# Evidence capture — round 2, corrected; round 3 X03 addendum

Historical round-2 assigned rows: **M09 and X03 only; 2/2 capture sets delivered**, with the Android timing qualification below. Pass 2b found one missing X03 external-live-sync sub-row; the round-3 addendum at the end supplies it. This is evidence collection, not a merge-readiness verdict. No previously accepted row was rerun.

For [#32816](https://git.cafe/pingdotgg/t3code/pull/32816), before is [20c3dd3](https://git.cafe/pingdotgg/t3code/commit/20c3dd3689782bfcb1a04332086d08017028c718) and after is [c94024d](https://git.cafe/pingdotgg/t3code/commit/c94024dfa563d3a722d1f21ddcb24a9bca08419b). Both were real apps built from their pinned checkouts, with independent fresh databases and `T3CODE_DEV_SEED=0`. All fixture turns used Replay (dev). A second real isolated after server supplied environment B. No live database, copied database, projection substitution, paid turn, or isolated component harness was used.

## M09 — tablet Files/split wrapper

**Android ordered steps:** open `ios-fork` with Files alongside it and the Git menu open; activate Merge back to source; await its response; reopen Git controls; hold the final menu for 10 seconds wall time. The retained after begins selected, cancels, stays on source with Files open, and reopens ordinary. Baseline begins ordinary, repeats merge, navigates to its parent, closes Files, and reopens the parent menu without a merge item. This different destination is expected baseline behavior, not an attempt to fabricate a selected baseline.

**Timing qualification:** Android gesture telemetry places the baseline activation at **4.062 s**, rather than the requested approximately 2 s, because of gateway latency; Git reopens at **19.095 s**. The action order and initial viewport/Files layout match. Its raw encoded duration is 21.066 s; Android omits the unchanged tail despite a 10 s final wall hold. Inspect the true last encoded frame for the reopened parent menu. This is not an exact timing match.

| M09 · Pixel Tablet · 1280×800 logical · DPR 2 · dark | Before | After |
|---|---|---|
| Whole-surface recording, 2560×1600 physical | [M09-before-tablet-files-final.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b381f28d30c35b87/M09-before-tablet-files-final.mp4)<br>[20c3dd3](https://git.cafe/pingdotgg/t3code/commit/20c3dd3689782bfcb1a04332086d08017028c718) | [M09-after-tablet-matched-dark.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/13c281fd59f18292/M09-after-tablet-matched-dark.mp4)<br>[c94024d](https://git.cafe/pingdotgg/t3code/commit/c94024dfa563d3a722d1f21ddcb24a9bca08419b) — retained from the accepted after-flow, not newly captured |

**iPad steps and timestamps:** true landscape Files split, `ios-fork`, Git menu initially open. Baseline activates the ordinary item at **2.568 s**, goes to parent and closes Files, then reopens Git at **11.150 s**. After activates the selected item at **2.613 s**, cancels without leaving the source/inspector, then reopens Git at **10.818 s** with the ordinary item. Each final menu is held for 9 s wall time; clips are approximately 23 s. The iPad exists in the pool and was leased successfully.

| M09 · iPad Pro 13-inch (M5) · landscape 1376×1032 logical · DPR 2 · dark | Before | After |
|---|---|---|
| Files/split recording, 2752×2064 physical | [M09-before-ipad-files-final.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/008c2754d4a6cabe/M09-before-ipad-files-final.mp4)<br>[20c3dd3](https://git.cafe/pingdotgg/t3code/commit/20c3dd3689782bfcb1a04332086d08017028c718) | [M09-after-ipad-files-final.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b52335a9135d7bc0/M09-after-ipad-files-final.mp4)<br>[c94024d](https://git.cafe/pingdotgg/t3code/commit/c94024dfa563d3a722d1f21ddcb24a9bca08419b) |

The iPad's AX tree incorrectly calls the thread toolbar covered while Files is open. After inspecting the screen, Git was opened at logical (732,54); the item itself was selected semantically. After cancellation it reopened at that same anchor. Baseline navigation closes Files, so its final menu is reopened semantically at the wider thread toolbar.

