# Issue 18300 — unified corrected evidence index

This is the COMPLETE replacement evidence index: C1–C6, closing-window proof, guards, original provenance, limitations and public independent audit. It supersedes the [original index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/256b99d7d63f0f2f/index.md) and [pass-2 index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6787baeb8c0e6947/index.md) for captions, timestamps and still/diff selection. Original recordings are retained unchanged.

## Explicit correction and supersession

The old C3 after seek at 14.6s and C4 after seek at 1.7s show **Stop 1**, not Stop 2. Their post-J Stop 2 captions are withdrawn. The old before stills extracted by seeking to 1.65s also show Stop 1; they are replaced too. ALL prior C3/C4 before/after stills and their gain-1 and gain-8 diffs/counts are superseded by the corrected pairs below. Old images remain historical artifacts only. A coincidentally identical T16 image/count does not validate an old frame selection.

The corrected full-resolution stills use exact zero-based decoded frame selection, not approximate seeking. Before frame 16 and after frame 17 in each original recording visibly show **Stop 2 of 3**, after the first J following prose focus, before the subsequent K. C3: before PTS 1.427433, after PTS 14.589900. C4: before PTS 1.427400, after PTS 1.513233. These are actual decoded PTS, not the old rounded seek requests. Both narrow revisions already navigate successfully; this pair demonstrates retained narrow behavior, not a narrow regression fixed by the patch.

All older narrated action times throughout this package are approximate **tool wall offsets**, not frame-accurate video timestamps. Sparse variable-rate native recording PTS can differ from wall time; a seek to the next frame may show the next key's state. C1/C2 and C6 legacy extraction times are seek requests, not guaranteed exact PTS. Only the corrected C3/C4 PTS above are asserted as exact frame timestamps here. C6 labeled reel segment offsets are encoded reel boundaries. Captions claiming old action offsets were seconds from the original video are superseded.

No retouching, fake DOM, new capture, server or browser was used for this correction. Source MP4 copies were verified byte-identical to the app-owned attachment originals. PNG extraction applies only decoded-frame selection: no crop, mask, alignment, scaling, color adjustment or overlays. Full-frame gain-8 diffs use signed RGB absolute differences; T16 filters on max raw channel delta >16 before gain, with saturated 8× amplification.

[Exact frame indices, PTS, original attachment IDs, SHA-256 and recomputed counts](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a868d330d4d597ca/corrected-stats.json).

### Corrected C3 — matched first-post-J Stop 2 pair

| Before: Stop 2 | After: Stop 2 | Gain 8, T0 | Gain 8, T16 |
|---|---|---|---|
| ![C3 before Stop 2](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ee671f38f84831b5/C3-before-corrected.png) | ![C3 after Stop 2](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c42eb871e4fa609e/C3-after-corrected.png) | ![C3 corrected T0 gain 8](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0280c5eb2fb1251a/C3-diff0-gain8-corrected.png) | ![C3 corrected T16 gain 8](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b0c7489da43afb32/C3-diffT16-gain8-corrected.png) |

Raw changed pixels: T0 157,882; T16 3,793 of 1,316,640.

### Corrected C4 — matched first-post-J Stop 2 pair

| Before: Stop 2 | After: Stop 2 | Gain 8, T0 | Gain 8, T16 |
|---|---|---|---|
| ![C4 before Stop 2](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/457c2b259bba80be/C4-before-corrected.png) | ![C4 after Stop 2](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a1859b75023e7078/C4-after-corrected.png) | ![C4 corrected T0 gain 8](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0244602a4f24ceed/C4-diff0-gain8-corrected.png) | ![C4 corrected T16 gain 8](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/75017d6bfda8ed05/C4-diffT16-gain8-corrected.png) |

Raw changed pixels: T0 45,406; T16 3,177 of 1,316,640.

## Independent proof audit (historical report, correction follows it)

[Read-only proof audit pass 2, unchanged](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c261cdb48f3377b4/18300-proof-audit-pass2.md). Its INSUFFICIENT verdict describes the package before this correction; no fresh independent sign-off is claimed. Its behavioral findings, explicit unverified boundaries and independent-rebuild statistics remain relevant.

![Independent candidate rebuild: Stop 2 with empty composer](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3f10b40c25cfc797/rebuilt-C1-after.png)

The independent C1 rebuild has 19,816 T0 / 801 T16 changed pixels of 5,184,000 against original C1 after, mean absolute per-channel delta 0.0207275/255. Cursor position and compression differ; observed reproducibility is not pixel identity. Report and image were inspected before public upload; no credentials or private conversation content were found.

## Rendering and comparison limits

Real Vite development app; Linux x86_64, Chrome 154.0.0.0, DPR 2. Desktop CSS viewport 1440×900, original video frames 2880×1800. Narrow web CSS viewport 390×844, frames 780×1688; this is not native mobile. System UI font stack: -apple-system, BlinkMacSystemFont, Segoe UI, system-ui, sans-serif; `fc-match sans-serif` reports DejaVu Sans; the exact Chromium fallback face was not independently resolved. Interface 16px, composer 14px, code 13px. Light/dark via prefers-color-scheme emulation, app follows system. Same task .t3, real linked PR, synthetic Stop 1/2/3 with no files. Stored guide head af68de7a2a51bf626d630916fdf711383f32fd03 is metadata, not either code SHA.

