fix(mobile): select file text across lines on Android#31286

Approvability

Verdict: Approved at 3f40f93

Macroscope's review found this PR approvable — Android source files gain cross-line text selection through a long-press, with Back returning to numbered rows. The change is limited to viewing and copying escaped local content, without changing product defaults or persistent data.

^{You can add or adjust custom eligibility rules. Learn more.}

Explain Change →

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

[GPT-6.1 Sol][Codex] on behalf of shivam.

Phone verification at 6d1310e04c6fc8b51a22aa0427e186dbb9b2642b, POCO F4 / Android 14, matching native fingerprint 56eae38da62e40f29b4dedcd0557b6371825776e, isolated worktree backend and Metro on corex.

Wrap off: cross-line native handle selection and Copy. The paste into an unsent composer exactly matched 65 UTF-16 characters, preserving the tab, emoji, and source newline. Recording at 3× speed.

Wrap off

Wrap on: selected across a blank line and the visually wrapped source line, then Copy. The unsent paste exactly matched 220 UTF-16 characters with the original source newlines. Cleared both pastes without sending a turn. Recording at 3× speed.

Wrap on

The large-file pass found a blocker in this PR. Entering Select text for an 832,000-byte / 16,000-line file caused Android's "isn't responding" dialog. dumpsys activity lastanr reports a 5,000 ms input-focus dispatch timeout. After Wait, the app recovered, scrolling and the menu worked, and exiting restored requested line 120 with wrapping on. The initial stall is unacceptable, so I returned the PR to draft and am investigating a fix. This recording keeps normal playback speed.

Large-file failure

Native fingerprint diff run 42023315261 completed successfully at 11:59:57 UTC on this exact head. Its fingerprint comparison and changed-platform upload both passed.

[GPT-6.1 Sol][Codex] on behalf of shivam.

Phone verification for head b9ae60812822b36c1fd0c0e0d429da312cd4c0be, POCO F4, Android 14, native fingerprint 56eae38da62e40f29b4dedcd0557b6371825776e, isolated worktree backend and Metro on corex. All recordings below play at normal speed.

The earlier 16,000-line ANR prompted a renderer change. Android continuous selection now uses escaped plain text in the existing WebView dependency, with JavaScript and network resources disabled. Normal numbered viewing remains the default. Refresh is available in File actions during selection.

Wrap off: File actions shows Enable word wrap, then Select text; drag the handle across two source lines and Copy. A plain-text paste into the unsent composer exactly matched 64 UTF-16 characters, including the source newline, tab and emoji.

https://github.com/user-attachments/assets/47b47c0f-cce8-4ddf-b5f9-74d403a98bfc

Wrap on: enable wrapping, enter Select text, drag across the blank line and visually wrapped source line, and Copy. The plain-text paste exactly matched 220 UTF-16 characters, preserving source newlines rather than adding visual-wrap newlines.

https://github.com/user-attachments/assets/4ef24458-bb39-408e-be8d-acec697ae79a

Current-head phone layout and cross-line selection:

832,000-byte / 16,000-line file: enter selection, scroll, open File actions, Refresh, leave selection and restore highlighted requested line 120. Repeat with wrapping disabled. No ANR appeared. This is a normal-speed usability recording, not a benchmark.

https://github.com/user-attachments/assets/d77e7972-91ec-44be-a519-69ea3d4122ca

Additional wrap-off large-file pass: cross-line handle selection and Copy, exit to requested line 120, then paste as plain text. The selected substring exactly matched the first two source lines, excluding the final period, at 102 UTF-16 characters.

https://github.com/user-attachments/assets/e7861603-5fea-4274-9ca1-0eb80a2afe2e

Draft attachment preview uses the same Android renderer. Cross-line selection and Copy produced the same exact 220-character plain-text result.

https://github.com/user-attachments/assets/47cfa09b-395d-45a8-bbfd-118fd21874a8

An additional 800,011-byte file with one very long line rendered and scrolled horizontally with wrapping off and vertically with wrapping on, without an ANR. The supporting recording is below. All unsent pastes and the draft attachment were cleared; no provider turn was sent.

https://github.com/user-attachments/assets/a6314074-e405-49e3-b945-98ec23a36f2d

Verification: vp run check:changed passed format, lint, both knip checks, affected typechecks, related tests and React Compiler. The separate Android React run passed both source-surface tests. Independent Codex review found two issues in the first fallback patch, an insufficient size gate and missing Refresh; both were fixed, and a fresh second review found no further findings.

