# Issue16105 final evidence

Issue [#16105](https://git.cafe/pingdotgg/t3code/issues/16105). Before [b5a226e49c](https://git.cafe/pingdotgg/t3code/commit/b5a226e49c52ecab4ddc2d678a7012d97e651a6a); after [8ac56474a3](https://git.cafe/pingdotgg/t3code/commit/8ac56474a31b840c4b149d58730a0a34cf98e3df). All canonical captures use these source revisions. No PR created; no production implementation edits during capture.

Eight canonical S1 images, four primary recordings plus two targeted supplements, four before/after/amplified-Diff0/DiffT sheets, exact authenticated app RPC observations for both search methods, committed self-contained synthetic fixture, and independent fresh-DB pixel rebuild. All required web matrix rows covered. Unverified surfaces/subcases are explicitly listed below.

## Environment and reproduction

Native T3 collaborative preview; Chrome154.0.0.0 on Linux x86_64, DPR2, visualViewport.scale1, UTC. CSS viewport1440×1000 or390×1000 (desktop Chrome at narrow width; **not native mobile**). App theme setting `t3code:theme=system`, with native prefers-color-scheme light/dark. Date frozen2026-10-21T12:00:00.000Z before SPA thread remount. Font `-apple-system, BlinkMacSystemFont, "Segoe UI", system-ui, sans-serif`; root16px, prompt14px, code13px. All panels closed, find closed, scrollTop0, F2 collapsed and F3 folded for S1; pointer CSS(2,500), no tooltip. Narrow activeElement BODY, editor bounding height0 (compact composer).

Native desktop still output1280×889; narrow output780×2000. Raw desktop videos2880×2000; delivered captioned videos2880×2160 (160px added black bottom band). Raw narrow videos780×2000; delivered780×2240 (240px band). App viewport content is retained; captions never cover it. Text narration, no audio. Video encoding is VFR; captions preserve media presentation time, not agent/tool elapsed time. Exact dimensions and raw/delivered durations appear per recording below and in each JSON sidecar.

Original homes use environment1dabb2cf-3ba8-4f1a-857f-2e4abd91687c. Final after-light390 replacement and I4 supplement use fresh isolated rebuild867b12ab-66f3-40ee-95ca-c7bdff7be1ac, same committed fixture and after SHA. Original before/after origins6789/7169; servers owned by parent. Exact build/navigation/freeze/settle recipe: [render-recipe.md](render-recipe.md). Full environment fields: [revision-metadata.json](revision-metadata.json).

Committed fixture source: [issue-16105.json](https://git.cafe/pingdotgg/t3code/src/commit/8ac56474a31b840c4b149d58730a0a34cf98e3df/apps/server/scripts/fixtures/issue-16105.json), [skeleton](https://git.cafe/pingdotgg/t3code/src/commit/8ac56474a31b840c4b149d58730a0a34cf98e3df/apps/server/scripts/fixtures/issue-16105-skeleton.json), [seeder](https://git.cafe/pingdotgg/t3code/src/commit/8ac56474a31b840c4b149d58730a0a34cf98e3df/apps/server/scripts/fixtures/seed-issue-16105.py). Run against stopped migrated isolated `.t3` homes. No preexisting UI-created thread is needed; sanitized projection rows include project/thread/run/node/message/item relationships. This is **projection-only static data plus actual RPC reads**, not provider-send/event-replay proof. Parent's send reproduction is separate. Seeder was also exercised twice against schema-only isolated DB with no skeleton; all20 typed payloads decoded.

## Canonical stills and pixel diffs

Diff0 = absolute RGB, zero tolerance; displayed amplification×8 clamps255. DiffT marks magenta iff max channel absolute difference>16/255. Full-frame images are unaligned/unmasked. Raw Diff0, x8, DiffT and metrics are bundled and individually linked in uploads.md.

