`**. Independently rehashed those bytes and all 565 contained files against the QA download receipt. Also checked all 564 SHA256SUMS entries and all 559 package input hashes; packaged `package.json` equals the archived metadata. + +Local reproducible evaluator: `inspect_independent_qa.py`, SHA-256 `46326eef94461bc87166c47ca34f05486648c59d734417c139ec54283e4af4b4`. Structured output: `independent-inspection.json`, SHA-256 `848d3b0894d76b8ed17e82760516f509e7be298dbf3fdbe8ceb3a9ccfc5bccf5`. Both reside alongside this report. Run from repository root: + +```powershell +& C:/Users/User/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/python.exe .local/prerelease37-integration-review-20260924/inspect_independent_qa.py +``` + +Prerequisites are the fetched immutable delivery, the exact Studio ZIP above, and previously delivered local `final04-complete-715036e.zip` plus `final04-complete-manifest.json` in this review directory. The evaluator pins all their expected hashes. It extracts evidence only into `independent-684061e`, verifies data without executing launch/network/build helpers, then runs source-reviewed `native-check.py` and `server-readback-check.py` with Python `-B`. Both exit 0 with empty stderr and JSON equal to their archived result. All 190 extracted files remain hash-identical and no file appears during either replay. These are offline replays, not additional native/server runtime trials or 23 newly executed checks. + +## Package, build and process binding + +The actual ZIP contents independently match the native/server run pins: + +| Artifact | Bytes | SHA-256 | +| --- | ---: | --- | +| `arcadia-studio.exe` | 119,396,352 | `2c03bd6cc3f1110c219e4f57ab94874a121590f83779239a93289b7170664399` | +| `arcadia-server.exe` | 40,228,352 | `6cebacc117e46fdab023c6402a5cebd6929fb55993ae8ff6b89cd1c6c1170237` | +| `libunwind.dll` | 190,464 | `aa282fa47bb2c4e210cea5e470cdbdca58568f0be0bc73e404d86ab9cd95ef6a` | + +Independently rehashed the retained final04 archive (`7cead0b6aa36a6af8fc9f8987186bfe0decd8fa2ebab094fc45387ae9daea632`) and all 462 members. All four package BuildEvidence pins match raw final04 summary/input/environment/relocation records. For each executable, the exact preserved **individual** build hash/size equals the ZIP bytes, the relocation record contains the literal Cargo artifact emitted in that build's raw output, the artifact is non-test, the output contains successful build-finished, the process exit is 0, and before/after source manifests equal the final04 manifest. Supplemental combined-build bytes are not substituted. This corroborates the supplied Cargo/source provenance; it is not an independent compiler rebuild or reproducibility claim. + +Native receipt download-proof/launcher pins match actual archived helper/receipt bytes, and executable/runtime pins match the exact ZIP above. All 37 capture receipts' timestamps fall inside their owning process interval and identify that executable path. Run 2 PID 8188 starts `22:02:26Z` and ends `22:11:16Z`; fresh run 3 PID 6904 starts `22:11:30Z` and ends `22:17:09Z`. Both record normal exit 0, unchanged binary/runtime and zero-byte stdout/stderr. Their child PATH is exactly `C:\Windows\System32;C:\Windows`, and cwd is the unrelated `C:\arcadia-package-qa-9e8r6t`. Existing user profile/recovery state remains; this is not a clean-OS/profile trial. + +## Native source and visual result + +The 9,662-byte native baseline, SHA-256 `2dae785d63a940094d8052f3dcddcb988c024b38df9780e026bccd516264dbfe`, equals the actual fixture inside the archived State/Traceability companion ZIP; that companion checksum passes. I independently located the unique Activate route waypoint needle inside `diagram readiness_view` and compared all five complete saved files against baseline with exactly one inserted waypoint object. No whitespace, number or newline normalization was used. + +| Saved snapshot | Added waypoint | Exact result | +| --- | --- | --- | +| `native-2/bend-saved.arcadia` | (530,385) | 9,705 