ARCADIA-Studio/harness/scenario-selection-evidence-20260924/reviews/native-checks.md

Independent native evidence review

Reviewed the Scenario selection native package and updated top report on 2026-09-24. No material finding within this bounded evidence audit. This is an audit of the recorded branch-specific exercise, not implementation acceptance or a new native run.

Integrity and provenance

• Ran python -B harness/scenario-selection-evidence-20260924/native/verify.py: PASS for archive integrity, build/source linkage, exact source deltas and clean close. Inspected prepare.py, verify.py, manifest, README, retained launch/capture helpers, and start/finish records.
• Independently checked the ZIP inventory, member lengths and SHA256 values: 99 members, no duplicates. All 99 members equal their retained original bytes in .local/scenario-selection-native/run-1, its helper directory, or the closed selection-build-1 gate. The archive is 2,622,962 bytes, SHA256 391a91b55ab2c4c4d7663026d38a6485bfa5876d1205f40456da508978a69df3.
• Archived build result and source proof are byte-identical to their logical files in the separately verified checks package. They link to ed55b1781d86ec49af5d2588cbd06f7b878d9009. The retained copied executable equals the build artifact pin: 119,545,344 bytes, SHA256 7e0b6e714c89e34baef38bdaf12ecd35456cd39d62cb6b1dcf06e3487b7824bc. The executable is intentionally not included in the native archive.
• Independently matched the baseline fixture to the tracked fixture bytes in the proven build input capsule. Launch records identify the isolated copied executable; every archived capture metadata record targets that process path and has a timestamp between recorded launch and close. The closed process has exit 0, unchanged executable and fixture, and empty stdout/stderr.

Exact model/source checks

Independently diffed baseline against the spacing snapshot. Exactly one line changes: saved diagram incident_readable (A), occurrence updated_report, send_y 252 to 281.8841247558594. All other lines remain identical, including saved B and semantic ordering. Selection-only, post-Undo, and final working snapshots each equal the entire 6,206-byte baseline (SHA256 74cd2c1dd5eed7eb19895a752bb8ff9ce93655f7638a5312c96f56218568ebb3).

The corrected packager/verifier locates A before the incident_wide section and then compares the complete expected file. This avoids the incorrect earlier assumption that the original ordinate appears only once across both representations. The parent reports that earlier packaging attempt failed before output creation; the corrected artifact validates. This packaging correction establishes no application defect.

Manual visual inspection

Inspected original screenshots 26/27, 31/32, and 34. The 26-to-27 pair shows count two becoming zero, primary Properties returning to Scenario, and visible geometry translating by (-50,-30). The 31-to-32 pair shows count two becoming one, first occurrence primary, and unchanged geometry. Screenshot 34 shows the second occurrence selected and shifted downward, the first unchanged, and spacing-updated feedback. These endpoints agree with the recorded exercise and source comparisons.

Screenshots establish visible endpoints; the precise first-press input sequence is the executing agent's observed native interaction, additionally supported by the separately recorded mounted tests. The standalone verifier correctly does not claim to automate visual interpretation or execute archived helpers.

Claim boundaries

The native README and updated harness/scenario-selection-20260924.md keep observed native behavior separate from mounted-test coverage. Native Shift, Redo, dirty Properties admission, stale queued events, group navigation, and fallback C are explicitly unexercised. Representation navigation is limited to retained singleton primary. Early targeting and minimized-window capture anomalies are disclosed; missing capture 45 is explained, not represented as a successful capture. The fixture's five existing checks are disclosed.

The updated top report accurately carries the final scoped gates and historical failures, keeps changed-input attempt 3 outside accepted checks, and distinguishes this #29 contribution from frozen #37. It does not claim completed group movement/removal, full workspace verification, integrated acceptance, or published-release acceptance. No Cargo, Git mutation, network access, or native input was performed in this review.