ARCADIA-Studio/harness/experiments/text-diagnostics-20260913/reviews/native-trial-2-review.md

Corrected diagnostic native trial — independent evidence review

Recommendation: accept the bounded Windows native correction for readable passive diagnostic hover, together with the source-bound gates and parent's final physical-fixture check. The original blank-tooltip failure is visibly resolved in the inspected corrected captures. This is not complete application, peer integration, packaging, Linux, minimum-window, or exhaustive interaction acceptance.

The original launch records GUI SHA256 9cff5a094e3015d95a2c5fdb8b1ceaf03aced6a09670de32a4ec5ff8dcc94900 (114,221,056 bytes), PID 1164, 2026-09-13 09:07:04.451663 UTC. Retained-child closure records normal exit 0 at 09:23:49.269580 UTC, unchanged trial/preserved binary and DLL, same-file identity and no receipt error. Independent readback confirms stdout/stderr are empty and match closure hashes. This review binds these original receipts; it does not rerun the process or independently rehash the executable. Parent owns the final disk-fixture equality check.

Viewed original JPEGs for ordinals 007, 011, 013, 016–019, 025–026, 030–031, 037–041, 043, 046–049, including both images at 040. Capture 040-0 is only a tiny partial image; the actual maximized view is 040-1. The folder contains 53 JPEGs and 49 metadata records. All their raw hashes, launch/closure/manifest and oracle records are bound in native-trial-2-review.json, SHA256 92da9da3dd9e184280b468ffecf6361f5e32244ae09f6f27b6cd2f2e328de51d. Only the listed subset was visually inspected.

Observed behavior

• 007 displays readable Error [missing_include], the missing file explanation and Source problems footer above the red include underline. 011 and 013 display both overlapping warning messages and footer at wrapped portions of the corresponding first/second declarations. Severity, code and content are visible, unlike trial 1's blank rectangles.
• 016 selects the first alias word while hover remains visible. 017 replaces only that alias with receive_changed; its old underline/tooltip is absent and the file is marked changed. 018 restores receive_request and the original warning presentation with Undo. This is a discriminating word-selection replacement check, not merely cursor equality. Screenshots do not measure the exact invalidation latency.
• 025–026 show a trailing standalone brace with a red token marker and readable Error [invalid_source] at 33:1, “Unknown declaration '}'.” 030–031 show the unfinished declaration's end marker and readable error at 33:38, “Expected a field, element declaration, or '}'.” The saved candidate oracle independently supplies the corresponding nonempty 1127..1128 and empty EOF 1164..1164 spans; screenshots establish the visible locations/content, not direct unsaved byte equality.
• 037 actually scrolls the editor while the drawer remains mounted: the second declaration is clipped at the upper editor boundary. 038 shows its correct readable warnings at the scrolled visible description; 039 restores the top. 040-1 shows reflow at the recorded 2245×1232 maximized window, and 041 returns to 1442×932. Underlines follow the changed visible wrapping. This is ordinary/maximized geometry evidence, not minimum-width acceptance.
• 046 shows the confirmation modal with no tooltip painted over it. 047 shows cancellation returning to the editor; 048 again shows readable hover. 049 begins another close confirmation from active hover and suppresses that tooltip. Underlines remain visible beneath the modal dimmer, consistent with hover-only suppression.

Oracle comparison and limits

Current-oracle-3 and candidate-oracle-3 both record PASS against the same source-input manifest (129607a08fea6f9eeae0d40982111c7513b243f6bb311bea8fd59026ce388076). Independent readback verifies each original stdout/stderr against its receipt. Missing/complete exits are 3/0 with two missing_include errors versus five complete warnings. Token/EOF exits are 1/1 with one invalid_source error each. Native missing status shows one error; do not claim exact GUI/discovery/compiler duplicate-count parity. Saved reports are lint-only and do not certify the unsaved GUI model or edits.

019 is labeled triple-click declaration line, but the image still selects only receive_request; it does not prove Line-mode selection. The initially attempted scroll 014 had all text fitting, and resize attempt 015 did not change dimensions; actual evidence is the later drawer scroll and maximize. The failed Restore capture/minimize episode followed by activate/get recovery must not be described as a clean direct Restore test. 043 says “No previous source location in this workspace” and remains on project.arcadia, so it is not a successful Back-return check after a Problem action. Those unsuccessful attempts do not invalidate the narrower annotation observations, but must stay outside accepted claims.

Some pointer placements used zero wheel events. The visible hover is real, but these actions do not isolate pure CursorMoved handling or prove an unchanged interaction clock; mounted input tests provide that separate evidence. This review does not establish native Line-mode preservation, exact stationary-pointer scheduling, screen-reader behavior, arbitrary bidi/EOF geometry, GPU performance, Save/Apply transactions, or final disk equality. No UI operations, CLI execution, source/archive edits, Git or Gitea actions were performed by the reviewer.