**What else changed:** baseline navigation changes the title, selected sidebar row, message history/lineage, content width, and inspector visibility. After keeps those source elements. Selected items add a check and cancel subtitle; wrapping and popup height change when the ordinary subtitle returns. iPad titles are normally ellipsized in the narrow Files column on both sides. Native OS clocks differ; server message times reflect device timezone. Android retained after was recorded earlier, so sidebar ages differ. The workspace is intentionally empty, and Files displays that state. No new static before/after pair or pixel diff is asserted; the prior audit waived unstable native paired stills.

## X03 — reconnect and environment isolation

A and B have identical `proof-fork`/`proof-parent` thread ids in separate real servers. **Only A** has a pending transfer. Both browser clients and the phone are paired to B through the real onboarding service. Switching uses the application's real router/navigation APIs; the environment picker UI itself is not being tested.

**Web steps, same order on both revisions:** show pending A; route to B and expose its ordinary action; return to A; stop the owned A dev runner and restart against the identical home; wait for actual connected state; show A's retained pending state; activate the action. After cancels and stays on the source with no pending transfer. Baseline has no selected styling and repeats merge, navigating parent.

**Video chapters (approximate encoded positions):** after A selected at **0.000/0.5044 s**, B first encoded at **15.1859 s** and readable at **15.4631–16.5151 s**, A selected again at **16.5151 s**, deliberate restart/splash beginning at **21.86 s**, reconnect content returning at **35.44 s**, and the final encoded frame is unselected on source. Before B is visible around **5 s**, A returns around **10 s**, restart/splash around **23–34 s**, A pending again around **35–42 s**, and activation navigates parent around **43–45 s**. The reconnect splash is the exercised lifecycle transition, not a settled endpoint. The recordings preserve their original variable frame timing; static tails can end at their last encoded change.

| X03 · Chromium 157/Linux · 1440×900 CSS · DPR 2 · dark | Before | After |
|---|---|---|
| Environment switch + restart + activation, 2880×1800 physical | [X03-before-web.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b581b798c7fbc0ca/X03-before-web.mp4)<br>[20c3dd3](https://git.cafe/pingdotgg/t3code/commit/20c3dd3689782bfcb1a04332086d08017028c718) | [X03-after-web.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/564a310be3c81de5/X03-after-web.mp4)<br>[c94024d](https://git.cafe/pingdotgg/t3code/commit/c94024dfa563d3a722d1f21ddcb24a9bca08419b) |

Actual observed reconnect responses: after at 17:59:30.432 UTC was `connected`, one pending transfer, `aria-pressed=true`; following activation it was false with zero pending and route still `proof-fork`. Before at 18:00:28.752 UTC was `connected`, one pending transfer, no `aria-pressed` attribute; activation selected `proof-parent`. The observation JSON is reconstructed from those tool responses and is explicitly not a continuous log.

**Native phone chapters:** Pixel 9, light. First recording: A selected at **0 s**, dismiss at **11.950 s**, B ordinary menu at **30.576 s**, dismiss at **42.322 s**, return to A and reopen selected at **63.964 s**. Second: starts selected; the actual server stop/restart shows “Reconnecting to fs-87” around **10–27 s**; dismiss at **32.209 s**, reopen selected after reconnection at **40.816 s**, cancel at **62.948 s**, reopen ordinary at **71.478 s**, remain on source. The combined file losslessly joins these two video streams at **71.498 s**; this recording boundary is disclosed, and both originals remain linked below. It contains no synthesized frames or overlays.

| X03 · Pixel 9 · 411.428571×923.428571 logical · DPR 2.625 · light | Before | After |
|---|---|---|
| 1080×2424 physical, actual Android API 36 app | Optional baseline not captured | [X03-after-native-combined.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bb75e0647e3c926c/X03-after-native-combined.mp4)<br>[c94024d](https://git.cafe/pingdotgg/t3code/commit/c94024dfa563d3a722d1f21ddcb24a9bca08419b) |

Original native chapters:

[X03-after-native-phone.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6bf72c70eedcb778/X03-after-native-phone.mp4)

[X03-after-native-reconnect.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e833c37e05c7cdd8/X03-after-native-reconnect.mp4)

**What else changed:** environment names, sidebar grouping/order/ages and source title distinguish A from B. Restart briefly shows the web splash or native reconnect notice. Selected styling, glyph, tooltip/subtitle and action feedback change with state; baseline activation changes destination and history. Browser details also show existing “Open in VS Code”/“Open in Files” preference text differences. The phone source title is ellipsized in its narrow header. Native OS clocks and server message times remain live. No static pixel comparison is claimed.