Current-head CI and Native fingerprint diff completed successfully. The evidence-gate-requested run 42023315261 also completed successfully on 6d1310e04c6fc8b51a22aa0427e186dbb9b2642b at 11:59:58 UTC.

Limits: selection starts at the top, omits gutters, syntax colors and target-line highlighting, and leaving selection restores the requested line rather than previous scroll position. Plain clipboard text was compared using Paste as plain text; the composer's separate rich-HTML paste path collapses source newlines. iOS, ZIP attachment navigation, and arbitrary far wrapped-line jumps were not retested. Web, desktop, providers and contracts have no changes.

[GPT-6.1 Sol][Codex] on behalf of shivam.

@coderabbitai review

The PR is ready for review again at b9ae60812822b36c1fd0c0e0d429da312cd4c0be. The Android renderer changed after the phone ANR; current-head phone evidence and completed CI are in the description.

[GPT-6.1 Sol][Codex] on behalf of shivam.

Evidence-gate follow-up on the POCO F4, Android 14, cached native fingerprint 56eae38da62e40f29b4dedcd0557b6371825776e, isolated backend and Metro on corex. All screen content is test data.

Before, clean main 298cb24140fdddebfbfaedad3a097e86ff3d9d04, using its own Metro checkout: long-press line 2 and drag the ending handle down toward line 3. The selection remains confined to line 2. The commit mentioned inside the source is fixture text, not the running main checkout.

https://github.com/user-attachments/assets/e36a3741-b15a-4c70-8f30-b28898edc28c

After, PR head 5b995ee455207cf5057c6c9840d61f2e0afc2210, wrapping off: enter Select text, extend handles across a source newline, Copy, and paste plain text into an unsent draft. The raw editable text matched the selected 58 UTF-16 code units exactly, including the newline, tab and emoji.

https://github.com/user-attachments/assets/742e7e22-b921-4b8f-93f9-7cb3fc19d6d7

Same head, wrapping on: extend through a blank source line and a visually wrapped line, Copy, and paste plain text. The raw editable text matched the selected 259 UTF-16 code units exactly. Both unsent drafts were cleared; no provider turn was sent.

https://github.com/user-attachments/assets/c06c575c-5211-4413-8b1e-fd9d9b695d69

The requested far-line check exposed a remaining failure on this head: requesting line 5,000 with wrapping enabled in an 848,000-byte / 16,000-line fixture lands near line 161. After entering and leaving selection it lands near line 241 rather than line 5,000. The scroll helper is identical to main. This recording documents the failure, not a passing restoration result. A focused follow-up is in progress; this PR is not ready yet.

https://github.com/user-attachments/assets/3d65fe9d-837c-45a4-8846-6ebdbbefa1b1

[GPT-6.1 Sol][Codex] on behalf of shivam.

Current-head Android verification at 04de14090d4282b26166c1c4ce55dcf78b4c45c4, POCO F4, Android 14, native development fingerprint 56eae38da62e40f29b4dedcd0557b6371825776e. Metro serves this checkout on corex, paired to isolated worktree state. Only test fixtures are shown. These are normal-speed recordings, not a performance benchmark.

Wrapping off: File actions > Select text, long-press, extend handles across source lines, Copy, and plain-text paste into an unsent draft. The raw composer value exactly matches the selected fixture substring, 57 UTF-16 characters, including the source newline, tab and emoji. The selection ends after the emoji.

https://github.com/user-attachments/assets/48c469b0-15a8-4d42-a06e-7707383002a1

Wrapping on: cross-line selection includes a blank source line and a long visually wrapped line, then Copy and plain-text paste. All 259 UTF-16 characters match the selected fixture substring exactly, including the blank line, tab and emoji. Both drafts were cleared without sending a turn.

https://github.com/user-attachments/assets/328857cf-a0b4-4e2e-b1cb-b5cbe89d9d0f

The 848,000-byte, 16,000-line file, wrapping on: open requested line 5,000, enter selection, scroll the continuous document, and leave selection. The numbered viewer settles at highlighted line 5,000. This replaces the failed far-line result recorded at 5b995ee.

https://github.com/user-attachments/assets/2001eda5-53f4-43c6-8e7c-8a1d89056b10

Same large file, wrapping off: enter selection, scroll, leave selection and return to requested line 5,000 in the native numbered viewer.

https://github.com/user-attachments/assets/e90e642c-f334-43a6-a077-26905eefe212

The original main restriction and the failed pre-fix far jump are recorded in https://github.com/pingdotgg/t3code/pull/31286#issuecomment-6660242971. Main recording uses clean 298cb24140fdddebfbfaedad3a097e86ff3d9d04 and shows long-press plus attempted extension to the next source row.

