ARCADIA-Studio/harness/process-group-removal-evidence-20260925/native-review.md

Independent bounded native evidence review: Process removal

Reviewed 2026-09-25 UTC. Source identified by the supplied build/native receipts: 572e96ba03522cfb822b13fe4abb50e857a5de48. This reviewer made no application inputs, source edits, Cargo invocations, Git mutations or Gitea writes. Parent owns execution and publication. Evidence root: .local/process-removal-native/run-1; local-only until published by parent.

Conclusion

No new blocking correctness or readability finding in the inspected OA three-member removal workflow. The evidence supports the complete-group review and removal, visible interpretation/caller consequences, cancellation, expired-review feedback, exact saved-source removal/Undo/Redo and fresh-process reopen. Native compact-width acceptance remains open: the 1605x872 maximized images and restored 1442x790 image do not establish a 1100px window. Accessibility and keyboard-only navigation are also not established by this evidence. The closed run-2 reopen addendum below completes the persistence review without extending these layout/input boundaries.

Evidence binding and its limits

• Inspected prepared.json, started.json, finished.json, build evidence/result.json, evidence/source-proof.json, capture.js, verify-captured.py, fixture files and every saved-source snapshot receipt/evaluator output. Independently recomputed the lengths/hashes of all 12 immutable entries in prepared.json; all match. This includes the executable, runtime DLL, oracle, helper, result and captured input-map/source-proof records.
• The build receipt identifies exit 0, unchanged captured inputs, 832 captured inputs, and desktop artifact SHA-256 a13190e41bdb2d05ee35f1e6e38d35ce9f7a8a46864d5a73522126c0529489ca. The native executable matches this artifact. The source-proof receipt reports exact captured inventory/Git blob agreement at the identified source. I checked the receipt bindings; I did not independently repeat the Git-blob comparator.
• Native run-1 starts PID 6524 from its isolated copied binary at 03:35:47 UTC; closes at 03:45:29 UTC with exit 0, unchanged immutable inputs and empty stdout/stderr. These are crash/log checks for this run, not assertions about all application behavior.
• Capture JSON records window identity, time, accessibility snapshot and image filenames/dimensions. The capture helper does not record each input. Consequently screenshots prove visible states; descriptions of clicks/keys and the single Undo/Redo count are parent-reported execution observations, corroborated by before/after states and source snapshots. Labels alone are not instrumented input evidence.

Directly inspected images and workflow findings

Images inspected with view_image: 17, 18, 19, 21, 22, 23, 24, 26, 27, 28, 30, 32, 38, 39, 40, 42 and 44 (the full-size -1.jpg variant when present, otherwise -0.jpg).

1. Captures 17/18/19 expose all ten effects across two pages and page-one scrolling: three selected removals (same-name Action/Control distinguished by @a/@c, plus @approved), one preserved/updated owner, explicit-before and implicit-after order, the two surviving ordered positions @e1/@e2, and caller definitions @caller/@outer. Full names and aliases are readable. The wording separates three removed records from one updated record and two affected caller definitions. It warns that reused topology, entry/exit connections and diagnostics may change and disclaims complete caller graph computation. This correctly avoids presenting the two caller definitions as a reference-instance count: @caller contains two references in the fixture. The transitive @outer row says it uses @assessment; the policy intentionally includes transitive use.
2. Page-one content scrolls while search, pagination, Cancel and destructive action stay available. Capture 18 reveals the otherwise lower after-order result; capture 19 exposes both ordered positions and both caller rows. No hidden part of this fixture's reported ten-effect set remains uninspected. This does not establish reachability for arbitrarily long names or much larger effect sets.
3. Capture 21 finds exactly @outer with full name and alias. Capture 22 retains that review and search after parent-reported Enter. Capture 23 shows zero matches and helpful recovery wording. The selection count and destructive action remain independent of the filter, as expected for reviewing the captured group.
4. Capture 24 after parent-reported Cancel and 27 after Escape retain three selected members, both same-name cards and the maintained arrow at the same visible positions; status says model/selection unchanged. A later saved snapshot is byte-identical to baseline. Capture 26 shows the whole group review after focused Delete; 30 shows it after focused Shift+Delete. Capture 28 explicitly does NOT show a review, despite its filename: parent reports that Shift+Delete immediately after Escape was ineffective until F6 restored canvas focus (29), then Shift+Delete opened 30. This supports the focused shortcut contract; do not claim automatic focus restoration after cancellation.
5. Capture 32 shows the restored window with the stale review warning, disabled removal button and available Cancel. Text and controls are readable. Its lower edge is clipped by the available capture/screen, so it cannot count as a fully visible compact-layout pass. Parent reports missing capture 31 followed a UIA restore action that yielded a minimized-window capture error; 32 was recovered through fresh enumeration/reactivation. Missing 34 followed a tool refusal to drag outside screen bounds; parent reports 35 unchanged. Neither absence is a native crash finding or evidence of successful resize.
6. Capture 38 shows the menu over unselected Receive report while the original group remains three selected; its action is explicitly Remove selected contributions / links. Capture 39 then reviews @a/@c/@approved, not Receive report. This is a meaningful non-primary/context-subject discriminator.
7. Capture 40 shows zero selection and owner inspection after confirmation. Only Receive report and Forward report remain, joined by the expected implicit arrow, and status reports three selected contributions/links removed in one change with referenced definitions preserved. Capture 42 restores the two removed cards and maintained arrow and reports model change undone. Capture 44 returns to the two exchange cards and reports model change redone. The status/canvas observations agree with the byte-exact snapshots below.