## Rendering, drivers and inspection

Web: Chromium 157, Linux x86_64; no device emulation; 16px default root font, 100% zoom, DPR 2. Native: actual Simlock Pixel Tablet and Pixel 9 on API 36 and iPad Pro 13-inch (M5) on iOS 27.2; **fontScale 1 on all three**. Logical dimensions/DPR came from the running app's React Native Dimensions/PixelRatio through agent-device CDP. iPad still screenshots from the tool are downscaled; the recordings retain full physical size. Exact geometry is in the JSON below.

Application JS Date was frozen to 2026-11-10T12:10Z. Native status-bar clocks and backend-generated timestamps are unaffected. Web Vite restart reloads the page and briefly restores the real clock before init is rerun, changing age labels. These are interaction recordings, not pixel-stable still pairs. There are no Diff0/DiffT columns because no new paired still claim is made.

Every published recording was opened through decoded whole-surface sequences and its actual last encoded frame. Native popup endpoints require that final-frame check because VFR samplers can miss them. The combined phone join was separately inspected. Superseded exploratory attempts are not offered as proof. No automated repository tests or CI/review certification was performed for this capture-only task.

[capture-drivers-round2-final.zip](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/432a63f95db4a231/capture-drivers-round2-final.zip)
[round2-execution-notes.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/4a1a3e1538e2db70/round2-execution-notes.md)
[capture-metadata.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/84b83dfef47af327/capture-metadata.json)


TRAN[capture-drivers-round2-final.zip](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/432a63f95db4a231/capture-drivers-round2-final.zip)
[round2-execution-notes.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/4a1a3e1538e2db70/round2-execution-notes.md)
[capture-metadata.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/84b83dfef47af327/capture-metadata.json)


[native-gesture-timings-public.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2c9ddcb787dd48a7/native-gesture-timings-public.json)
![M09-ipad-files-after-end.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1bc02a9d4c139ecd/M09-ipad-files-after-end.png)


The PNG above is an unpaired final frame extracted from the iPad after recording, labelled **After cancellation — source and Files preserved**. It is not an independently captured still comparison.

## Scope and cleanup

Round 2 visited both assigned capture sets, but the subsequent audit identified the external live merge/cancel sub-row as missing. Round 3 adds that sequence below; retained footage is unchanged. Qualifications: Android baseline timing is not exact; native phone before is optional and absent; native environment switching uses navigation APIs, not the picker UI. The retained Android after was not re-recorded. Remaining rows and previous evidence were left alone.

All three backend/web runners and both Metros stopped through their tracked sessions. No listeners remained on 5775/13815, 5939/13979, 6733/14773, 8081 or 8082. All three owned browser tabs closed. iPad close succeeded; both Android close attempts reported the leases already inactive, then both sessions disconnected successfully. The local CDP session had no remote lease. The created baseline worktree was removed; the supplied after checkout remains clean at [c94024d](https://git.cafe/pingdotgg/t3code/commit/c94024dfa563d3a722d1f21ddcb24a9bca08419b). Scratch evidence and private logs remain in the assigned directory; no raw evidence was deleted. Disk available: 238 GB.

AI-assisted; verified with the tests and tools listed above.

FRICTION:

tool | Device gateway intermittently returned Remote daemon unavailable; Android leases expired during early uploads and were recovered before capture | More reliable gateway/lease renewal during uploads would avoid setup loss.

tool | Native CDP target selection through tailnet gateway returned 401 | Local Metro CDP through a separate named session worked; document that supported route.

tool | iPad AX falsely marked the thread toolbar covered by a fullscreen toolbar | Screenshot-guided coordinate fallback at the visible anchor was required.

tool | Bounded batch rejected record:start despite listing record support | Explicitly document that recording must be a top-level command.

tool | Android VFR omitted unchanged tails and ordinary sampling missed popup endpoints | Inspect true final encoded frames and retain gesture/stop telemetry.

tool | Android command latency moved the intended approximately 2 s tap to 4.062 s | Server-side timed device actions compatible with recording would provide exact timing.

tool | Browser tool writes generated media to its own artifact area | An explicit output directory would keep everything directly under assigned scratch; only owned media was copied.