The new scroll regression reproduces native clamping and fails with the old helper. It also asserts that intervening rows are not measured. An independent source review found no code findings. A stalled native layout can leave a jump pending until another content-size event or cancellation; the retry budget is not a wall-clock timeout. iOS runtime verification and current-head CI are still pending, so this is not a ready report.

The primary independently repeated vp run check:changed at 04de14090d4282b26166c1c4ce55dcf78b4c45c4; it completed successfully. Full output:

[GPT-6.1 Sol][Codex] on behalf of shivam.

Macroscope's stalled-layout finding is fixed in abd08af6244aebc3168ffaf778c62406cf16b4aa. The scroll helper now keeps a bounded timer fallback while awaiting layout, cancels a superseded timer when a layout event retries the jump, and resets the budget only when measured rows advance. Exhaustion clears the pending target and reserved tail.

Extended the existing stalled-target test. It fails at 04de14090d with expected 21576 to be 0; with the fix all seven helper tests pass, including the far-jump case whose layout is slower than the timer. A later content-size event also cannot restart an exhausted jump. An independent round 2 review found no findings.

Repeated the affected native flow at exact head abd08af6244aebc3168ffaf778c62406cf16b4aa, POCO F4, Android 14, native fingerprint 56eae38da62e40f29b4dedcd0557b6371825776e, Metro and isolated backend on corex. The 848 KB / 16,000-line file opens at requested line 5,000 with wrapping enabled. The recording shows selection entry, scrolling, exit to highlighted line 5,000, then a manual drag that stays at rows 5,004 onward.

https://github.com/user-attachments/assets/bf66e244-0bc3-4ebc-a299-58c89b409768

Only the scroll helper and its regression test changed since 04de14090d. The wrap-off/on cross-line Copy and large-file recordings in https://github.com/pingdotgg/t3code/pull/31286#issuecomment-6660718359 cover the unchanged selection renderer.

vp run check:changed completed successfully before commit; commit hooks also passed:

iOS verification is still waiting for the matching native client build on shiv. This is not a ready report. No performance benchmark or stalled native-layout fault injection was performed on the phone; the stalled-layout cleanup is proven by the regression test.

[GPT-6.1 Sol][Codex] on behalf of shivam.

The shared-cache authentication failure is bypassed using the matching local simulator archive returned by the shiv job. No cache credentials were changed.

Verified current JavaScript head abd08af6244aebc3168ffaf778c62406cf16b4aa on the Device panel's iPhone 17 simulator, iOS 27.0, with Metro and isolated worktree backend on corex. The cached arm64 development client is version 2.2.0, build 1, actual native fingerprint e8f78d5e730353790c0a572575c387141109cfa4. The repository prepare helper selected it on shiv at exact 04de14090d4282b26166c1c4ce55dcf78b4c45c4; only the scroll helper and its test changed afterward, with no native-input changes. This is a reused cache artifact, not a fresh Xcode compilation.

The recording shows the 848,000-byte / 16,000-line fixture at requested line 5,000 with wrapping off, File actions without the Android-only Select text item, enabling wrapping and landing at highlighted line 5,000, a manual scroll that stays where dragged, then opening a small file and returning to the large file at requested wrapped line 5,000. It is a continuous normal-speed excerpt with the first 150 seconds of idle menu time removed, resized for upload. The complete source recording decoded without errors after recovering its file transfer outside the command's 90-second limit. Only test data is visible.

https://github.com/user-attachments/assets/3cf0a2c2-0241-4b4f-b2ad-b2455dc0bb2d

Final settled wrapped line 5,000:

An independent artifact comparison found all native sources and the other 117 fingerprint inputs matched. The Linux patched keyboard-controller directory also contained generated Android outputs and unrelated ISSUE31223 JavaScript diagnostic listeners from another task. Those listeners were present in Metro; source-viewer code was the exact PR head. The shared package was left untouched.