Exact source and history evidence

Independently recomputed all snapshot SHA-256 hashes in addition to inspecting the four evaluator results (each PASS, exit 0). Full fixture source inspection confirms that the v2 expected file deletes @a, @c and @approved, changes assessment steps to [@e1,@e2], and retains all work/exchanges, both repeated caller references, transitive caller and opposite-layer sentinel records, with preserved writer trivia.

Saved state	Bytes	SHA-256	Expected
cancelled-baseline	2609	9965f4024584d64ac81442724aeb77b018b0cfae94e50bd412c5dcbbac49f6bf	Complete baseline
removed	2258	b4a2b07ad4c796f555c61945ca42b1c221b454b35488c60fefc86f96f7aa71ba	Complete v2 removal oracle
undo	2609	9965f4024584d64ac81442724aeb77b018b0cfae94e50bd412c5dcbbac49f6bf	Complete baseline
redo / final working file	2258	b4a2b07ad4c796f555c61945ca42b1c221b454b35488c60fefc86f96f7aa71ba	Complete v2 removal oracle

The saved-source comparisons establish exact full-file persistence at these checkpoints. Together with parent-reported single Undo/Redo actions and screenshots, they support the one-change history workflow. The source comparator itself does not introspect parser state, transaction count or history stack. This native case does not independently exercise SA removal, qualified-selection refusal, every ContextualLink purpose, a write failure, or every semantic discriminator in the broader automated suite.

Residual acceptance gaps and small observations

• Native 1100px compact layout and complete modal visibility at that width remain untested; do not close that gate based on restored 1442px capture.
• Parent created membership with context-menu Add, so native Shift+click selection is not established by this run. The failed early selection attempts 07/08 should not be counted as successful; parent reports selection took after maximize at 10.
• Accessibility snapshots expose only the window/system chrome, not application controls; no screen-reader semantics or tab-order acceptance follows. Search caret and focused shortcuts provide only the specific focus evidence above.
• No keyboard-only review completion, native failure recovery, or native SA group-removal evidence is included. Existing automated/source-review results must retain their distinct scopes.
• The visible plural agreement strings (for example "1 records updated", "1 definitions", "1 adjacent ordered pairs") are minor copy defects, not a correctness or reachability blocker. No source change is proposed as part of this bounded review.
• Packaged launch, integration and publication require their own closed records. Run-2 is a fresh native process using an isolated copied executable/runtime; it does not establish execution from the eventual distribution archive.

Closed run-2 reopen addendum

After parent reported closure, inspected run-2 prepared/started/finished records, snapshot receipt/evaluator result, and actual full-size captures 09-reopened-implicit-order-1.jpg and 10-reopened-saved-1.jpg. Independently recomputed every prepared immutable file length/hash: all 12 match. Both final working file and independent reopened snapshot hash to the exact 2258-byte v2 removed oracle (b4a2b07ad4c796f555c61945ca42b1c221b454b35488c60fefc86f96f7aa71ba). The snapshot evaluator passed with exit 0.

The prepared record names run-1/snapshots/redo.arcadia as initial source with its matching removed-source pin. Run-2 launched a new process, PID 7248, at 03:45:38 UTC from its separately copied binary and isolated cwd. Capture 09 displays 01 Assessment with zero selection, only Receive report and Forward report joined by the implicit arrow, owner inspection, and the source-opened status. Capture 10 displays all changes saved / all source files saved. The full file remains byte-identical after this reopen/save. The process closed at 03:47:52 UTC, exit 0, empty stdout/stderr and immutableUnchanged true; parent reports normal Alt+F4. These observations support fresh-process load/render/save preservation for the actual removed native result. No parser/history-stack introspection or broad model completeness claim is inferred from them.