bytes; `ea05d0e409687c7943dda0b93bc34036eb524f5f5033b912e71ad2eaf643a158` | +| `native-2/move-undo-saved.arcadia` | (526,310) | 9,705 bytes; `7f1fc18265a79b7d452e4a3de3f17671684e38c0f727cbdec15833196f46f114` | +| `native-2/move-redo-saved.arcadia` | (530,385) | Exact equality with bend-saved | +| `native-2/after-cancel-saved.arcadia` | (530,385) | Exact equality with bend-saved | +| `native-3/reopened-saved.arcadia` | (530,385) | Exact equality with bend-saved | + +The original (500,310) waypoint remains. All other source bytes—including semantic declarations, IDs, endpoints/attachments, node positions, labels, other routes and all other views—remain original. This is whole-source preservation evidence; I did not run a semantic parser or inspect an in-memory model/history stack. + +I viewed the full-window originals below, not their 16x16 icon companions. Paths are under `independent-684061e/` beside this report: + +- `native-2/09-add-bend-1.jpg`, `10-undo-added-bend-1.jpg`, `11-redo-added-bend-1.jpg`: visible selected-route square handles change two → one → two with corresponding Undo/Redo status. Add-bend Undo/Redo is visual evidence; no saved original/no-bend snapshot was invented. +- `native-2/14-undo-bend-move-1.jpg`, `16-redo-bend-move-1.jpg`: added handle returns near the original, then moves downward again. The byte oracle establishes coordinates. +- `native-2/19-reference-view-1.jpg`: distinct “04 Readiness — untouched reference” view, without the edited view's two bend handles. Full-byte comparison independently establishes its preservation. +- `native-2/24-transition-source-picker-1.jpg`: dropdown open inside Edit transition. `25-first-escape-picker-only-1.jpg`: dropdown closed, form still present with retained fields. +- **`native-2/26-second-escape-cancels-form-1.jpg`: form is still present.** I inspected this original at full resolution. The optimistic filename is false as an outcome description, and the delivery correctly overrides it in README/reports. It must not be cited as second-Escape cancellation. +- `native-2/28-explicit-cancel-1.jpg`: form closed; edited route remains. The subsequent saved file is unchanged. No form fields had been changed, and no internal history-stack behavior is inferred. +- `native-3/05-persisted-bend-selected-1.jpg`: fresh-process view 03 retains the downward bend and two square handles. `native-3/08-state-guide-topic-1.jpg` shows rendered States, modes & transitions guide content. +- `native-2/04-walkthrough-loaded-1.jpg`: bundled water-pressure walkthrough visibly loaded from the unrelated launch cwd, with 94 objects/62 links/1 model check. + +Checked the actual immutable 715036e Escape branches (`crates/oa-desktop/src/app.rs:2742`, `behavior/states.rs:283`, `app/escape.rs:18`) and issue-35 contract. They preserve the ordinary transition form's explicit Cancel policy; there is no generic second-Escape state-draft cancellation branch. Thus observed retention is consistent with the selected requirement that later Escape follows the surrounding workflow's existing policy. + +At 56% Fit, the State fixture labels are visibly small/crowded, especially in the reference view. Its displayed 12 model checks are retained fixture incompleteness. The record does not establish broad editing usability, readability, accessibility or complete State behavior. Guide rendering does not establish a network-disabled environment. + +## Captured server reconciliation + +Read the archived runtime helper before evaluation; did not execute it. Independently checked all 11 request methods/paths, body Content-Lengths, response status sequence `[200,200,201,200,200,200,412,200,200,200,200]`, and JSON relationships rather than trusting `status: PASS` alone. + +- Readiness and served OpenAPI succeeded; response 2's JSON exactly equals `docs/openapi.json` from the actual pinned Studio ZIP. +- Import request/GET/source GET preserve complete entry/file content, original CRLF, Unicode em dash/comment and stable model UUID. Allocated resource `230d7234-8150-4252-a4bd-85a3b7b2308c` differs from