ARCADIA-Studio/harness/release-canvas-native-d61f1f3-independent-review-20260912.md

# Independent review of closed d61 native paint evidence

Reviewed `harness/release-canvas-native-d61f1f3-20260912.md` and the original archive at `harness/experiments/canvas-native-d61f1f3-20260912/`. Scope: the closed Windows trial of `d61f1f38779ab70a77efcde523ae31c432aefcfd`, activation-border continuity, selected identities, minimum-window evidence and claim boundaries. No app input, compiler/test execution, Gitea mutation or tracked edits. This report is the only new review artifact.

**Bounded conclusion: no actionable finding against the reported paint correction or its stated scope.** The archived original images independently show the former erased border bands repaired, and the Inspector states agree with the claimed call/reply selections. This is evidence for this particular correction and fixture, not the later lifecycle union, palette acceptance or complete application acceptance.

## Identity and preservation evidence

Read manifest, launch, closure, stage manifest, both checkpoint records/logs, fixture identities and launcher/checkpoint procedures. Independently recomputed **all 43 payload sizes and SHA256 hashes**, with no discrepancy. Independently compared complete bytes of all four baseline/preflight/final-Save files; all three copies match for each file.

Launch identifies source d61, PID 10792 and executable SHA256 `c990fa05772c60666ecf96ab808e6140e57542ad448ddd092b977be0b587136f`. The staged manifest agrees and records executable size 113,939,456. Closure refers to the same PID, exit 0, 23:51:20 UTC, after launch at 23:46:20. The launch procedure checks staged artifacts and same-source full/Clippy/GUI results before launching under a private profile. All three archived stage results identify d61, exit 0, and no changed inputs. I did not rerun those gates or the native executable.

The final checkpoint at 23:51:02 precedes closure. Its identified oracle returns `PASS same: complete Project and exact SourceBundle restored`; retained warnings are two explicitly reported `unallocated_leaf` completeness warnings, not errors. The checkpoint's oracle identity is correctly separated from the native executable's source. The final `bound-saved-0.png` says “All source files saved.” These results support the documented preservation outcome. The root's recorded input sequence supplies the Ctrl+S/click actions; static screenshots independently establish visible result states, not the entire input history.

## Visual findings checked against original images

| Witness | Independent visible result |
|---|---|
| `oa-outer-former-label-shoulder-0.png`, `oa-outer-r2-label-shoulder-0.png`, `oa-outer-r1-label-shoulder-0.png` | OA nested scenario; C1 selected in tree and Inspector, name C1 420 kPa, alias `bound_1040`. The outer left border is continuous through the C2, R2 and R1 label bands; the inner execution still occupies its distinct offset interval. |
| `oa-c2-exact-receive-endpoint-0.png` | C2 380 kPa, alias `bound_1030`; inner execution and C2 line highlighted. |
| `oa-r2-exact-send-endpoint-0.png`, `oa-r1-exact-send-endpoint-0.png` | R2 378 kPa `bound_1060` and R1 418 kPa `bound_1050`, respectively; selected dashed reply visible. These endpoints did not result in the outer C1 Inspector state. |
| `oa-reentrant-outer-label-shoulder-0.png`, `oa-reentrant-inner-0.png` | OA reentrant outer C1 `bound_3040` and left inner C2 `bound_3030`; reverse paths and both respective execution outlines remain visible. |
| `sa-outer-former-label-shoulder-0.png`, `sa-r2-exact-send-endpoint-0.png` | SA nested outer C1 `bound_2040` versus R2 `bound_2060`; repaired outer border and separate inner execution are visible. |
| `sa-minimum-100-0.png` | Actual image size 1102×732, scale display 100%; right participant header partly clipped. The release note acknowledges this rather than claiming all content fits. |
| `sa-minimum-fit-0.png`, `sa-minimum-fit-outer-shoulder-0.png`, `bound-saved-0.png` | Same outer dimensions, toolbar displays 66%. Fit restores the header within the canvas. Shoulder-selected and saved states show C1 `bound_2040`, with a continuous outer border. Small text/unselected outlines remain faint. |

Arrowheads and text remain visible in these witnesses. Cursor/blue pointer halo covers a small local area in individual screenshots, so multiple selected-outer witnesses were used to inspect the same border bands away from the pointer. I did not infer concealed pixels from a single cursor-covered endpoint.

## Discriminating read-only pixel cross-check

After visual inspection, read pixels from original PNGs with Pillow without modifying or generating screenshots. In the two d61 OA selected-outer images `oa-outer-former-label-shoulder-0.png` and `oa-outer-r1-label-shoulder-0.png`, checked each row in y=432…452, 524…544, and 616…636 at x=661…664. A row counts as having the teal border when any pixel satisfies `R < 85`, `70 < G < 180`, `G > R + 35`, and `B > R + 25`.

Both d61 images have **zero missing rows in all three 21-row bands**. For a negative control, inspected the original 83 image `harness/experiments/canvas-native-83f53c7-20260912/evidence/oa-outer-activation-selected-0.png`: same 1442×932 dimensions, same nested fixture and `bound_1040` selection. That image visibly has gaps and the same read-only scan reports **21 missing rows in each band**. The discriminator therefore detects the original defect rather than accepting a generic selected border elsewhere. The scan is narrowly coordinate/fixture/palette-specific, not a general renderer acceptance test or a monitor contrast measurement.

## Scope assessment

The release note is appropriately bounded: it does not convert the same-source automated gates into native evidence for unexercised workflows. It explicitly excludes a new self-call/ALT/PAR/lifecycle-marker result, architecture/state Save/Undo repetition, palette correction, lowest-zoom readability, complete OA/SA release acceptance and Linux. Its minimum-window clipping/66% faintness statements agree with the images.

The evidence supports retaining the stated d61 native paint acceptance. f659 introduces additional lifecycle markers, lane rendering and header clipping and still needs its own combined gates/native witnesses. Issue #26 remains open on the demonstrated faint unselected geometry and small text; this review did not test a new palette.