CI run https://github.com/pingdotgg/t3code/actions/runs/42048519202 and Native fingerprint diff https://github.com/pingdotgg/t3code/actions/runs/42048880809 both completed successfully at the full head above. The primary's check:changed output and current-head Android flow are in https://github.com/pingdotgg/t3code/pull/31286#issuecomment-6660978607. All review findings are resolved.

Limits: iOS was a simulator, not a physical iPhone. ZIP attachment navigation was not repeated. These recordings establish usability, not a performance benchmark. Stalled-layout cleanup is covered by the focused regression test, not native fault injection. The app session and simulator were closed, and the owned Metro, backend and SSH forwards were stopped.

[GPT-6.1 Sol][Codex] on behalf of shivam.

Rechecked the base move for head abd08af6244aebc3168ffaf778c62406cf16b4aa. Main advanced through the reported 268e5810cd69393be362d5062d456f45e21b8492 to 2d0efd35840dd3118939cd27b096340ea522a541; I checked the latter directly from refs/heads/main. The pull-request REST/GraphQL base SHA still returns the older 298cb24140fdddebfbfaedad3a097e86ff3d9d04.

The only overlap is apps/mobile/vite.config.ts. Main removes two list entries for deleted tests; this PR adds its two file-viewer test entries elsewhere. A clean git merge-tree against current main produces tree 6c3a7e6b171755b7c5bbfe5bce4b64352408055b. Its mobile configuration is exactly the head configuration minus those two obsolete entries, with both PR entries preserved once each. Comparing it to the prior checked merge tree shows no new changes in apps/mobile/src/features/files/.

Temporarily applied only that combined configuration and ran the three affected test files with T3_MOBILE_TEST_PLATFORM=android. All nine tests passed. Restored the original configuration; the worktree is clean, the branch and PR head are unchanged. No runtime changes result from this overlap, so no new device recording was needed. Existing POCO F4 and iPhone simulator evidence remains scoped as recorded in https://github.com/pingdotgg/t3code/pull/31286#issuecomment-6660978607 and https://github.com/pingdotgg/t3code/pull/31286#issuecomment-6662512903.

Command, from apps/mobile:

The live PR is open, not a draft, cleanly mergeable, and has no auto-merge enabled. Required checks pass at the unchanged head, and all review threads are resolved. The completed fingerprint run is https://github.com/pingdotgg/t3code/actions/runs/42059512799. No Julius takeover was found.

[GPT-6.1 Sol][Codex] on behalf of shivam.

Rebased the seven commits without conflicts onto main 7bc2a40ced9094e2b5fd2f076e1bfeab734f8aed. New head is ac80d42602414146fa5b0c7f5f3edb9f4a25e123. The 648-file base move overlaps the file-screen header/layout and test configuration. A separate read-only compatibility review found no code findings, and the rebased tree exactly matches its reviewed merge tree c8122fcbd344d502a93e28f73117ed9393f839f7.

Ran vp run check:changed 7bc2a40ced9094e2b5fd2f076e1bfeab734f8aed against the eight PR files:

The base also changed native review-diff, highlighting and composer modules. I am preparing matching new Android and iOS development clients and will repeat the affected phone and simulator interactions. Earlier recordings remain historical evidence, not verification of this rebased combination. This head is not ready yet.

[GPT-6.1 Sol][Codex] on behalf of shivam.

Repeated Android verification at ac80d42602414146fa5b0c7f5f3edb9f4a25e123 after the base rebase. POCO F4, Android 14, development native fingerprint 91f90d1d53ec7e79d90fd83538ace1a0c7f7b393. The repository installer confirmed that exact client compatible. Metro and the isolated worktree backend run on corex. Only test fixtures are shown; no provider turn was sent.

Wrapping off: enter Select text, long-press and extend a handle across the source newline, then Copy. The first pass's raw pasted editor value matched 93 UTF-16 units including a tab and emoji. Its pasted view is brief, so a second continuous recording below shows a clearer paste. That second selection ends just before the emoji and matches all 91 UTF-16 units exactly, including the newline and tab.

https://github.com/user-attachments/assets/0c07e2d7-bfe1-42f1-8148-12cca1a99fea

https://github.com/user-attachments/assets/549cc3ba-e902-4107-8e22-102497850f7e

Wrapping on: extend across a blank source line and a long visually wrapped line, then Copy and plain-text paste. The raw pasted value matched all 302 UTF-16 units, including the blank line, tab and emoji. Drafts were cleared after comparison.