model UUID `93200000-0000-4000-8000-000000000001`. +- Conditional PUT changes the single project-name string only, returns `changed: true`, increments revision 1→2 and returns a new strong ETag. Readback model equals the original complete model with only `name` replaced; exact edited source is preserved. +- Stale PUT carries the original ETag and distinctive `STALE MUST NOT PERSIST` source. Response is 412/revision_conflict. Subsequent accepted GET/source output equals the intended version, excluding that rejected write. +- Accepted GET equals fresh-process GET, archived accepted envelope and store JSON. Store bytes hash `c928995daf0c1ff3963850aba9ad533196423fa2112ad99554758825f73fc5af`; final ETag is consistent across edit, readback and reopened endpoints. Original/final revisions and ETags match the recorded result. +- Runtime result has 23 distinct named checks, two distinct owned server processes (4176 and 2188), exact packaged executable/runtime identities and loopback-only bind `127.0.0.1:61860`. Reviewed stdout/stderr contain normal startup text and no reported crash diagnostic. Version/help exit 0; version prints `arcadia-server 0.2.0`. + +Both server exit-1 results are explicitly imposed Windows TerminateProcess after acknowledged requests. This supports persistence of acknowledged writes across that restart, not graceful Ctrl+C shutdown, in-flight-write recovery or power-loss durability. The 23 checks are the recorded runtime helper's checks, not 23 HTTP calls or 23 new trials in this review. Authentication, remote networking/TLS, concurrency/load, Linux and broad API semantics remain outside this evidence. + +## Final disposition + +The delivery's corrected, bounded conclusions are supported by raw package/build/process/source/HTTP records and selected original screenshots. No acceptance claim needs expansion. Preserve the explicit second-Escape correction, source715036e/package hashes, forced-server-stop limitation, existing-user-profile qualification and readability limitation in final release evidence. This review supplies additional independent evidence for release-owner judgment; it does not close the full human checklist, issue 37 or the user application goal. diff --git a/harness/reviews/prerelease37-native-gui-20260924.md b/harness/reviews/prerelease37-native-gui-20260924.md new file mode 100644 index 00000000..25f3c097 --- /dev/null +++ b/harness/reviews/prerelease37-native-gui-20260924.md @@ -0,0 +1,31 @@ +# Native GUI evidence admission review — 2026-09-24 + +No new material admission finding. The published evidence supports the bounded Windows packaged workflows and saved-state claims, with the observational limits below. This is an independent read-only evidence review, not a new native execution or full application acceptance. + +## Identity and method + +Read README, independent-review and manifest through Gitea MCP at `8a82c0da1fd7130807878cc74598545ef38b5904:harness/prerelease37-windows-gates-20260924/native-gui`. Downloaded only its capped manifest/archive. Manifest SHA256 `acc5a1f1add9bc8be0711372224d8941ac14dfdc6018f7f6cc38b5bb798cc0b1`; archive 10,357,627 bytes, SHA256 `f18c21d262ca78bc9c2d8a339c3389fa31c50a81187cc53cc472f3253cda2815`. Independently checked all 351 unique exact members, lengths, SHA256 and CRC against manifest. Original stage manifest SHA256 `c90371bee1c2b2ae850b5d8e64ab9f1368f541ef51104e3b888a721e05d49c13` matches. + +Reproducible local data inspector: `inspect_native_gui.py`, output `native-gui-inspection.log`, adjacent to this report. Run using Python with assertions enabled (`python -B inspect_native_gui.py`). It reads ZIP/JSON/Git blobs, extracts only explicitly allowed evidence text/data/images into `native-gui-inspection`, and never imports or executes archived recipes or binaries. No candidate edits, Cargo, product launches or remote writes were performed. The first extraction refused JPEG files absent from the local inspector allowlist; adding JPEG as evidence data resolved that local inspector-only omission. No original