ROWS: 2/2 capture sets; M09 timing qualification; X03 native before optional/not captured.

## Round 3 addendum — X03 live external merge/cancel

**Assigned/captured: 1/1 sub-row.** This closes the specific external-client sequence identified by pass 2b. Earlier accepted coverage is retained above; no other fixture-matrix row was recaptured. The pass-2b timing correction is applied to the round-2 web chapter paragraph above. Any older execution note saying B appears around 3 s is superseded by that corrected paragraph.

**Setup and action order:** observer A is already mounted on `proof-fork`, with its Lineage control exposed and ordinary. B is a separately paired incognito browser context using the same server and synthetic thread IDs. B issues the actual mergeBack command; A is untouched. B then issues cancelMergeBack on the after revision; A remains untouched. No click, navigation, reload, DOM styling, subscription replacement, or projection mutation occurs on A during either recording. Its read-only click observer records zero clicks. B alone runs the real commands through the application's command layer and authenticated transport. The B tab is not included in these recordings, as requested.

**After web, actual encoded frames:** A is ordinary at **0.000000 s**; B's merge causes the first selected frame at **10.738233 s**; B's cancel causes the first ordinary frame at **23.544400 s**; A is still on source in the final encoded frame at **39.344833 s**. Video duration **39.378133 s**. The wall-clock command responses were 18:45:04.874 UTC (merge, sequence 162) and 18:45:17.591 UTC (cancel, sequence 163). A's actual aria-pressed mutations followed at 18:45:04.882 and 18:45:17.596 UTC. Wall timestamps are separate from the decoded media positions.

**Before web:** A is ordinary at **0.000000 s** and remains ordinary through the final encoded frame at **29.330800 s**; video duration **29.364100 s**. B's successful merge response is logged at **18:45:33.375 UTC**, sequence 162, during the recording window (18:45:18.919–18:45:47.563 UTC). A's parent projection then contains a real pending transfer from `proof-fork`, while its control has no selected appearance or aria-pressed attribute. There is no visible transition to assign a fabricated media timestamp to. The baseline has **no cancelMergeBack command**, so B cannot cancel on that side; this is the expected before limitation. The accompanying command transcript establishes the external action while the video establishes the unchanged observer.

| X03 external live sync · Chromium 157/Linux · 1440×900 CSS · DPR 2 · dark · no device emulation | Before | After |
|---|---|---|
| Whole-surface Browser panel recording · 2880×1800 physical | [X03-external-before-web.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/437eb17978a6d362/X03-external-before-web.mp4)<br>[20c3dd3](https://git.cafe/pingdotgg/t3code/commit/20c3dd3689782bfcb1a04332086d08017028c718) | [X03-external-after-web.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e043de36cc434761/X03-external-after-web.mp4)<br>[c94024d](https://git.cafe/pingdotgg/t3code/commit/c94024dfa563d3a722d1f21ddcb24a9bca08419b) |

The following is an **After** excerpt from actual encoded frames: whole-surface thumbnails plus additional Lineage detail crops. Labels carry the original frame timestamps; no video frame was altered or synthesized.

![X03-external-web-encoded-states.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b5353d5b2c0bbdaa/X03-external-web-encoded-states.png)

**Native observer:** an actual Pixel 9 app is paired to the same after backend. Its Git menu starts open and stays open throughout, with no tap, dismissal, navigation or reload during the recording. External B issues a new merge (18:47:26.056 UTC, sequence 165), then a cancel (18:47:58.902 UTC, sequence 166). Ordinary subtitle is visible in encoded frames at **0.000000, 0.722444 and 15.223967 s**; the selected subtitle “Merged back. Tap to cancel” first appears at **15.256589 s**, and remains through **48.074100 s**; the ordinary subtitle returns at **48.105622 s** and remains in the actual last frame at **50.723478 s**. Native capture duration **51.791422 s**, nine encoded frames. Android omits the unchanged tail of the **64.8 s wall recording window**. All nine frames were decoded and visually opened; the selected and cancelled boundary frames were also viewed at full surface.

