85 lines
11 KiB
Markdown
85 lines
11 KiB
Markdown
# Independent final consolidated E/F packet review
|
|
|
|
Review checkpoint: source HEAD `d0896ced75fea73bcda6bd14f259a34503840704`; documentation and raw evidence staged afterward for archival. This is a new packet review after the earlier source, repair, and pre-reopen evidence reviews. It does not relabel those historical reviews as having examined later data.
|
|
|
|
Read-only scope: reviewed README, validation, staged STATE/DECISIONS changes, inventory/source manifest, portable verifier, copied launch/exit records, raw failures, locked readings, native observation notes, and the later independent saved-F checker/results. Ran only the portable offline verifier without optional flags and read/hash checks. No Rust/helper/document edits, staging, commit, GUI, Cargo, or Gitea action. Only this report was written.
|
|
|
|
## Recommendation
|
|
|
|
**Accept the packet for archival as a bounded, non-promoted experiment. No material evidence or final-claim blocker found.** Neither E nor F establishes dependable unselected label ownership. The source repair and conservation evidence are useful, but they do not turn the experiment into a production readability improvement. Keep the wider issue open.
|
|
|
|
The README, validation, STATE, and DECISIONS consistently express that boundary. E's correct lack-of-direction observation is correctly treated as appropriate for an unoriented boundary exchange; it is not falsely graded as an error. F's wrong exchange readings are preserved. Both readers' plausible wrong conversation ownership remains the relevant negative observation. Separate single-agent screenshot readers are not presented as a controlled human or causal comparison.
|
|
|
|
## Inventory and portability
|
|
|
|
I independently ran:
|
|
|
|
```text
|
|
C:/Users/User/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/python.exe harness/experiments/wrapped-label-20261001/evidence/verify.py
|
|
```
|
|
|
|
It exited 0 with PASS: 83 inventoried files, exact pristine E/F SVG conservation, complete F Project delta, and reopened-F logical geometry. The inventory totals **3,476,900 bytes**, excluding the inventory file itself. The source manifest declares 300 source files at d089. The verifier requires the actual evidence-file set to equal the inventory plus itself, so omitted/extra local packet files are not silently ignored. Scoped `-text` attributes preserve raw bytes through Git.
|
|
|
|
The verifier does not execute historical preparation/launch scripts. It loads checker definitions without invoking their main functions and supplies packet-local inputs; no native process, renderer, or bytecode cache is launched/written. Historical absolute paths remain preserved as provenance. The README properly directs portable replay through the wrapper rather than promising those historical launchers work from a new checkout.
|
|
|
|
The initial staged-byte verification failure, supplied by the parent, was due to ignored log files being absent from the index. Parent reported force-adding only the scoped evidence directory and rerunning the unchanged verifier. That is a packaging repair, not grounds to change the inventory or relax byte equality. This review's independent default run checks filesystem packet bytes; the parent's final `--index --source` checkpoint is recorded separately below.
|
|
|
|
## New reopen evidence, timing, and claims
|
|
|
|
The earlier final evidence review said same-profile reopening was then absent. The consolidated README and validation explicitly state that it predates the fresh reopen runs and link the later saved-F review. That correctly preserves the historical checkpoint without leaving its limitation to be mistaken for a final verdict on newer artifacts.
|
|
|
|
The later independent checker requires the exact original and saved Project hashes, the exact single authored target route, and complete parsed Project equality after removing it. It pins the original target path before deriving the authored label anchor/box/line positions through f32 arithmetic and earlier Iced font metrics. It compares every edge/path record and all SVG bytes outside only the moved target label. The five corruption controls and result are present. It uses explicit `require` checks rather than Python assertions for its logical oracle. This is materially stronger evidence of saved-F export geometry than simply comparing a screenshot with a remembered position.
|
|
|
|
Copied saved inputs under both reopen profiles are exact `0ad4015d...` files, and reopened-F SVG is `5b1106b9...`; exact hashes are below. The independent default verifier reran the logical saved-F check successfully from the packet. This establishes model-delta and logical-export preservation, not glyph/raster parity.
|
|
|
|
The ordinary Baseline process actually opened the saved file from the preceding F folder through Recent Open, rather than its unused identical local copy. The parent observation and validation call this out. Its watcher alone checks its own pinned copy; the explicitly recorded actual-loaded-path observation and post-close hash are therefore necessary provenance. The packet includes the actual F-folder input and its independently matching hash. No Baseline SVG is claimed. The one-line Baseline rendering remains a parent-observed native result; the independent saved-F checker does not claim to establish it.
|
|
|
|
The parent observation distinguishes tool-observed events, captures, and the unretained raw event transcript. F's last-line click is specifically marked tool-observed only. E's two line selections have captures; E editing is not claimed. F JSON export remains dirty until Undo, so it is not described as Save All or an ordinary source-file save/clean round trip. Failed chooser attempts and automation ambiguity are retained rather than diagnosed as application bugs without evidence.
|
|
|
|
## Gate scope and retained failures
|
|
|
|
`commands.json` explicitly records the parent-observed filtered command:
|
|
|
|
```text
|
|
cargo test --locked -p oa-desktop -- diagram:: diagram_usage:: fit_ --test-threads=1 --nocapture
|
|
```
|
|
|
|
The raw result and validation say 126 passed, 0 failed, 0 ignored, 1,426 filtered out. They do not claim a full desktop/workspace suite. Clippy is identified as desktop all-targets with `-D warnings`; the build is the development GUI binary. The command manifest acknowledges that logs alone do not prove invocation scope. An empty formatting log is paired with a fresh parent-recorded successful command, and the earlier missing-log summary error is disclosed.
