# Fourth capture evidence for [#32936](https://git.cafe/pingdotgg/t3code/pull/32936)
**Rows done: 2 / 2, using I2’s explicitly allowed “before control unreachable” fallback.** F5 includes the optional detail-column companion. This report is capture evidence, not a merge recommendation.
Before [ca796d5211](https://git.cafe/pingdotgg/t3code/commit/ca796d52119592de771d8ccfa955d70dba84d11e); after [f76e21be6c](https://git.cafe/pingdotgg/t3code/commit/f76e21be6cdcf0936a28521badaf66e4a85bb9cc). Environment: fs-86 Linux VM; two independently leased **iPad (7th generation), iOS 18, landscape-right, light**, default system text size **large**. Screenshots are 1080×810 at 1×. The OS minor version was not recorded. Both cached native clients have fingerprint `5fa891c99bf29870cc7d56112e3458d95a948393`. Before/after run their own Metro bundles (18081/18082), fresh independent server homes (15228/14456), and transparent proxies (17057/17058). Every server was started with **T3CODE_DEV_SEED=0**. All content is synthetic Replay (dev) data created through RPC. No live database was read, copied, seeded, or modified.
Every selected still has two raw captures at least 500 ms apart with **zero full-frame pixel differences**, including the clock. Seven selected stills are covered by the [stability manifest](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/91ce69fbcb3b1cbf/stability-manifest.json) and [raw pairs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/39b05f6371cdb244/raw-stability-pairs.zip). All published stills, both variants of each diff, and the recording contact sheet were opened and visually inspected before upload. No selected frame exposes an agent-card model label: the parent’s Worked block was collapsed before F5, and I2 shows a finished-task link. No image or video redaction was necessary. Private configuration, credentials, unselected diagnostics and the failed preliminary captures are excluded.
Diff0 is ImageMagick compare at **0% fuzz**, amplified binary magenta (#ff00ff) changed pixels against #151515. DiffT uses **5% fuzz**. Both use `-compose src`, complete frames, no resampling, no masking and no crop. Their diffs retain incidental clock, runtime-duration and edge-rendering changes.
## Row 1 — F5 unavailable placeholder — iPad 7 / iOS 18 / landscape / light
**Exact state in both frames:** an uncached synthetic child opened while its environment is disabled, rendered by `contentPresentation.kind === "unavailable"`: **“Messages not cached”**, with **“Reconnect this environment to load the conversation.”** wrapping beneath it. This is the ThreadFeedPlaceholder path, not the separate “Thread unavailable / current mobile snapshot” fallback.
The right inspector is 260 pt wide (x=820…1080); its solid header ends at y=70. Before, the title begins around y=290, centred using the column area including the covered header. After, it begins around y=325, a **35 pt downward shift**, centring the block in the area below the header. The block is pixel-identical after translating it down 35 pt: before crop `220x180+840+250` versus after `220x180+840+285` compare at 0 differing pixels. [Measurement transcript](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/103626867e221504/placeholder-measurements.json).
**Trigger, identical on each side:** create a separate parent/child with [offline-fixture.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8d25a0a7f1b76bac/offline-fixture.mjs); open only the parent while connected; turn off **Settings → Environments → fs-86 → Enable fs-86**; reopen the cached parent; expand Worked and open **Inspect uncached conversation** in the inspector; collapse Worked; capture without focusing a composer. The child detail had never been opened while connected. The environment’s disconnected notice is real app UI.