Browser Date was frozen after setup to 2026-10-23T01:44:00Z. This did not retroactively replace already-rendered relative ages. C6 therefore has incidental sidebar-age differences, and real host metadata changed independently (Merge versus Resolve conflicts). These are not clean pixel-equivalence proofs. Core desktop sidebar is closed; narrow sheet covers it; C6 sidebar is open. Scroll is at the top of the guide. All frames are unmodified extracts of original recordings; no DOM screenshot fabrication. Preview saved screenshots proved stale/downsized, so they are not used in the tables.

All current diff tables use gain 8 with raw T0/T16 filtering as defined above. Earlier gain-1 diffs are superseded.

## Current provenance, C5, closing and guards


C5 pending→arrival is now recorded through the real app and real server transport on both [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc) and [e8de9986ed](https://git.cafe/pingdotgg/t3code/commit/e8de9986ed52d2b241d5eef1c556e1fbe04d20ec). Closing-during-exit is independently recorded against the prior candidate [8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3), served from fixture commit [b9aec92ec9](https://git.cafe/pingdotgg/t3code/commit/b9aec92ec95dc1bf77734734b6494790234fd6f6), and fixed production [e8de9986ed](https://git.cafe/pingdotgg/t3code/commit/e8de9986ed52d2b241d5eef1c556e1fbe04d20ec). These are two different comparisons. Final guard reel is [fee1f231f6](https://git.cafe/pingdotgg/t3code/commit/fee1f231f67bd18feac0814c9cd34944cd48a9c3); its only difference from [e8de9986ed](https://git.cafe/pingdotgg/t3code/commit/e8de9986ed52d2b241d5eef1c556e1fbe04d20ec) is removal of a test, not production code.

Historical source: [previous index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/256b99d7d63f0f2f/index.md) and [matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3bcf454692f7dd5c/issue-18300-proof-audit-pass1.md). C1–C4/C6 recordings remain at their original revisions; corrected C3/C4 frames below come from those same recordings; they have not been relabeled as final-head captures. No production edits, fixture edits, commits, push, PR mutation, comment submission, generation, or provider launch by this test agent.

## C5: matching action sequence

Cold setup: remove only this owned origin's `t3code:right-panel-state:v2` immediately before reload. After chat is ready, install [the transport hold](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2422a2c7a5d8c35f/c5-transport.js), then open the real linked PR and select Guide. The real `pullRequests.reviewGuide` request starts on opening the PR, before selecting Guide. Hold its outgoing WebSocket send, verify exactly one held request and no guide, record the real spinner, click non-editable chat, release the original send unchanged, then await real Stop 1. No response substitution, DOM guide fabrication, component harness, fake generation, or production patch.

The web runtime used here has no persisted EnvironmentQueryStore provider; a reload alone did not help earlier attempts because they paused after the PR had already queried the guide. The hold-before-first-PR-mount recipe reliably establishes the interval regardless of cache assumptions. No unrelated origin storage was cleared.

| Evidence | Before [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc) | After [e8de9986ed](https://git.cafe/pingdotgg/t3code/commit/e8de9986ed52d2b241d5eef1c556e1fbe04d20ec) |
|---|---|---|
| Original recording | [C5-before-51273020db.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/546dbea4b3cc90b1/C5-before-51273020db.mp4) | [C5-after-e8de9986ed.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ee307fb6e0536ed4/C5-after-e8de9986ed.mp4) |
| Click chat while spinner present | Yes | Yes |
| Real guide arrives, then first j | Stop 1; draft `j` | Stop 1; draft `j` |
| Click Stop 1 prose, then second j | Stop 1; draft `jj` | Stop 2; draft `j` |
| Final frame, unmodified video extract | ![Before: Stop 1 and composer jj](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/41e74f8f2b68aa0f/C5-before.png) | ![After: Stop 2 and composer j](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/54feb3639e3b0c0f/C5-after.png) |

Pacing is the same sequence, not identical timing: before chat click 2.757s, release 4.882s, first j 7.141s, prose 10.286s, second j 11.362s; after 2.728s, 4.792s, 7.181s, 9.653s, 10.742s. These are tool wall offsets, not frame-accurate video timestamps. Native recording startup/encoding can shift PTS. Both final stills are extracted at video 12.5s and visually checked. The final state is held for 3 seconds before stopping.