evidence was changed. + +## Independently reconciled + +- Frozen source is `715036e132677dd2bc2fb37ed316d93dc9987aec`. Entire stage source inventory equals the archived final04 inventory already independently checked against all 3709 immutable Git input blobs in FINAL04-COMPLETE-REVIEW.md. The baseline fixture was additionally read from this exact local Git commit and byte-compared: B SHA256 `2dae785d63a940094d8052f3dcddcb988c024b38df9780e026bccd516264dbfe`. P is exactly B minus the single predefined Shutdown complete placement source line. +- Independently deep-copied complete expected B: removing only Activate's unique `(500,310)` waypoint yields complete C; removing only placement `ea202922-0033-4000-8000-000000000216` for element `...000000000037` yields complete P. All frozen typed-validation outputs equal their expectations. Expectations are bound in the staged immutable manifest before launches; no GUI edited result was used to construct an oracle. +- All ten complete actual Project JSON objects exactly equal expected B/C/B/C/P/P/P/P/C/P, using strict canonical JSON. Each receipt binds the exact stage manifest/source, accepted=true, no failure, unchanged original before/after inventory, actual output digest and exact archived source bytes. Each converter/snapshot receipt has successful exit, expected helper/runtime hash, matching argv, input/output digests, and empty stderr. All five P source inventories match checkpoint04; all three C inventories agree, including reopened08. Final working C/P bytes match those checkpoints. +- Rehashed actual local release ZIP `5554e32533e3cb2c33fb4143a5a0902733d5a8d47577303cb2842d268e7bc76d` and every one of its 565 member lengths/digests against stage package inventory. GUI identity remains individual-build `2c03bd6cc3f1110c219e4f57ab94874a121590f83779239a93289b7170664399`; runtime `aa282fa47bb2c4e210cea5e470cdbdca58568f0be0bc73e404d86ab9cd95ef6a`. Snapshot `6a1f438ab8f122251846f46cf1c5440ec7b0362b34fd61b97286d470bad11b74` and supplemental converter `c318ea38111688b6ce0d878b3e9e7dfebe0b3732d0c094bb4580fb39dbcc3d26` match already reviewed source/build provenance in PACKAGE-SERVER-REVIEW.md and FINAL04-COMPLETE-REVIEW.md. +- run01 PID6516 and run03 PID2044 bind the admitted extracted executable, exact argv, stage manifest and archived launcher hashes. Both close receipts match PID and exit0, package_unchanged=true, with empty raw stdout/stderr. Child environments are exactly equal: private LOCALAPPDATA and PATH limited to package plus System32; run03 changes only unrelated cwd among these launch settings. All 44 capture metadata records identify the admitted executable and include an actual full-size image member. + +## Original refusal and recipe boundary + +Read stage/check recipe, PLAN, both launcher versions, observations and cwd incident. The stage/check algorithm compares entire model objects and closure fingerprints; its success is not inferred from screenshots. The frozen recipe's historical “prepared/unexecuted” text is retained preparation provenance, not the current execution report. + +run02 is explicitly a pre-Popen refusal at the original launcher's empty-cwd assertion. `reviews/cwd-incident.json` retains the assertion/attempt record and original inventory, including literal `%SystemDrive%/ProgramData/Microsoft/Windows/Caches` with three file digests. It is not counted as a successful launched process. Independently versioned v2 launcher SHA256 `09ba3aad00643d9915453797656347ba1084a3b9a75187964587a733ea19fc4b` requires fresh run>=3 cwd, preserves the original incident and verifies its old inventory before launch and after exit. Raw run03 inventories equal the incident. Cache-writer attribution remains unproven. There is no evidence substitution or need to discard the original refusal. + +## Visual review and limits + +Viewed 18 full-size archived screenshots: 03,04,06,08,10,12,14,26,27,29,30,31,32,35,36,41,42,44. They corroborate the populated example and guide/return, authored square disappearing/restoring/disappearing, armed