| Before | After | Diff0 | DiffT (5% fuzz) |
|---|---|---|---|
| 
[ca796d5211](https://git.cafe/pingdotgg/t3code/commit/ca796d52119592de771d8ccfa955d70dba84d11e)
SHA-256: a4acf233424c9e277b1ab8a3dd057d924efa0547f8ffb73a2b061a7baf45bc32 | 
[f76e21be6c](https://git.cafe/pingdotgg/t3code/commit/f76e21be6cdcf0936a28521badaf66e4a85bb9cc)
SHA-256: 5b9efeea5b1157fa2282d652e3eb563f376d6e528165a0cb2d4544c5c654016a | 
[ca796d5211](https://git.cafe/pingdotgg/t3code/commit/ca796d52119592de771d8ccfa955d70dba84d11e) → [f76e21be6c](https://git.cafe/pingdotgg/t3code/commit/f76e21be6cdcf0936a28521badaf66e4a85bb9cc)
SHA-256: 6aa332b0be6ce9b17dad4685a0e36193ba0b9d55ca03cb91fac2494b01c5942a | 
[ca796d5211](https://git.cafe/pingdotgg/t3code/commit/ca796d52119592de771d8ccfa955d70dba84d11e) → [f76e21be6c](https://git.cafe/pingdotgg/t3code/commit/f76e21be6cdcf0936a28521badaf66e4a85bb9cc)
SHA-256: c88a5af787ecd50e4002a347d7045bdfc6ae873baa8a56a0368172d74add15b9 |
**Optional detail-column companion:** open the same uncached child’s thread deep link while still disconnected. Same placeholder, now in the 735 pt detail column. The title starts around y=312 before and y=347 after, also **35 pt lower**; translated placeholder crops compare at 0 differing pixels (`500x160+470+270` → `500x160+470+305`). The before sidebar’s “No threads yet” label is hidden under its own header, while after it is visible; this is an additional inset change in the same PR, not different data.
| Before | After | Diff0 | DiffT (5% fuzz) |
|---|---|---|---|
| 
[ca796d5211](https://git.cafe/pingdotgg/t3code/commit/ca796d52119592de771d8ccfa955d70dba84d11e)
SHA-256: f1fe6b1e1e3060249edffda256f9ff2aab530f0b2a945e8c6216909da70684ab | 
[f76e21be6c](https://git.cafe/pingdotgg/t3code/commit/f76e21be6cdcf0936a28521badaf66e4a85bb9cc)
SHA-256: 20df0ba6c65c083e2409bbedbaef2cacd3410f28bee0c8253d60231e3a145c35 | 
[ca796d5211](https://git.cafe/pingdotgg/t3code/commit/ca796d52119592de771d8ccfa955d70dba84d11e) → [f76e21be6c](https://git.cafe/pingdotgg/t3code/commit/f76e21be6cdcf0936a28521badaf66e4a85bb9cc)
SHA-256: ca09dbbb5f3cc57bc405e3a8fac53a7daa1eed94fe60305a399d211cc95ee52e | 
[ca796d5211](https://git.cafe/pingdotgg/t3code/commit/ca796d52119592de771d8ccfa955d70dba84d11e) → [f76e21be6c](https://git.cafe/pingdotgg/t3code/commit/f76e21be6cdcf0936a28521badaf66e4a85bb9cc)
SHA-256: 3ede74c9bfa880c0a2af855d37a4475fcfdfb6ad6261fa49362967a59f0252b3 |
## Row 2 — I2 inspector history error and retry — iPad 7 / iOS 18 / landscape / light
**Before frame: control unreachable.** Inspector is at the top of the initial bounded history window, CHILD ROW 04. “Load earlier activity” is positioned at y=33 under the solid header and is not visible or tappable there. A semantic tap resolves to (950,33), hits the header/Close inspector instead, and sends no history request: the proxy stays armed. Reopen the inspector and return to the same loaded-page start for the selected still. No before error state is claimed.
**After frame: error visible below the header.** At the same loaded-page start, the control is at y=103, followed by the wrapped, contract-valid error: **“The environment failed to answer this request (orchestration_thread_history_failed).”** The proxy deliberately fails one earlier-page HTTP request with EnvironmentInternalError status 500 and then returns to pass-through. CHILD ROW 04 remains below that error. The same seed script created both 13-turn child histories. The before and after stills intentionally compare unreachable-control versus reachable-error states under the brief’s allowed fallback; they are not the same error state.