## Closing during exit: prior candidate, not original base

Set the actual client `panelAnimationDurationMs` to 400 and reload; DOM confirms `data-panel-animations=true`, `--panel-animation-duration:400ms`, and reduced motion false. Start at engaged Stop 2 with empty composer. Press Linux Mod+Alt+B (Ctrl+Alt+B), immediately press J. Passive window-capture telemetry observes trusted input and actual mounted panel; it dispatches no events and changes no app state.

| Evidence | Prior candidate [8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3), fixture-only head b9aec92ec9 | Fixed production [e8de9986ed](https://git.cafe/pingdotgg/t3code/commit/e8de9986ed52d2b241d5eef1c556e1fbe04d20ec) |
|---|---|---|
| Original recording | [closing-before-b9aec92ec9.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5025a14f86b7ef09/closing-before-b9aec92ec9.mp4) | [closing-after-e8de9986ed.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5c4e2e155f6fbf28/closing-after-e8de9986ed.mp4) |
| Trusted J after close keydown | 86.3 ms | 249.3 ms |
| Guide still mounted at J | Yes, Stop 2, `data-guide-tour-keys=jk` | Yes, Stop 2 |
| Settled composer | Empty: J lost | `j`: J released to chat |
| Raw passive trace | [exit-before.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5893233159ffe188/exit-before.json) | [exit-after.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e30ce9882ab88195/exit-after.json) |

Do not claim the prior video proves a visible Stop 3: the recorded defect is lost composer input and a retained reservation marker during exit. Different J delays both fall inside the real 400 ms window; this is behavior evidence, not an animation pixel-equivalence comparison. Final-head production equals the fixed capture.

## Required guard completion

Before guard reel: [guards-before-51273020db.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9b5ec44d2f314250/guards-before-51273020db.mp4). Final-head reel: [guards-after-fee1f231f6.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/eca8724d42928ab7/guards-after-fee1f231f6.mp4). Logs: [guards-before-log.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/89eb248081631c43/guards-before-log.json), [guards-after-log.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d681427191a5086e/guards-after-log.json). Modifier reengagement/clear operations are visible and no text is submitted. The earlier exploratory after877/working-tree reels are superseded by the final-head reel.

| Matrix row | Current evidence / remaining boundary |
|---|---|
| 1 initial mount | Prior C1–C4 initial keys retained. Separate first mount with body focus remains unverified. |
| 2 stepping/clamping | Prior C1–C4 retained; new C5 final reengagement confirms current production stepping. |
| 3 chat handoff | C5 both revisions plus prior C1–C4. |
| 4 keyboard focus / editable | Final reel: Guide button → 15 Shift+Tab presses → Message textbox; j/k type `jk`, Stop 2 unchanged; clear and prose reengagement restores stepping. Before keyboard-only sequence not separately captured. |
| 5 panel boundary | Prior active PR header path retained; arbitrary standalone surface/chrome permutations remain unverified. |
| 6 modifiers | Both reels: Shift+J/K, Alt+J/K, Ctrl+J/K, Meta+J/K never move Stop 1. Shift types uppercase J/K in composer; Ctrl+K opens the real palette. Final log offsets 1.394, 3.965, 6.022, 7.677, 9.458, 11.586, 16.112, 18.173 seconds. |
| 7 competing overlay | Final wide real More pull request actions menu: j/k leave Stop 1 and composer unchanged; Escape + prose + j reaches Stop 2. Base comparison ends at Stop 2 with menu still open, demonstrating old routing differs; no claim that base passed overlay exclusion. Narrow after overlay remains prior C3. |
| 8 palette | Both reels: real Ctrl+K palette search gets `jk`; Stop 1 and composer unchanged. Escape and prose reengagement restore current guide keys. |
| 9 narrow sheet ancestor | Prior C3/C4 retained; no fresh narrow capture for the closing-only change. |
| 10 hidden/closed | Prior Summary/sheet sequences retained; new actual wide closing race directly exercises caller activity propagation. |
| 11 pending→arrival | COMPLETE C5 before/after, real cold query and chat click before completion. |
| 12 pending/null | Real pending spinner now captured C5; settled null remains prior C6. A supplemental pending attempt with Guide button focus did not insert j into composer and is excluded from claimed chat-typing proof; no guide reservation marker existed. |
| 13 no-guide run/error/permission | Unverified; no generation or permission changes invoked. Matrix identifies these as unchanged wrapper branches, not separate minimum captures. |
| 14 zero/one | Prior C6 retained. |
| 15 banner permutations | Prior populated no-banner guide retained. Stale/failed/running rewrite banners unverified. |
| 16 editable inside guide/callback | Prior Ask-about-stop context/composer check retained. No excerpt files or inline comment field in fixture; no new inside-guide editor evidence. |
| 17 model picker / terminal | Final reel: actual model search gets `jk`; Stop 1/composer unchanged, Escape and prose restore guide navigation. No terminal launched; terminal boundary unverified. |
| 18 IME / consumed events / 229 | Unverified actual IME, defaultPrevented, isComposing and legacy keyCode 229. No synthetic keyboard-event recording claimed. |

Final reel modifier portion ends around 20s; palette/menu/recovery is complete by wall 26.462s. Then keyboard transfer, editable j/k, clear/reengage, real model picker, and final recovery run through wall 52.408s. Pauses are retained in originals.