https://github.com/user-attachments/assets/00b92460-abcc-4330-93bd-22d188882a0a

Large fixture, 848,000 bytes and 16,000 lines: with wrapping on, open requested line 5,000, enter selection, scroll, leave selection and restore highlighted line 5,000. A subsequent manual drag stays at the new position.

https://github.com/user-attachments/assets/33709bf2-094f-40da-b848-c0b067096db7

With wrapping off, the same file returns to requested line 5,000 after selection. This also exercises Refresh while selecting. These are normal-speed interaction recordings, not timing benchmarks. A transient empty canvas while toggling wrapping settles to file content.

https://github.com/user-attachments/assets/1e2e0683-bb8d-4091-9409-c04795d07eac

Current-head CI completed successfully, including Native fingerprint diff run 42144287927. The scoped check output is in the preceding rebase report. The original main restriction and prior red/green scroll regression remain in the historical evidence linked in the PR body.

Not ready yet: a matching fresh iOS integrated pass is outstanding. The Android attachment was imported through the picker, but tapping its composer chip did not open the preview, so that preview interaction remains unverified after this rebase.

[GPT-6.1 Sol][Codex] on behalf of shivam.

Completed the remaining Android attachment preview check at ac80d42602414146fa5b0c7f5f3edb9f4a25e123, on the same POCO F4, Android 14, native fingerprint 91f90d1d53ec7e79d90fd83538ace1a0c7f7b393, Metro and isolated backend on corex. Opened a test attachment through its existing thread attachment route, enabled wrapping, long-pressed and dragged a handle across the source newline, then Copy. Disabled wrapping, left the preview, and pasted as plain text into an unsent draft. The raw editor value exactly matches the selected 90 UTF-16 units, including the newline and tab. Cleared the draft without sending.

https://github.com/user-attachments/assets/5fef51ff-ce95-45d6-92c9-a4b823acba08

This supersedes the attachment-verification gap in the preceding comment. The source was a 333-byte test fixture copied into isolated attachment storage. The composer chip interaction was bypassed using the existing attachment deep link.

The fresh iOS run is unnecessary for this PR under fleet's current verification rules. device-route --base 7bc2a40ced9094e2b5fd2f076e1bfeab734f8aed reports iOS simulator or Android phone: one is enough, and full-send says shared mobile code needs one phone. This PR adds no iOS-specific behavior; the current-head F4 recordings cover the shared scroll helper and the reporter's platform. The earlier iOS recording remains historical evidence, with no claim of a fresh iOS pass at this head.

A separate read-only review checked the subsequent 75-file base move through c52df24d285a072ad9c3f7ab5b3656c0be1cb0f5. The merge is clean, and all eight PR files are unchanged in combined tree f3c332744881b8254111ae2b6927f3e74ca85ea3. No code findings. New main changes affect PR discussions and composer input, with no overlap in the file viewer or scroll helper.

[GPT-6.1 Sol][Codex] on behalf of shivam.

Verified current head de78e46b72eeaa27ae12302df15c1c68207d9a99 on POCO F4, Android 14, native development fingerprint ef469ea7401d69afda88e4de10e16a396759a9b9. Metro and the isolated worktree backend ran on corex. All displayed files and the empty thread are test fixtures. No provider turn was sent.

With wrapping off, File actions > Select text, long-press, drag across the source newline, Copy, Done selecting text, and plain-text paste. The raw pasted editor value exactly matched the selected 59 UTF-16 units, including newline, tab and emoji. The recording trims only the initial idle time.

https://github.com/user-attachments/assets/6ffa7dd8-ef5d-434d-9560-a587f0512544

With wrapping on, selection spans a blank source line and several visual wraps. Copy and plain-text paste exactly matched 246 UTF-16 units, including blank line, tab and emoji. Leaving selection returned to highlighted requested line 20. All unsent drafts were cleared.

https://github.com/user-attachments/assets/039d06b3-5b1f-4b02-8cbb-6f6b86f9e869

The 848,000-byte, 16,000-line fixture supports entering selection, scrolling and Refresh, with wrapping on and off. Leaving selection returned to requested line 40. The first recording ends before the final unwrapped canvas settles; the second records the settled unwrapped return. These show interactions, not a performance benchmark.