| Before | After | Diff0 | DiffT (5% fuzz) |
|---|---|---|---|
| 
[ca796d5211](https://git.cafe/pingdotgg/t3code/commit/ca796d52119592de771d8ccfa955d70dba84d11e)
SHA-256: 405df40b5498cd4d9bee1f9ecb29f03b386a4bdef7c49dc76d46e5aac70ee984 | 
[f76e21be6c](https://git.cafe/pingdotgg/t3code/commit/f76e21be6cdcf0936a28521badaf66e4a85bb9cc)
SHA-256: ddd84b2f25a94b30a6d641f7977756351c506c6495b5f595fd81db9589093b9c | 
[ca796d5211](https://git.cafe/pingdotgg/t3code/commit/ca796d52119592de771d8ccfa955d70dba84d11e) → [f76e21be6c](https://git.cafe/pingdotgg/t3code/commit/f76e21be6cdcf0936a28521badaf66e4a85bb9cc)
SHA-256: 0c7dbb07e71b49a9055236de7a619c814cf18b6210a3624e52f92e8e7ec28a17 | 
[ca796d5211](https://git.cafe/pingdotgg/t3code/commit/ca796d52119592de771d8ccfa955d70dba84d11e) → [f76e21be6c](https://git.cafe/pingdotgg/t3code/commit/f76e21be6cdcf0936a28521badaf66e4a85bb9cc)
SHA-256: c1afb5fd3303452665bcaa3d64aec2d5ce3b80aad68de2d701813c07d19314a7 |
**After retry recording:** [Watch retry succeed](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e2d9bf4d0b750666/after-I2-retry.mp4) — H.264, 2160×1620 landscape, 30 fps, 14.93 s, audio removed, native landscape pixel dimensions retained after applying source rotation. The raw recorder used HEVC 1620×2160 plus rotation metadata and emitted about 14.94 s. Driver actions span 21.05 s: native recording omits some static tail time. Contact-sheet labels are encoded player times; the action timestamps below are UTC command dispatch times and must not be treated as identical clocks.
Driver steps on 2026-11-10 ([exact timestamped actions](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e3abe76d5140d195/recording-steps.json)):
- **13:30:19.557 UTC (+0.00 s):** start recording with the wrapped error and retry control visible.
- **13:30:21.899 / 13:30:24.481 (+2.34 / +4.92 s):** matching settled error screenshots.
- **13:30:26.618 (+7.06 s):** tap the load control at (895,103); request passes, and both error and load control disappear.
- **13:30:29.882 / 13:30:32.375 (+10.33 / +12.82 s):** two inspector pans upward through loaded history.
- **13:30:34.696 (+15.14 s):** accessibility snapshot confirms CHILD ROW 01 is present.
- **13:30:35.769 / 13:30:38.401 (+16.21 / +18.84 s):** matching settled success screenshots.
- **13:30:40.607 (+21.05 s):** stop recording.
Success still SHA-256: `fa6eb3383b98dd585c06141c2e6fc43984c161694eb91cbf709604a39299ed87`; after [f76e21be6c](https://git.cafe/pingdotgg/t3code/commit/f76e21be6cdcf0936a28521badaf66e4a85bb9cc).
## What else changed and limits
The supplied base-to-head source diff contains five mobile inset/layout files, 13 insertions and 16 deletions; there is no dependency upgrade. The accepted F5 full-frame comparison also contains a 5.5s versus 5.3s parent Worked label and a bottom-edge rendering difference. I2’s parent Worked labels are 5.2s versus 5.5s. These Replay timings are incidental. Both comparisons otherwise use the same fixture driver, theme, orientation and device size. The optional detail comparison includes the sidebar empty-label inset change described above.
**Unverified:** the before I2 error text, because its covered control cannot trigger the request through the UI. This is the brief’s allowed fallback, and the before still is explicitly labelled control unreachable. The after failure, wrapped error, successful retry and actual first row are verified. No requested row is omitted.
Cold opens normally prefer a socket snapshot, so a proxy 404 on the HTTP bounded endpoint did not produce F5. Deleting an open child produced a different fallback. Neither attempt is accepted evidence; the native offline/uncached state above exercises the requested placeholder directly. The first I2 fault response had an invalid error body and produced a long undeclared-status message; it was superseded by the contract-valid response before the selected still and recording.