## Amplified diffs and comparability

For every pair, `D=abs(beforeRGB-afterRGB)` per channel using signed integer subtraction. Diff0: `min(8*D,255)` at zero threshold. DiffT16: retain that gain-8 image only where `max(D)>16`, otherwise black. Gain is 8, not the previous gain 1. No masking, alignment, scaling, gamma correction, or color normalization. Threshold is applied before amplification. Counts are from raw differences. Both are PNGs; original stills are unmodified video extracts.

C5 compares [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc)→[e8de9986ed](https://git.cafe/pingdotgg/t3code/commit/e8de9986ed52d2b241d5eef1c556e1fbe04d20ec), 2880×1800 frames from 1440×900 CSS, DPR 2, light. Same Chrome 154/Linux host, system sans stack and 16/14/13 interface/composer/code settings as the earlier capture. Date is not frozen in this pass. Sidebar closed, guide at top. The after frame has an incidental branch-changed banner and Merge control whereas before has no branch banner and Resolve conflicts. Host metadata is live; PR/guide head is not served code SHA. Full-frame diffs include these differences, cursor/caret, H.264 compression and intended stop/draft changes. This is not clean pixel equivalence. The older pairs keep all limitations from the original index, including C6 sidebar ages and null draft mismatch.

| Pair / revisions | Zero threshold, gain 8 | T16, gain 8 | Raw changed pixels T0 / T16 |
|---|---|---|---|
| C5, [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc)→[e8de9986ed](https://git.cafe/pingdotgg/t3code/commit/e8de9986ed52d2b241d5eef1c556e1fbe04d20ec) | ![C5-diff0-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/f653ca732a504d60/C5-diff0-gain8.png) | ![C5-diffT16-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8e2c14c0a72f5726/C5-diffT16-gain8.png) | 179,848 / 23,018 of 5,184,000 |
| C1, [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc)→[8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | ![C1-diff0-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/7f4b4f407a329829/C1-diff0-gain8.png) | ![C1-diffT16-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/fae5a75e70b5d35c/C1-diffT16-gain8.png) | 84,449 / 14,499 of 5,184,000 |
| C2, [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc)→[8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | ![C2-diff0-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/03680f9e17c5f80d/C2-diff0-gain8.png) | ![C2-diffT16-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/87dc7c3a0a1bdc15/C2-diffT16-gain8.png) | 256,498 / 31,274 of 5,184,000 |
| C3, [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc)→[8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | ![C3-diff0-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0280c5eb2fb1251a/C3-diff0-gain8-corrected.png) | ![C3-diffT16-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b0c7489da43afb32/C3-diffT16-gain8-corrected.png) | 157,882 / 3,793 of 1,316,640 |
| C4, [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc)→[8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | ![C4-diff0-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0244602a4f24ceed/C4-diff0-gain8-corrected.png) | ![C4-diffT16-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/75017d6bfda8ed05/C4-diffT16-gain8-corrected.png) | 45,406 / 3,177 of 1,316,640 |
| C6-null, [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc)→[8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | ![C6-null-diff0-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1816a02e052fc8c8/C6-null-diff0-gain8.png) | ![C6-null-diffT16-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/18000cd2f76c7236/C6-null-diffT16-gain8.png) | 191,983 / 15,297 of 5,184,000 |
| C6-zero, [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc)→[8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | ![C6-zero-diff0-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b11b9ac81b6fe475/C6-zero-diff0-gain8.png) | ![C6-zero-diffT16-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/578ea6c8030cfa81/C6-zero-diffT16-gain8.png) | 177,660 / 14,568 of 5,184,000 |
| C6-one, [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc)→[8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | ![C6-one-diff0-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/06bbf8fd7ac57961/C6-one-diff0-gain8.png) | ![C6-one-diffT16-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/066a7611097f1989/C6-one-diffT16-gain8.png) | 176,003 / 21,670 of 5,184,000 |
| C6-three, [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc)→[8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | ![C6-three-diff0-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/dbbafe04b3a962ec/C6-three-diff0-gain8.png) | ![C6-three-diffT16-gain8.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/64f32add9e2ae1c6/C6-three-diffT16-gain8.png) | 104,931 / 9,564 of 5,184,000 |

## Historical capture runtime handoff and teardown

All four owned dev-runner terminal sessions (77750, 7885, 2103, 34644) exited with Ctrl+C before source switches or teardown. Final port owners were server PID 855904 on 15360 and Vite PID 855891 on 7320; both `/proc` entries are gone. Final `ss -H -ltnp` has no listener on either port. Only the task's ignored `.t3` was served; live userdata was never opened read-write.

Owned tab `tab_e_f91513c6-0b52-4b7b-8605-0906569268b8` closed through T3 lifecycle. Final preview_status returns no tab/available false. All recordings stopped before closure. No other tabs or servers touched.