| X03 external live sync · Pixel 9 · actual Android API 36 · 411.428571×923.428571 logical · DPR 2.625 · fontScale 1 · light | After |
|---|---|
| Whole-device 1080×2424 physical, source and popup remain mounted | [X03-external-after-native.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/95175eff81287daf/X03-external-after-native.mp4)<br>[c94024d](https://git.cafe/pingdotgg/t3code/commit/c94024dfa563d3a722d1f21ddcb24a9bca08419b) |

![X03-external-native-encoded-states.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/7c5b64fc97e9ea4d/X03-external-native-encoded-states.png)

**What else changed:** the web control turns blue on pending and ordinary after cancellation; the conversation, route, scroll position, Lineage row and sidebars remain in place. Native subtitle wrapping changes menu row/popup height and moves the More row; source content and navigation stay in place. The OS minute changes from 18:47 to 18:48. Source titles ellipsize normally. Backend message times are live and differ between baseline and after, while both browser/app JS clocks were frozen at 2026-11-10T12:10Z. Web composer caret blinks; native OS clock stays live. Before's aria-pressed attribute is absent rather than false. The synthetic workspace is intentionally not a git repository, shown consistently by native Git controls. No additional ghost elements, stale overlay, spinner, skeleton, wrong viewport or unexpected clipping was observed. These are interaction recordings, not pixel-matched still pairs; the figures are video excerpts, so no Diff0/DiffT pair is claimed.

**Rendering environment:** web Chromium 157 on Linux x86_64, Browser panel, 1440×900 CSS, DPR 2, 2880×1800 encoded, dark, no emulation, 16px root font, 100% zoom. Native Simlock Pixel 9/API 36, actual Android dev client and after-checkout Metro, dimensions/DPR/fontScale above, light, no browser emulation. Real apps at their pinned revisions; fresh isolated homes with T3CODE_DEV_SEED=0 and the same seed driver; Replay (dev) only. All credentials and private runtime operational identifiers stay out of uploads.

**Inspection:** all web encoded frames were decoded using passthrough timing, then opened as whole-surface sequences. Actual initial, selected, cancelled and final full frames were inspected separately. Native all-frame sequence, selected/cancelled boundary full frames and final encoded endpoint were inspected. No trim, video overlay, interpolation or fabricated frames. The media checksums and exact encoded frame timestamp lists are uploaded below.

[capture-drivers-round3.zip](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/094a651cf295299a/capture-drivers-round3.zip)
[capture3-metadata.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/03b6ae54d0e565d2/capture3-metadata.json)
[round3-execution-notes.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/362212752d4c6f26/round3-execution-notes.md)

The driver bundle includes every seed/observer/command/native navigation/inspection script used or written in this round, plus the exact command/observer transcript and encoded frame lists. It excludes pairing credentials and raw private logs. No automated repository tests, CI certification, code changes, commits, pushes, PR edits or paid provider turns were performed.

**Limits:** no assigned sub-row remains missing. No new iOS or native-before claim; the narrow supplemental request requires the web before/after pair and one native source observing the external sequence. The native supplemental recording is light, as in the retained accepted round-2 phone footage. Earlier accepted reconnect/isolation footage remains unchanged. This is evidence delivery, not an independent merge-readiness audit.

**Cleanup:** both backend/web runners and Metro were stopped through captured terminal sessions. No listeners remained on 5775/13815/5939/13979/8081. Four owned browser tabs closed; Pixel 9 lease closed and proxy disconnected; local CDP had no remote lease. Created baseline worktree removed. Supplied after checkout remains clean. No raw evidence deleted. Disk available **239 GB**. Finished within the 35-minute budget.

AI-assisted; verified with the tests and tools listed above.

FRICTION:

tool | preview_open rejected about:blank; omission opened a blank tab | Document blank-tab syntax.

tool | preview_evaluate rejects surrounding whitespace | Trim script input.

tool | Native AX snapshot raced installation | Wait for install completion; retry then succeeded.

tool | pair --tailscale attempted unavailable HTTPS serve | Direct tailnet backend port worked; no manual serve changes.

tool | Android omits unchanged tail | Inspect actual encoded frames and disclose wall/encoded durations.

ROWS: 1/1 assigned external-live-sync sub-row captured.
UNVERIFIED: none within assigned sub-row; no broader audit verdict claimed.