## Drivers and cleanup
[Reproduction and pairing notes](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/210b295f5fd86f5d/pair-notes.md) · [History seed](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b66c62f44bd40adf/seed-rpc.mjs) · [Offline fixture](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8d25a0a7f1b76bac/offline-fixture.mjs) · [Fault proxy](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/28dd9d90d485f194/proxy.mjs) · [RPC driver](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/32200c958b9de529/rpc.mjs) · [Timestamped device actions](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/50335bb4bc0d480c/action.py) · [Settled capture driver](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/33e062b70570c9af/capture.py) · [Retry recording driver](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0268fad4d38abb47/record-retry.py) · [Diff driver](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/70156f79582b609b/make-diffs.py) · [Stability packaging](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5b25782fae5e178b/verify-package.py) · [Bounded health-cache shim](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6c5b1afdf62c0964/cached-health.mjs) · [Session workspace wrapper](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/fc285742e563f737/agent-device-root).
[Cleanup verification](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e1916faa207adbcf/cleanup-report.json): both cap4-before and cap4-after closed successfully and then disconnected; all six server/Metro/proxy ports have no listeners. All captured server process groups were stopped by tracked PIDs; both Metro port owners had already exited when rechecked. The before worktree was removed; the lead’s after worktree is clean and preserved. Scratch remains at `/tmp/lead-32936-r1/capture4/` as explicitly requested, overriding the brief’s generic delete-scratch paragraph. No source edit, commit, push, PR creation or PR comment was made. Disk available: 236 GiB.
FRICTION:
- infra | native-client transfer/install through the gateway took 246–250 seconds per device | remote-download installs or warm cached clients.
- tool | iOS open-confirmation prompts and pairing links arriving before first bundle load required accepting prompts and pairing again after readiness | await a running app before pairing.
- tool | iPad coordinate injection lands about 60 pt right, including semantic targets; the environment-switch tap expanded its containing row | calibrated coordinates and verification of unchecked state.
- tool | devices inventory returned duplicate-request-id errors | reliable read-only inventory; requested device identity was selected explicitly by pool open.
- brief | deletion yields a separate fallback, and HTTP404 is bypassed by socket-first cold opens | use the explicitly supported offline/uncached trigger for F5.
- tool | a fading scroll indicator rejected one before-I2 frame pair | accept only a subsequent zero-difference pair.
- tool | two simultaneous capture drivers chose the same timestamped directory | include the capture name in each new directory.
- brief | explicit Keep scratch conflicts with generic final deletion text | follow the explicit Keep instruction.
AI-assisted; verified with the tests and tools listed above.
## Verbatim ghdrop output
The following uploader outputs are unchanged. Videos remain ordinary links.
### assets-upload.md
```markdown
[after-I2-retry.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e2d9bf4d0b750666/after-I2-retry.mp4)
```
### scripts-upload.md
```markdown
[seed-rpc.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b66c62f44bd40adf/seed-rpc.mjs)
[offline-fixture.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8d25a0a7f1b76bac/offline-fixture.mjs)
[proxy.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/28dd9d90d485f194/proxy.mjs)
[rpc.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/32200c958b9de529/rpc.mjs)
[action.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/50335bb4bc0d480c/action.py)
[capture.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/33e062b70570c9af/capture.py)
[record-retry.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0268fad4d38abb47/record-retry.py)
[make-diffs.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/70156f79582b609b/make-diffs.py)
[verify-package.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5b25782fae5e178b/verify-package.py)
[stability-manifest.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/91ce69fbcb3b1cbf/stability-manifest.json)
[raw-stability-pairs.zip](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/39b05f6371cdb244/raw-stability-pairs.zip)
[recording-steps.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e3abe76d5140d195/recording-steps.json)
[cleanup-report.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e1916faa207adbcf/cleanup-report.json)
[placeholder-measurements.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/103626867e221504/placeholder-measurements.json)
[pair-notes.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/210b295f5fd86f5d/pair-notes.md)
```
### helpers-upload.md
```markdown
[cached-health.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6c5b1afdf62c0964/cached-health.mjs)
[agent-device-root](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/fc285742e563f737/agent-device-root)
```