Checkout restored to `fix/issue-18300` at [fee1f231f6](https://git.cafe/pingdotgg/t3code/commit/fee1f231f67bd18feac0814c9cd34944cd48a9c3), clean `git status --porcelain`. Existing task SQLite remains three-stop fixture. Client-only panel duration was left at 400 ms in this owned origin; no production setting changed. No fixture change needed, so no test-agent commit.

Scratch: `/tmp/issue-18300-pass2/` contains scripts, downloaded/generated frames, all new selected recordings, JSON logs, URL manifests, discarded exploratory artifacts, and this index. Earlier scratch remains `/tmp/issue-18300-evidence/`. App-owned original recordings remain under `/home/ubuntu/.t3/userdata/attachments/` with IDs returned by T3; do not delete live userdata wholesale. [Capture metadata](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d1236d13819212a0/capture-metadata.json) records exact revisions and action observations. Public uploads are retained for review.

## FRICTION

| Friction | Resolution / limitation |
|---|---|
| Earlier C5 pause was too late | Query starts on PR open, not Guide selection. Hold the real send before first PR mount; verify held request plus actual spinner. |
| Panel state reset got repersisted | Remove only owned-origin panel storage and reload in the same evaluate call; clearing it and waiting lets the app write it back. |
| Browser API shape | evaluate rejects surrounding whitespace; use trimmed script. Modifiers use explicit modifier array. Invalid locator/aborted setup attempts were discarded. |
| Video final-paint lag | Initial C5 attempt ended before Stop 2 reached the recording; replaced with 3-second final hold and visually checked frame. Sparse native video PTS means wall action times are not exact video timestamps. |
| Source changed during exploratory guards | Final guard reel recaptured at [fee1f231f6](https://git.cafe/pingdotgg/t3code/commit/fee1f231f67bd18feac0814c9cd34944cd48a9c3). Production [e8de9986ed](https://git.cafe/pingdotgg/t3code/commit/e8de9986ed52d2b241d5eef1c556e1fbe04d20ec); earlier working-tree/after877 exploratory reels are superseded. |
| Closing before provenance | Compared prior fixed-core candidate [8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) against [e8de9986ed](https://git.cafe/pingdotgg/t3code/commit/e8de9986ed52d2b241d5eef1c556e1fbe04d20ec), not original base. Both trusted J delays are inside 400 ms but differ. |
| Live host metadata | Mergeability and branch warning differ; disclose rather than masking full-frame diffs. One transient GitCafe EAI_AGAIN occurred during restart; PR real-route capture subsequently loaded. |
| Gain 1 was not amplification | Recomputed all earlier source-frame differences at gain 8 plus raw T16 filtering. Old source revision labels retained. |

No full-suite checks were run by this capture agent. This deliverable is integrated real-client evidence, not backend/protocol correctness proof.

## Retained C1–C4/C6 originals and corrected comparison tables

## C1 — 1440×900, light

Approximate tool wall action offsets (NOT frame-accurate video PTS; pauses retained):


**Before**: 0.1s initial j; 0.9s initial k; 2.1s prose; 3.0s boundary j; 3.9s boundary j; 4.7s boundary j; 5.6s boundary k; 6.4s boundary k; 7.2s boundary k; 8.4s chat; 9.2s chat j; 10.1s clear select; 11.1s clear; 12.2s header; 13.2s header j; 14.3s composer; 15.2s editable j; 16.0s editable k; 16.9s clear select; 17.9s clear; 19.5s Summary; 21.5s chat; 22.5s hidden j; 23.4s clear select; 24.3s clear; 25.6s Guide; 26.5s reopen j; 27.3s reopen k.


**After**: 0.1s initial j; 1.2s initial k; 2.6s click prose; 3.7s boundary j; 4.8s boundary j; 5.9s boundary j; 7.1s boundary k; 8.2s boundary k; 9.2s boundary k; 10.6s chat handoff; 11.9s chat j; 13.0s clear draft select; 14.2s clear draft; 15.5s active PR header reengage; 16.6s header j; 28.0s composer focus; 29.1s editable j; 30.1s editable k; 31.1s clear select; 32.2s clear; 33.4s prose restores; 34.4s restored k; 35.5s ordinary a; 36.5s clear select; 37.6s clear; 38.9s prose reengage; 40.0s Shift J; 41.1s clear select; 42.1s clear; 43.4s prose reengage; 44.4s Alt j; 45.8s Summary hides guide; 47.1s chat transcript; 48.1s hidden j; 49.1s clear select; 50.2s clear; 51.6s Guide reopen; 52.6s reopen j; 53.6s reopen k; 64.5s palette j; 65.6s palette k; 66.8s dismiss palette; 68.4s reengage prose; 69.5s guide j after palette; 70.6s guide k; 72.1s Ask editor.