source/target status reverting to Select/move, Dispatch query and filter surviving popup dismissal, dialog cancellation restoring clean canvas, reopened cleared State and Show-unplaced offering Shutdown complete, and final all-changes-saved feedback. State labels remain tiny/congested at the shown 56% view; #36 remains open scope, not accepted as fixed. + +Images establish visible states, not the causal input between captures. The published operator observations supply the Escape/outside-click/Alt+F4 sequence and fresh-process Open actions; mechanical receipts independently establish saved contents and process identities. Exit0 alone does not establish normal-close cause, and full Project equality does not expose transient focus/undo internals. No broader accessibility/DPI/performance, held-drag/reconnect, physical network disconnection, Linux or full specification claim is warranted. Independent downloaded-package runtime QA and release-owner synthesis remain separate. + +For parent visual review, all full paths start `C:/Users/User/Documents/Git/ARCADIA-Studio/.local/prerelease37-integration-review-20260924/native-gui-inspection/native/captures/`. Selected exact filenames: `08-activate-selected-1.jpg`, `10-clear-bends-1.jpg`, `12-undo-1.jpg`, `14-redo-1.jpg`, `26-source-armed-1.jpg`, `27-source-escape-1.jpg`, `31-filter-open-suggested-1.jpg`, `32-filter-escape-dialog-retained-1.jpg`, `35-first-escape-all-retained-1.jpg`, `36-second-escape-cancelled-1.jpg`, `41-run3-state-cleared-bend-1.jpg`, `44-run3-saved-before-normal-close-1.jpg`. Index0 is often a 16x16 auxiliary surface; use index1 for those captures. diff --git a/harness/reviews/prerelease37-parent-visual-20260924.md b/harness/reviews/prerelease37-parent-visual-20260924.md new file mode 100644 index 00000000..d739a246 --- /dev/null +++ b/harness/reviews/prerelease37-parent-visual-20260924.md @@ -0,0 +1,35 @@ +# Parent visual cross-check of packaged trials + +2026-09-24. Parent Codex-956atf independently opened seven original full-window +JPEGs from the two published evidence archives. No application was launched or +controlled on this machine. These observations supplement the separate raw +source/model/process reconciliation; images alone do not prove persistence or +history integrity. + +Builder delivery `8a82c0da1fd7130807878cc74598545ef38b5904`, archive SHA256 +`f18c21d262ca78bc9c2d8a339c3389fa31c50a81187cc53cc472f3253cda2815`: + +- `native/captures/31-filter-open-suggested-1.jpg` shows the Add existing dialog, + Dispatch query and open Suggested filter. Capture32 closes the dropdown while + retaining the dialog, query, filter and empty-result guidance. +- `native/captures/36-second-escape-cancelled-1.jpg` shows the ordinary canvas, + no chooser, all changes saved and Cancelled/Select-move feedback. +- `native/captures/41-run3-state-cleared-bend-1.jpg` shows reopened State view03 + with Activate selected, endpoint circles and no authored bend square. The + displayed route labels are small and overlap nearby content at 56% Fit. + +Independent-machine delivery `684061ec0f486825886c077a038f06d17896392e`, archive +SHA256 `cb084be5f316f1b3be45bc5493fe0cbd6076052198c09f9dbac441b8f2588e63`: + +- `native-2/26-second-escape-cancels-form-1.jpg` actually shows the Edit transition + form still open, source dropdown closed. Its intended-outcome filename is not + accepted as evidence that the form closed. +- `native-2/28-explicit-cancel-1.jpg` shows the form closed, Activate selected and + its two square bend handles, including the lower added bend; all changes saved. +- `native-3/05-persisted-bend-selected-1.jpg` shows the same two-handle route in + reopened view03 and source-opened feedback. Exact coordinates and source + preservation require the separately reviewed five-file byte oracle. + +The captures support the bounded described visual states. They retain the known +State label crowding and do not establish general readability, accessibility, +every intermediate input event or completion of the human preview checklist.