https://github.com/user-attachments/assets/a692f692-5c1d-47f1-bbb5-1fa89f925b29

https://github.com/user-attachments/assets/6a276e08-36c8-4e85-b181-dea8036c44dd

The shared Android attachment renderer also supports cross-line selection and Copy with wrapping enabled. The first clip shows that selection and Copy, then wrapping disabled. The separate paste clip shows the selected 59 UTF-16 units and clearing the draft. The raw value exactly matched the source selection.

https://github.com/user-attachments/assets/3bfcd8ec-84ed-42a6-9d01-eb704cc0f708

https://github.com/user-attachments/assets/249d7c93-c03e-44e1-9824-ed3d9d27f922

fleet-job check 31286 passed at this exact head, merge base 0f4a13071090e89c6d7eccdb0eb98b3cde1823ee, command vp run check:changed origin/main, exit 0:

Native fingerprint diff run 42357403824 completed successfully at this head. All 33 reported checks are success or intentionally skipped, with none pending or failing.

The before recording on clean main remains the baseline for the row-selection restriction. Earlier ac80 recordings are historical. The removed far wrapped-line scroll rewrite is not part of this diff; main's existing distant wrapped-line jump limitation remains. Selection starts at the top, without gutters, syntax colors or target highlighting. Done requests the linked line through main's unchanged helper, rather than preserving the prior scroll position. No far-line or timing guarantee is claimed.

Shared mobile verification requires one iOS simulator or Android phone under fleet's device-route rules; the reporter's Android platform and that requirement are covered by the F4. The draft workspace entry shares the reviewed file screen but was not separately driven. The attachment was opened through its existing deep link.

Final independent code and evidence review found no blocker, should or nit findings. The live PR remains open, non-draft and mergeable, with auto-merge disabled and every review thread resolved. Its reported base is 0f4a13071090e89c6d7eccdb0eb98b3cde1823ee; current main 83a3674ffd84ffb33de768210c806cb1ac53fd80 adds only web scrollbar CSS, with no mobile overlap. Current head remains de78e46b72eeaa27ae12302df15c1c68207d9a99.

[GPT-6.1 Sol][Codex] on behalf of shivam.

The description-check warning compares the current description against CodeRabbit's stale file summaries from 04de140, while automatic code review is paused. At current head de78e46b72eeaa27ae12302df15c1c68207d9a99, source-file-scroll.ts and its test are byte-identical to PR base 0f4a13071090e89c6d7eccdb0eb98b3cde1823ee. Neither is in the six-file final diff. The removed tail-reserve implementation and its stalled-jump warning therefore do not apply to this PR. Main's existing distant wrapped-line limitation remains explicitly documented.

Added the repository's Problem, Change, Scope and approval, and Verification headings to the PR description. The current description matches the final diff. No code changed. Fresh independent review and the current-head phone recordings and check results cover the remaining selection change.

[GPT-6.1 Sol][Codex] on behalf of shivam.

Android evidence for the long-press flow. POCO F4, Android 14, development native fingerprint fe047aa373fabca2de1892d5d68134782e4d25bf, installed through the repository installer. Metro and the isolated worktree backend ran on corex, using test fixtures only. No provider turn was sent.

The selection, Copy and large-file recordings were made with e20df26ead21b1d549c334af26e96f52252662d4. Subsequent commits add the accessibility action, navigation handling and mode reset, and hide the unused highlighting indicator. They do not change the WebView HTML, anchor scrolling or clipboard script. The final two recordings exercise those affected paths at acda47d822a8bbdf2aa59ffcc9e9c01fd21ea086.

Wrapping off: long-press a numbered row, long-press again, extend across the newline, Copy, Back, then normal Paste into an unsent composer. The raw editor value matches the selected source prefix exactly, all 59 UTF-16 units including newline, tab and emoji. Entry was near the first rows; this clip does not claim a particular return line.

https://github.com/user-attachments/assets/9b61ba0c-fca9-4a01-8c17-ed515d01a626

Wrapping on: selection crosses the blank line and several visual wraps. Normal Paste matches all 246 UTF-16 units exactly, preserving blank line, tab and emoji. Back restores highlighted line 1. All drafts were cleared after comparison.

https://github.com/user-attachments/assets/fe0757c0-79c0-4b01-95ca-5e95a7beaf05

Large fixture, 848,000 bytes and 16,000 lines: long-press near requested line 40 with wrapping enabled, enter selection and scroll. The first clip shows entry and scrolling; the supplement shows the settled numbered viewer at highlighted line 40 after Back. These are interaction recordings, not timing benchmarks. Reading this WebView's 16,000 accessibility anchors exceeded the driver's observation limit, so no hierarchy-read latency is presented as app performance.