After additionally: ~58s Ctrl+K opens palette; 64.5s j, 65.6s k, 66.8s Escape; 68.4s prose; 69.5s j→Stop 2, 70.6s k→Stop 1; 72.1s Ask about this stop; ~74–80s composer j/k then Ctrl+A Backspace, cleared without sending. These final unlogged actions remain visible in the original clip.


| Before recording | After recording |
|---|---|
| [C1-before.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bfcd7d9c7a4f8d2c/C1-before.mp4) | [C1-after.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/44e7843675e2db61/C1-after.mp4) |


Comparison frames: before 3.6s, after 4.1s. These legacy seek requests select desktop before Stop 1 plus chat j and after Stop 2 plus empty chat; they are not exact decoded PTS.


| Before [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc) | After [8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | Diff0 (T=0, gain=8) | DiffT (T=16, gain=8) |
|---|---|---|---|
| ![C1-before.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1f7b6a5387b847af/C1-before.png) | ![C1-after.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/97d9252b96d5890c/C1-after.png) | ![C1-diff0.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/7f4b4f407a329829/C1-diff0-gain8.png) | ![C1-diffT.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/fae5a75e70b5d35c/C1-diffT16-gain8.png) |


Changed pixels: T0 84,449/5,184,000; T16 14,499/5,184,000.

## C2 — 1440×900, dark

Approximate tool wall action offsets (NOT frame-accurate video PTS; pauses retained):


**Before**: 0.1s initial j; 0.9s initial k; 2.1s prose; 3.0s boundary j; 3.9s boundary j; 4.7s boundary j; 5.6s boundary k; 6.4s boundary k; 7.2s boundary k; 8.4s chat; 9.3s chat j; 10.1s clear select; 11.1s clear; 12.3s header; 13.2s header j; 14.3s composer; 15.5s editable j; 16.4s editable k; 17.2s clear select; 18.1s clear; 19.4s Summary; 20.6s chat; 21.5s hidden j; 22.4s clear select; 23.3s clear; 24.6s Guide; 25.4s reopen j; 26.3s reopen k.


**After**: 14.2s initial j; 15.1s initial k; 16.2s prose; 17.1s boundary j; 18.0s boundary j; 18.9s boundary j; 20.1s boundary k; 21.0s boundary k; 21.8s boundary k; 23.0s chat; 24.0s chat j; 24.9s clear select; 25.8s clear; 27.0s header; 27.9s header j; 29.1s composer; 30.1s editable j; 31.0s editable k; 31.9s clear select; 32.8s clear; 34.2s Summary; 35.4s chat; 36.4s hidden j; 37.3s clear select; 38.2s clear; 39.7s Guide; 40.6s reopen j; 42.3s reopen k.


| Before recording | After recording |
|---|---|
| [C2-before.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bb32cbc80db21466/C2-before.mp4) | [C2-after.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9b30ab3223eb190a/C2-after.mp4) |


Comparison frames: before 3.6s, after 17.5s. These legacy seek requests select desktop before Stop 1 plus chat j and after Stop 2 plus empty chat; they are not exact decoded PTS.


| Before [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc) | After [8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | Diff0 (T=0, gain=8) | DiffT (T=16, gain=8) |
|---|---|---|---|
| ![C2-before.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/900b22dd021baaa1/C2-before.png) | ![C2-after.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/f4a3e8babb72a56c/C2-after.png) | ![C2-diff0.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/03680f9e17c5f80d/C2-diff0-gain8.png) | ![C2-diffT.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/87dc7c3a0a1bdc15/C2-diffT16-gain8.png) |


Changed pixels: T0 256,498/5,184,000; T16 31,274/5,184,000.

## C3 — 390×844, light

Approximate tool wall action offsets (NOT frame-accurate video PTS; pauses retained):


**Before**: 0.4s prose ancestor focus; 1.3s key j; 2.2s key k; 3.1s key j; 4.0s key j; 4.9s key j; 5.8s key k; 6.7s key k; 7.6s key k; 9.0s Summary hidden; 10.3s Guide restored; 11.2s restored j; 12.1s restored k; 13.1s close sheet; 14.2s chat transcript; 15.2s chat j; 16.1s clear select; 17.0s clear; 19.2s reopen sheet; 20.3s prose; 21.3s reopen j; 22.2s reopen k.


**After**: 13.2s prose ancestor focus; 14.2s key j; 15.2s key k; 16.2s key j; 17.2s key j; 18.1s key j; 19.1s key k; 20.1s key k; 21.1s key k; 22.7s Summary hidden; 24.7s Guide restored; 25.7s restored j; 26.7s restored k; 27.8s close sheet; 29.0s chat transcript; 30.1s chat j; 31.1s clear select; 32.1s clear; 33.7s reopen sheet; 34.9s prose; 35.9s reopen j; 36.9s reopen k.


