ARCADIA-Studio/harness/experiments/text-workspace-initial-review-20260913/focus-selection-review.md

Find focus-only correction: independent static review

PASS — no actionable production correctness finding within the bounded delta. Review uses the immutable focus-selection-candidate-1 copies, not the parent's transient worktree during its planned old-controller red run. Candidate manifest sizes/hashes verified. No builds, tests, GUI actions or source edits performed.

restore_source_editor_focus still admits only available navigation and an exact live cursor, then creates the same fresh reveal ticket with interaction/raw clock/snapshot/path. Clearing source_jump is an explicit focus-only intent. source_viewport retains every existing guard, consumes the admitted ticket, and mutates Content only when a genuine Some(cursor) navigation request exists. Existing guarded_focus still rechecks raw input at operation execution and requires one current editor target and no foreign focused widget. Focus-only restoration neither reconstructs Content nor roundtrips the lossy Word/Line cursor representation.

The unchanged intentional-navigation branch still validates exact cursor, moves to it and optionally selects the line; invalid cursor returns before focus. Other production jump-clearing paths inspected also revoke reveal through source_interaction/reset, or are behind the generic interaction invalidation. Thus the new None case does not turn a cancelled navigation request into a valid focus request.

The four new tests address the actual prior evaluator blind spot: semantic selection() and typed replacement bytes, rather than equal flattened Cursor endpoints. They use real mounted mouse double/triple click actions, Find input focus, actual Close button, captured Escape input with explicitly injected existing bridge message, and the actual guarded widget operation. Word and line controls plus the deliberately lossy old cursor roundtrip discriminate the native failure. Accepted Document identity/history remain unchanged while draft bytes receive the replacement.

Evidence limits are accurately disclosed: the CPU-mounted Content is distinct from App's renderer-typed Content; the initial hit-tested caret is copied before semantic selection, and subsequent actual selection/edit actions are relayed to App. Production GPU layout/scroll, full Ctrl+F subscription and the Escape subscription bridge are not end-to-end covered by this fixture. Native replay remains necessary. Debug-only test cfg is consistent with existing shared views::source_tests::FontLayout; it does not claim release-profile testing.

The new later-raw-input case also reopens Find (new interaction target ID and foreign focus), so by itself it does not isolate raw-clock rejection. This is a limitation, not an uncovered production change: existing unchanged source_navigation/interaction/tests.rs::focus_rechecks_clock_between_inspection_and_application_and_missing_target_is_inert isolates clock invalidation between operation passes, while the new test exercises the complete newer-Find refusal scenario. Explicit navigation test directly queues a reveal after resetting Content, so it supplements rather than replaces existing navigation query/reveal regressions.

Parent's old-production/new-regression red run, corrected green/full gates and fresh native replay are outstanding evidence at review time. No native acceptance is claimed.

Bound inputs

• crates/oa-desktop/src/app/source_symbols.rs: 21062 bytes; SHA-256 aa4f1423e759fdeaf7815d656b40189be176f4e8659f3da50debb5d85ed3af6c
• crates/oa-desktop/src/app/text_workspace.rs: 21658 bytes; SHA-256 723ccc8b31c0e2ea8d34841359bf5df086ef866b2bf5c65eefce276c20015bb0
• crates/oa-desktop/src/app/text_workspace/focus_selection_tests.rs: 15091 bytes; SHA-256 074dd8e281b5bf08baffc1667024007c742472bc0a72b2fdb6b1823478824e44
• Attempt-4 archived original app/source_symbols.rs: 20674 bytes; SHA-256 1fb3709bb973e7dcfbcde56054d028e47d770a62f3b7c8b6fac411543a234b7b
• Historical attempt-4 input ZIP SHA-256 e1e566295bbbd6c1d571f7c8908b7344397293496481790f0f8dacab7eba5366.