ARCADIA-Studio/harness/experiments/text-workspace-followup-review-20260913/native2-evidence-review.md

Native Text trial 2 — independent evidence review

Bounded evidence supports the seven selection-replacement probes, saved rename transactions, and the inspected compact Text interactions. This is not final #13 acceptance: capture 051 demonstrates the inherited simultaneous Find/drawer Escape defect; the next source correction is outside this trial. Review performed read-only on 2026-09-13; no GUI, compiler, application, or oracle was executed by this reviewer.

Identity and original records

Original trial: .local/text-workspace-native-20260912/trial-text-2. PID 5284 launched 2026-09-13T03:42:39.425514Z and closed normally at 04:18:28.799966Z, exit 0. Both logs are empty; closure records unchanged binary and DLL. Development executable SHA256 4a1d09c108fd5062f01bc7059064fa7dc839e03d0cc2a3b1e12f8b02c8115a24, 114,171,904 bytes. This was an isolated development launch, not packaged-release acceptance. Elapsed session time is not a performance measurement.

The adjacent native2-evidence-review.json binds all 105 original metadata states and all 112 original JPEGs individually, plus seven saved-disk receipts and the launch/closure. Every metadata-referenced JPEG exists; the JPEG set is exact with no duplicate references. Receipt SHA256: a2b0a6b3bec5bd88497f8de89baf3a652b5210b7c6a3480dd600bc8faa438722.

Observations

Directly inspected JPEG states 007, 010, 014, 016, 018, 022, 024, 025, 028, 033, 035, 037, 039, 043, 047, 050, 051, 064, 068, 072, 086, 089, 096, 099, and 104. Other captures are byte-bound, not claimed as individually visually inspected.

• Seven result images show the intended whole replacement: Find word/Close oa_contact_probe (010), word/Escape oa_control_probe (014), line/Close description "line_close_probe"; (018), line/Escape description "line_escape_probe"; (022), Back word oa_control_back_probe (028), file roundtrip word oa_control_file_probe (033), and file roundtrip line description "line_file_probe"; (037). Selected-origin examples 007/016/024/035 corroborate actual word/line selections rather than merely caret movement. These demonstrate selected-byte replacement for these seven probes; they do not establish preservation of Word/Line drag-extension mode after real navigation or every Unicode selection shape.
• At the recorded 1102x732 outer window, 025 shows approximately 306 pixels of editor height with the default drawer and three complete reference result rows (one declaration, two uses), with the next use near the lower edge. Do not describe this as three complete lexical-use rows. Enlarged drawer 039 leaves approximately 246 pixels of editor. Scrolled references in 043 and explicit hide/reopen in 047 show the retained lower group position and readable full oa/participants.arcadia path. Stale is visible before the alias. Retained stale rows are visibly pale; this review makes no contrast-compliance claim.
• 050 shows the first Escape has removed Actions while Find and the drawer remain. Despite its filename, 051 shows both Find and the drawer closed after the second Escape. The filename is an inaccurate historical label and must not be treated as a passing assertion.
• Representative saved states 064/068/072 show oa_restore_pressure / original oa_restore_work / oa_restore_pressure respectively, clean save messages, and consistent visible declaration UUID. Reopen capture 089 independently shows oa_restore_pressure in graphical properties. Small capture 086 shows assess_damage while the comment and description still contain the original prose @assess.
• Scope-stress 099 shows Partial and Collision before the long alias, two declarations/two uses, full Unicode 救援/assess.arcadia path, and combined repair conditions in the status strip. 096 shows diagnostic navigation selecting the missing include, with the full failure explanation wrapped in Problems. 104 shows the prior representative graphical model retained and properties blocked by pending source drafts. Invalid source loaded as repair drafts is not a successful valid-project load or model replacement.

Saved-disk and archive verification

All seven original verifier receipts report PASS, exit 0, physical byte equality, and no oracle mutation; their raw stdout/stderr hashes independently match the recorded hashes. Five representative checks cover baseline before opening, renamed save, undo save back to baseline, redo save, and final closed state. Two simple checks cover renamed save and final closed state. Oracle stdout reports exact model/identity with 30 reachable representative sources or four simple sources. The separate physical maps contain 37 and four files respectively. Independent current readback confirms all 41 final physical work files equal their preauthored expected trees. Scope-stress's four physical model files equal its manifest's authored after-hashes. These are saved-disk assertions, not an oracle inspection of a live unsaved editor model.

Versioned archive directory: C:/ar-text-9e8r6t/harness/experiments/text-workspace-native2-20260913. Manifest SHA256 3ddabcf26828afe2aee882bdfbac26d0f3e6003e6c22f1575bb5bfdf12dcb01a, 70,577 bytes. Independently verified all 367 unique payload files / 13,631,658 bytes against the full manifest, exact on-disk membership including only manifest as additional file, and 366 payload files against their independently located originals. The archived initial trial2 helper is historical evidence at its manifest hash; no separate pre-correction live original remains. An initial reviewer mapping to archive_initial.py failed and was rejected because that file belongs to trial1; it was not accepted as provenance. Corrected helper equals the live helper. The parent's original Windows Unicode-decoding failure and corrected explicit UTF-8 helper remain distinct historical files; resumed identical copies are not a new native run.

The nested attempt-5 input ZIP passes full CRC reading (526 members). This is archive integrity, not a rerun of its compiler gates. Payload excludes executables, DLLs, profiles, temporary directories and caches. Baseline/expected trees are archived; final work equality is supplied by the original receipts and this independent current readback, rather than claiming a separately archived final-work tree.

Limits and disposition

Initial empty-text small/scope path-entry captures were failed input attempts later recovered, not successful opens. The scope-stress project was already invalid on load; close/Discard without manual edit or Save is the operator's recorded sequence, consistent with the unchanged physical readback, but disk equality alone cannot prove the absence of transient input. Screenshot metadata/accessibility does not constitute a complete key-event log. Existing Escape correction requires its own source gates and targeted behavior evidence. No full-app, release, timing, accessibility, or final UX acceptance is asserted here.