After additionally: ~38s open More pull request actions; j/k while menu open preserve Stop 1 and draft; Escape, prose click, j/k recover navigation. The shell-sheet focus was read as sheet-popup after prose clicks.


| Before recording | After recording |
|---|---|
| [C3-before.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/964811bc5860cd30/C3-before.mp4) | [C3-after.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/35688def61683b41/C3-after.mp4) |


Comparison: exact decoded before frame 16 at PTS 1.427433 and after frame 17 at PTS 14.589900; both visibly Stop 2 after the first J following prose focus. These replace the old 1.65s before and rounded after seek extracts.


| Before [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc) | After [8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | Diff0 (T=0, gain=8) | DiffT (T=16, gain=8) |
|---|---|---|---|
| ![C3-before.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ee671f38f84831b5/C3-before-corrected.png) | ![C3-after.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c42eb871e4fa609e/C3-after-corrected.png) | ![C3-diff0.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0280c5eb2fb1251a/C3-diff0-gain8-corrected.png) | ![C3-diffT.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b0c7489da43afb32/C3-diffT16-gain8-corrected.png) |


Changed pixels: T0 157,882/1,316,640; T16 3,793/1,316,640.

## C4 — 390×844, dark

Approximate tool wall action offsets (NOT frame-accurate video PTS; pauses retained):


**Before**: 0.4s prose ancestor focus; 1.3s key j; 2.4s key k; 3.4s key j; 4.3s key j; 5.2s key j; 6.1s key k; 7.0s key k; 7.9s key k; 9.9s Summary hidden; 11.3s Guide restored; 12.2s restored j; 13.1s restored k; 14.1s close sheet; 15.2s chat transcript; 16.2s chat j; 17.2s clear select; 18.2s clear; 19.8s reopen sheet; 20.9s prose; 21.8s reopen j; 22.8s reopen k.


**After**: 0.3s prose ancestor focus; 1.3s key j; 2.3s key k; 3.3s key j; 4.3s key j; 5.3s key j; 6.2s key k; 7.2s key k; 8.2s key k; 9.9s Summary hidden; 11.4s Guide restored; 12.5s restored j; 13.5s restored k; 14.5s close sheet; 15.7s chat transcript; 16.9s chat j; 17.9s clear select; 18.9s clear; 20.5s reopen sheet; 21.7s prose; 22.8s reopen j; 23.8s reopen k.


| Before recording | After recording |
|---|---|
| [C4-before.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d2636973c56fa81d/C4-before.mp4) | [C4-after.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/071063eaba728dd7/C4-after.mp4) |


Comparison: exact decoded before frame 16 at PTS 1.427400 and after frame 17 at PTS 1.513233; both visibly Stop 2 after the first J following prose focus. These replace the old 1.65s before and rounded after seek extracts.


| Before [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc) | After [8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | Diff0 (T=0, gain=8) | DiffT (T=16, gain=8) |
|---|---|---|---|
| ![C4-before.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/457c2b259bba80be/C4-before-corrected.png) | ![C4-after.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a1859b75023e7078/C4-after-corrected.png) | ![C4-diff0.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0244602a4f24ceed/C4-diff0-gain8-corrected.png) | ![C4-diffT.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/75017d6bfda8ed05/C4-diffT16-gain8-corrected.png) |


Changed pixels: T0 45,406/1,316,640; T16 3,177/1,316,640.

## C6 — 1440×900, light; labeled fixture-state reels

Each segment begins only after stopping the server, seeding the named state, restarting and reloading. These edits are not product transitions. In each segment: click heading/prose, j, k, then clear composer using click/Ctrl+A/Backspace. The first null-before segment retains jj from a discarded C5 attempt, so its j/k append to that draft; this is an explicit fixture-reset limitation. Subsequent segments start empty.


**Before segment timestamps**: 0.00s null (6.76s); 6.76s zero (7.89s); 14.65s one (6.83s); 21.48s three (6.75s).


**After segment timestamps**: 0.00s null (7.27s); 7.27s zero (6.92s); 14.19s one (7.23s); 21.42s three (6.74s).


| Before recording | After recording |
|---|---|
| [C6-before.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/7055471ad7f520b3/C6-before.mp4) | [C6-after.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9d09bf1c341b43ee/C6-after.mp4) |


### Null segment

Comparison frames at 4.4s within each original segment; j/k are paced approximately 1.6s apart. Null/zero/one keep the guide unchanged and route printable input to chat. Three-stop before stays Stop 1 and types jk; after steps 1→2→1 and leaves chat empty.


| Before [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc) | After [8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | Diff0 (T=0, gain=8) | DiffT (T=16, gain=8) |
|---|---|---|---|
| ![C6-before-null.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3a4faf57bb75949c/C6-before-null.png) | ![C6-after-null.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/57286aba595ad9bb/C6-after-null.png) | ![C6-null-diff0.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1816a02e052fc8c8/C6-null-diff0-gain8.png) | ![C6-null-diffT.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/18000cd2f76c7236/C6-null-diffT16-gain8.png) |


Changed pixels: T0 191,983/5,184,000; T16 15,297/5,184,000.