|
|
|
|
The E0063 fixture compile failure and the later five-pass/one-fail stale-wrapped-geometry regression are preserved. The source/repair reviews and repaired raw/presentation split are not rewritten to imply first-pass success. The SVG oracle's initial selfcheck setup error and revised successful controls are also retained.
|
|
|
|
No new Rust correctness issue was identified by this packet review; the prior repair review still covers d089. Native different-edge guard refusal, held-pointer cancellation, E edit lifecycle, glyph/raster parity, generalized metrics/placement/persistence/performance, nesting, accessibility/DPI, Linux and broader modeling acceptance remain excluded. Those exclusions fit archival after a negative reading result; they are not undisclosed unfinished acceptance gates for a promoted feature.
|
|
|
|
## Remaining nonblocking caution
|
|
|
|
The wrapper and original F model helper use Python assertions, so the documented normal Python invocation matters: `python -O` would disable some checks. The SVG and later saved-F logical check functions retain explicit exception guards, but that does not rescue disabled inventory/index/source assertions. The packet documents this limitation for the original model helper; do not describe the wrapper as hardened against optimized-Python misuse. No such invocation was used in the reviewed successful run.
|
|
|
|
Source-manifest checks bind recorded files to Git objects; process receipts bind expected hashes/profiles to parent launches. They are not binary reproducibility attestations or a native interaction replay. Current documentation avoids those overclaims.
|
|
|
|
## Fingerprints at this review checkpoint
|
|
|
|
| Artifact | SHA-256 |
|
|
| --- | --- |
|
|
| README.md | 79f621b58fe0b9a5b16d3aa9d5b1c89994b79f4f314d4d97ff9419e7603f7a1a |
|
|
| validation.md | ad608b10ea9634fde22823cd841a5e7bdee5aa5b67faf21a2975f3726e505e8f |
|
|
| evidence/evidence-files.json | 57f137d47d73f0b28ed248c974815d10454339c35d73eb86a550b34d3061fd39 |
|
|
| evidence/verify.py | 282f8d95355d764d09ceea5662be2aa381b17ed863b5fd442d37599817616c23 |
|
|
| evidence/reviews/wrapped-native-observation.md | 11a295fd1f60dad6c57b674f810a16d2e91b40bf2f84592cadc51ae5f4c3d114 |
|
|
| evidence/scripts/wrapped-reopen-independent-check.py | 8d718db9a52e894e8160436eab2d57998f211b71f4485b3543c6badde36b7815 |
|
|
| evidence/native-E/first-reading.md | 9de823d0d43e7753f2799e203a68332b7361c1c6c2f8f18baece5c60baa01c66 |
|
|
| evidence/native-F/first-reading.md | 2867ba22c3195e7b77cf696fffbdcc65ca6e8a74dc212f38407ef797216b5f8e |
|
|
| evidence/reopen-F/working/moved-F.json and reopen-Baseline copy | 0ad4015d06cf513cd8f9cfe3a155c07590ba09585002277d6eed7a78dd3b6a57 |
|
|
| evidence/reopen-F/working/reopened-F.svg | 5b1106b9f8cdb6030e29d657f8aac1920e1a554e868bab4fce3e250a42469cd5 |
|
|
| evidence/reviews/wrapped-reopen-independent-result.json | 8804ba0f50500ae92c1dd2c175ca241aa23070db6a0221793d6a85d67cd27e0f |
|
|
| evidence/reviews/wrapped-reopen-independent-selfcheck.json | 7e22a0379be1ecf095130ab179525525b850f50ee905e804b607db3bc92e023b |
|
|
|
|
The copied locked-reading hashes exactly match their pre-packet originals reviewed earlier. The packet inventory supplies remaining file hashes; the verifier confirms them. This report itself will be archived after that inventory checkpoint, so adding it to the inventoried directory requires inventory regeneration and another staged-byte check rather than pretending the 83-file result covered the new report.
|
|
|
|
## Final staged/source checkpoint received before completion
|
|
|
|
Parent confirmed the unchanged `verify.py --index --source` run completed at d089 with exit 0/PASS: inventory 83; E and F each retain the exact 6,977-byte remainder SHA256 `5603ada145dbb255eb4dfd583258427057a9312443685d47f278ef76756ef8ed`; complete F model delta and reopened-F logical geometry true; `index_checked: true`; `source_manifest_checked: true`. All 300 pinned source-blob checks succeeded. This optional-flags result is parent-run evidence, distinguished from my independent default run above.
|
|
|
|
The first index failure remains in chat history rather than as an inventoried file: ignored log files had not been staged. Parent force-added only the scoped packet evidence directory, reran the full checker unchanged, and obtained the result above. No Rust/source change was made for that repair.
|
|
|
|
Parent will archive this review as `final-review.md` at the experiment root, outside `evidence/`. That preserves the 83-file raw-evidence inventory and avoids a self-referential review/inventory cycle. This review's acceptance is for archival of that verified packet and explicitly non-promoted experiment; it does not claim the later review file was part of the earlier inventory result.
|