ARCADIA-Studio/harness/experiments/state-label-layout-20260924/native-review.md

#36 native visual review: 70bdcdf8

Final scope update: the addendum below completes the fresh-process reopen and process-exit review. The original snapshot's pending statements are retained as history; the unmaximized-input limitation remains open.

Independent read-only review of original JPG/JSON captures01–15 under .local/state-label-trial-70bdcdf8-20260924/raw, the frozen B/W/U inputs, completed save checkpoints and harness/state-label-layout-contract.md. All15 full application images were inspected at original resolution; the16×16 secondary captures are not useful application evidence. No GUI input, product/helper execution, build, source/tooling edit or Gitea operation was performed. Fresh-process reopen and final process closure are outside this review snapshot.

The recorded run01 launcher identifies source 70bdcdf8fec1728e971cd90ad7f79a1df99777ff, PID5328, preserved GUI SHA256 0401341388025544185455ad5520aa2c1b79ca1387d1a57c2bb0bddb8b23754a. All15 capture metadata records identify that executable path. Independently reread manifest SHA256 is 7f1b6cfa8f6ca0c36b0b867f8a2ccf555cea292e8df878af629dac5b63d36ee6. This is evidence attribution, not a newly rerun executable/build check.

Observed result

No clear title/summary self-overlap, wrapping or ink clipping in the required B/W/U labels was visible in the maximized captures at displayed68% and82%. The text is small at those scales; these screenshots support the concrete cases, not universal glyph containment, perfect legibility at every scale, or pixel-exact font metrics.

• B, captures03–07: Ready and for review occupy separate rows, followed by the External · Completion [request_ready] summary. Captures04 and06 show State transition properties with stable alias between; capture05 shows State action properties with alias action_a and the exposed entry / Record request row selected. Thus the recorded sequence demonstrates distinct observable targets for the covered transition label and exposed row. Capture07 shows All source files saved. The broader authored diagram remains crowded and overlaps node territory; this is not evidence of global collision avoidance.
• W, captures08–11: the input is exactly40 W characters, verified from the pinned expected Project. The title occupies one visibly wider line at both scales, with the summary on the next line. Its rightmost ink appears within the visible white label extent; it is not shortened to the old scalar-width box. Capture09 shows between selected after the parent's described far-right/padding interaction. A static after-capture confirms the selected object, not the exact preceding pointer trajectory. Capture11 shows saved status.
• U, captures12–15: the input is exactly Ready\r\n漢字 é 👩‍🔧\n (combining acute, ZWJ sequence and terminal LF verified from the expected Project). The canvas shows two inked title rows and an extra blank cell before the summary, consistent with three title cells plus one summary cell. The second row's marks appear contained; capture13 shows between after the described blank-cell interaction, and15 shows between in the summary-selection sequence. At this screenshot size I do not claim independently verified shaping of every emoji/code point, nor infer preservation of CRLF from pixels.

Preserved ambiguous observation: captures01/02 are unmaximized1442×842 at displayed79%. Capture02 clearly has state_a selected after the parent's attempted click at(660,534). The coordinate/geometry cause has not been resolved by this review. This attempt must remain recorded as unresolved, not silently reclassified as a successful label hit or generalized into a demonstrated product defect. Successful maximized observations do not establish an all-window-size hit pass. Displayed percentages are rounded; no exact world-space movement oracle is inferred from them.

Save evidence checked separately

At inspection, checkpoints 01-B-saved, 02-W-saved, and 03-U-saved each have accepted receipts. I independently hashed the checkpoint actual.json and pinned expected file: each full Project file is byte-identical to its expected B/W/U file and the manifest/receipt hashes match. Each checkpoint has nine source files, matching the nine baseline members; every saved file's bytes match both its corresponding immutable baseline and its receipt SHA256. Consequently these pure saves preserve complete Project values and authored source bytes, including title separators and authored label geometry. This conclusion uses saved artifacts, not the screenshots' saved-status text. No claim is made for the unused M movement oracle.

Limits and retained history

Parent reports core68 passes, desktop1540 passes/seven existing ignored, formatting and strict Clippy on the candidate; this visual review does not independently rerun or audit all gate logs. Preserve the earlier desktop gate03 E0382 failure: the initial Shift/mixed regression moved row into its selection and reused it. Candidate70bdcdf8 fixes that test-only ownership issue with row.clone(); the failed attempt is not a runtime failure and is not erased by the later pass.

The bounded maximized visual observations and three exact pure-save checkpoints support this label repair. Fresh-process reopen, final close/log evidence and any generic-canvas native claims require their own remaining evidence. Group(+40,+40), preview/settled equality and one-step history are assigned to the fixed-oracle production-input regression; no native group displacement is inferred post hoc. The unmaximized79% target ambiguity remains an explicit limitation unless independently resolved.

Final addendum: fresh process, reopen and closure

Inspected all seven original full-size captures18–24 without repeating the earlier15-image review. Their metadata consistently identifies the pinned GUI path, window263106 and maximized1605×872 dimensions. Run02 launch records identify a fresh PID6100 with the same70bdcdf8 source and GUI hash as run01 PID5328.

Reopened B at displayed68% retains its separate two-line title and following summary in18;19 shows State transition properties with alias between after the parent's recorded summary click(737,518), and20 shows saved status. Reopened U in21/22 retains its two inked title rows and blank-cell separation before summary. Reopened W in23/24 retains the single wide title line and separate summary. No new visible required-case clipping or title/summary self-overlap was found. These are reopen observations in the second process; W/U are not additional fresh-process label-hit trials.

Independently reread and hashed checkpoints 04-B-reopened, 05-U-reopened, 06-W-reopened: all accepted; each actual.json is byte-identical to its corresponding manifest-pinned expected full Project and receipt hash. Each saved source directory again contains nine files, equal in count to baseline and receipt; every file matches its baseline and receipt SHA256. Across the six reviewed save/reopen checkpoints, all complete Projects and source bytes remain exact. No semantic, authored route or title normalization change is inferred or permitted by these comparisons.

Both close receipts match their launch PID and source pin, report observed exit0, null failure, and immutable_inputs_unchanged: true. I independently checked all four stdout/stderr files are zero bytes; their recorded hashes are the empty-file SHA256. These receipts establish successful observed process exit and the launcher's unchanged-input check. The specific normal Alt+F4 close action is parent-observed evidence, not something this read-only reviewer performed or derives from exit0 alone.

The parent additionally reports that restored run02 Home Recent attempts at(1060,744) and the offset-hypothesis point(1052,713) did not open B; after maximizing, the actual Open point(1147,736) did. This broadens the unresolved observation beyond a State-label target, but does not prove a coordinate-mapping cause. Preserve all attempts. Native hit claims remain limited to the successful maximized trials; the initial79% attempt is still not passed or retrospectively diagnosed.

Final bounded verdict: no new actionable defect found in the reviewed maximized B/W/U presentation, six exact pure-save/reopen checkpoints, or recorded exits/logs. Fresh-process persistence evidence is now complete for these variants. Keep the restored-window input uncertainty, rounded-scale limitation, synthetic movement/history test scope, and absence of universal Unicode/all-window hit guarantees in the published evidence. This addendum does not certify a broader archive/collector or generic-canvas native trial.