### Zero segment

Comparison frames at 4.4s within each original segment; j/k are paced approximately 1.6s apart. Null/zero/one keep the guide unchanged and route printable input to chat. Three-stop before stays Stop 1 and types jk; after steps 1→2→1 and leaves chat empty.


| Before [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc) | After [8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | Diff0 (T=0, gain=8) | DiffT (T=16, gain=8) |
|---|---|---|---|
| ![C6-before-zero.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6498887bf02e14bf/C6-before-zero.png) | ![C6-after-zero.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a709b7a844313bb5/C6-after-zero.png) | ![C6-zero-diff0.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b11b9ac81b6fe475/C6-zero-diff0-gain8.png) | ![C6-zero-diffT.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/578ea6c8030cfa81/C6-zero-diffT16-gain8.png) |


Changed pixels: T0 177,660/5,184,000; T16 14,568/5,184,000.


### One segment

Comparison frames at 4.4s within each original segment; j/k are paced approximately 1.6s apart. Null/zero/one keep the guide unchanged and route printable input to chat. Three-stop before stays Stop 1 and types jk; after steps 1→2→1 and leaves chat empty.


| Before [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc) | After [8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | Diff0 (T=0, gain=8) | DiffT (T=16, gain=8) |
|---|---|---|---|
| ![C6-before-one.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d291345f3187e1b3/C6-before-one.png) | ![C6-after-one.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c114d0623ce6560b/C6-after-one.png) | ![C6-one-diff0.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/06bbf8fd7ac57961/C6-one-diff0-gain8.png) | ![C6-one-diffT.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/066a7611097f1989/C6-one-diffT16-gain8.png) |


Changed pixels: T0 176,003/5,184,000; T16 21,670/5,184,000.


### Three segment

Comparison frames at 4.4s within each original segment; j/k are paced approximately 1.6s apart. Null/zero/one keep the guide unchanged and route printable input to chat. Three-stop before stays Stop 1 and types jk; after steps 1→2→1 and leaves chat empty.


| Before [51273020db](https://git.cafe/pingdotgg/t3code/commit/51273020db3245f4ec29514a419d29b5dcb012dc) | After [8772272735](https://git.cafe/pingdotgg/t3code/commit/8772272735c47b7bd912eaca918315052b1b16e3) | Diff0 (T=0, gain=8) | DiffT (T=16, gain=8) |
|---|---|---|---|
| ![C6-before-three.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/34325cbd32e9a41f/C6-before-three.png) | ![C6-after-three.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/425c6d8ac0e0b545/C6-after-three.png) | ![C6-three-diff0.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/dbbafe04b3a962ec/C6-three-diff0-gain8.png) | ![C6-three-diffT.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/64f32add9e2ae1c6/C6-three-diffT16-gain8.png) |


Changed pixels: T0 104,931/5,184,000; T16 9,564/5,184,000.

## What else changed in the comparisons

- Real PR host mergeability refreshed independently: Merge versus Resolve conflicts. Its metadata branch, file counts and stored guide head are unrelated to local code provenance.
- C2/C6 after can show the existing Branch changed from main warning because the source checkout is on fix/issue-18300; the detached before revision can suppress it. It is outside the guide.
- C6 sidebar relative ages were rendered before Date freezing and differ. The first null-before draft starts with jj, as disclosed above.
- Cursor location, editable focus, button focus rings, placeholder state, H.264 compression and guide position/draft after the tested key can differ. Diffs retain all of them; no broad masking hides them.
- Preview snapshot cached/downsampled stills were rejected. All published stills were extracted from original video, at full DPR resolution.
- No production fixes. Only reproducible seed.py, fixed fixture data.json and usage README were committed. No PR-only captures were committed.

## Scope and verification

Web inline panel and narrow web sheet were captured. Desktop reuses web but was not independently captured in Electron. Native mobile, provider runs, RPC behavior changes, remote/relay/tunnel modes and scheduled tasks are outside this routing diff. Source Control settings gate was enabled in the task home. Command palette was exercised only as an input-owner boundary.

Seed script: syntax compiled; running-server write refusal checked; three-stop seeding and bootstrap applied twice with one fixture thread and three stops verified. Loops cover two fixed bootstrap rows and their fixed columns, O(1) relative to repository/database size; stops bounded at three; argparse choices only. No repo-wide checks or browser harness substituted for recordings.

## Correction-run FRICTION and preservation

- Sparse PTS and next-frame seeking caused both old narrow sides to select Stop 1. Fixed by frame-index extraction and visual verification of matching first-post-J Stop 2 states; all original recordings retained.
- The checkout was not the reported clean candidate when this correction began: HEAD was `07f8aee277e9f6fdb76c0f27af6c39350a2f9214` and untracked `app.json` existed. No checkout switch, cleanup, source/test/fixture edit or commit was performed. Historical capture teardown claims above describe those capture runs, not this correction run.
- All correction work stayed in `/tmp/issue-18300-corrected`; no browser/server started, no push or forge write, no new capture, no test run. The four original tests and fixtures were left untouched.
- Remaining evidence gaps are explicitly preserved in the current row matrix: before keyboard-only transfer and actual IME/consumed-event/legacy-229 remain unverified. Correcting artifacts does not fill them.
- Original public URLs are immutable. This replacement index supersedes inaccurate claims but does not delete historical uploads.