https://github.com/user-attachments/assets/0cf71144-00a8-4c49-a358-cb293e870ac4

https://github.com/user-attachments/assets/4c68a58b-8e5e-445c-b348-0fd58ebef2b0

Wrapping off: enter selection at line 5,000 and Back to the settled numbered viewer at line 5,000.

https://github.com/user-attachments/assets/49ae3b44-8afd-4c30-ac87-87665beb8b30

Android attachment source preview, same renderer: extend across a source newline, Copy, return to the empty thread and normal Paste. The raw 59 UTF-16 units exactly match the selected prefix, including tab and emoji. The attachment was opened through its existing deep link. The pause before Paste is untrimmed.

https://github.com/user-attachments/assets/57621f93-3dd9-44f2-a31d-26f1326fd124

Final-head mode reset: enter selection in a Markdown file, choose Preview, then Source. Source returns to numbered rows rather than reopening the prior selection. This verifies the handler reset added in acda47d822 after independent review found that changing the scope alone was insufficient.

https://github.com/user-attachments/assets/7e904967-4dbb-4df5-b7ff-5b8e53c9dff2

Final-head accessibility: invoke the row's advertised long_click accessibility action, observe the anchored selection text, then Back to numbered rows. This is an accessibility action invocation, not a claim of a full TalkBack walkthrough.

https://github.com/user-attachments/assets/d5706f5a-c6c6-4607-a431-dd90a96cce11

The main before recording shows the original single-row restriction. Main's distant wrapped-line jump helper is unchanged: Back requests the pressed line but can miss a far target. This is a known limitation, not merely a settling delay. No guarantee of far wrapped-line restoration or previous-scroll preservation is claimed. Return to chat in split layout and the draft entry were reviewed in code but not separately driven. The required shared-mobile phone pass and reporter's Android platform are covered by the F4; no fresh iOS run is claimed.

[GPT-6.1 Sol][Codex] on behalf of shivam.

Final verification at head acda47d822a8bbdf2aa59ffcc9e9c01fd21ea086, base d6c2b09c52e96a71e22124a734190cc0d14ca25f. fleet-job check 31286, job jljdtkrplsgakevf, ran vp run check:changed origin/main, exit 0. I read its completed result; its Boat VM is confirmed gone.

CI run 42451757376 and all required checks passed at this head, including completed Native fingerprint diff. Macroscope approved this exact commit. Every review thread is resolved. A fresh independent second review confirmed the mode reset, corrected far-line claims and uploaded phone evidence, with no remaining findings.

CodeRabbit's description warning about pending results is addressed in the body. This focused bug fix uses CONTRIBUTING.md's obvious-bug exception; no separate maintainer approval is claimed. Its older stalled-scroll warning and summary still describe the removed 04de140 rewrite. source-file-scroll.ts and its tests are unchanged from current base. The actual known distant wrapped-line limitation is explicitly disclosed, rather than claimed fixed.

The test app is uninstalled, the F4 released, and the isolated backend and Metro stopped. Nothing enables automatic merging.

[GPT-6.1 Sol][Codex] on behalf of shivam.

Rechecked head 0321ec053a85e58788726a3dc5d7a578c9b10c18, base 09a54f01bc0ae83865347a965ced50f8c8ec2953. The new merge adds only main's thread-search ranking change in paletteSearch.ts and its test. All mobile files are byte-identical to previously verified acda47d822a8bbdf2aa59ffcc9e9c01fd21ea086; neither the new base nor head changes the file-viewer entry, selection, clipboard, attachment or navigation paths. The F4 recordings and independent reviews still cover this final diff.

fleet-job check 31286, job bcd2s43gvdup44jg, ran vp run check:changed origin/main on this exact head and base, exit 0. I read the result. Its Boat VM is confirmed gone.

Current-head CI passes. Native fingerprint run 42452569484 completed successfully on this head, superseding cancelled run 42452542020. Macroscope approved this exact commit. Review threads remain resolved; the PR is open, non-draft, mergeable and auto-merge is disabled. CodeRabbit's updated summary adds no new actionable finding and still references the removed scroll rewrite. The existing distant wrapped-line return limitation remains disclosed.