### light, 1440 CSS pixels

Before: [before-light-1440.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/336dce92ea5dcd4d/before-light-1440.png) · After: [after-light-1440.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2aea29855c419bb5/after-light-1440.png)

![light-1440-comparison.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c0a4253e1f1abca4/light-1440-comparison.png)

### dark, 1440 CSS pixels

Before: [before-dark-1440.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2d8cfb2e0d9ea769/before-dark-1440.png) · After: [after-dark-1440.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3c4a456b347da8d3/after-dark-1440.png)

![dark-1440-comparison.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/dd21d9f280b59b64/dark-1440-comparison.png)

### light, 390 CSS pixels

Before: [before-light-390.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/374e12c9ad0968cc/before-light-390.png) · After: [after-light-390.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/177af4a8ff262269/after-light-390.png)

![light-390-comparison.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5db6fc10d325d7e3/light-390-comparison.png)

### dark, 390 CSS pixels

Before: [before-dark-390.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a76ea05d5e2eb0ef/before-dark-390.png) · After: [after-dark-390.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/97e4d1dc99db2788/after-dark-390.png)

![dark-390-comparison.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/36cb740e8165defa/dark-390-comparison.png)

Desktop normalized: header, assistant and floating scroll button pixels match exactly. Differences are only F1 y67–294 and F4 bottom sliver y871–887. Desktop light Diff0=9604px, DiffT=4578px; dark9732/4702. Narrow F1 gains one23CSS-pixel line, moving later rows46device pixels; assistant crop translated46pixels matches exactly. Translucent composer background and lower-row cropping follow that reflow. Exhaustive accounting, including all capture-state differences and superseded material: [visual-differences.md](visual-differences.md). No unexplained pixels are being attributed to the fix.

## Matrix coverage

| Row | Status / evidence |
|---|---|
| S1 prose, parentheses, quote, list, actual GFM table, ~solo~, ~~old~~, bold/link | Verified8canonical stills above, same F1/F1A8-line fixtures including required blank line before table. User single tildes literal after; double tilde stays struck. |
| S2 F2 inline file chips, literal HTML, both fences, tail | Both revisions desktop-dark and narrow-light recordings at initial chapter; F2-expanded stills in uploads.md. Code fences keep literal tildes both sides. |
| I1 raw F1 path, old, ~solo~, next/previous, close/reopen | Both widths/both revisions recordings; after raw/solo1match vs before0. old9matches unchanged (includes folder substrings). Actual highlighted DOM ranges in raw tool-clock logs; settled media frames/chapter anchors below. |
| I2 collapsed F2 find/reveal plus manual toggle | Before0/0, no auto-reveal; after1/1 active highlighted tail and expanded row. Desktop after around202.3–206s; new narrow around36.5–44s. Manual expand/collapse recorded. |
| I3 automation summary/body and find auto-reveal | Before body strips single tildes and0find; after literal path and1find. New narrow begins folded, manual expand/collapse then automatic reveal around60s. Desktop dedicated I3 supplement at9.7–13.5s proves reveal/highlight. |
| I4 legacy visible path, hidden output, context-label exclusion | Desktop before around58.9–64.9s: visible0, exclusions0. After I4 supplement: visible1+highlight held53.5–63.8s, hidden output0held67.3s onward, terminal label0held75.7s onward. Original after main dropped intermediate hidden-output frame; supplement closes that gap. |
| C1 assistant/default markdown | Same text assistant in all8S1 images; original single-tilde strike retained. Desktop entire assistant pixels exact; narrow translated assistant crop exact. |
| C2 double-tilde strike | User and assistant ~~old~~ retained in stills and automation body recordings. |
| U1 native Electron | Unverified; shared web renderer verified in Chrome only. |
| U2 native mobile | Untouched and unverified; known follow-up. Narrow web is not a native mobile fix claim. |
| Generated structured terminal-context visual subcase | Unverified: no real UI-generated terminal context record available. F2 empty-record context/plain link is an extraction control only, never labeled generated-chip proof. F4 legacy terminal chip is real rendered legacy syntax and separately covered. |

## Recordings: narrated chapters in actual media time

Chapter timestamps are interval starts in the delivered MP4; settled frame anchors are stated separately. Captions are burned into added black padding. JSONs in recordings/ use encoded media seconds. The archived `raw/*-tool-clock.json` elapsed timestamps are diagnostics only and are **not valid seek times**. In particular original after desktop agent elapsed59.396s encoded230.72s; waiting/capture transport time must not be treated as app latency. Source MP4s retained in recordings/raw; no hidden cuts or speed changes.

### before-dark-1440

[before-dark-1440.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/db09e257546b89b2/before-dark-1440.mp4)

Raw 2880×2000, 64.93s; delivered 2880×2160, 65.35s. Environment `1dabb2cf-3ba8-4f1a-857f-2e4abd91687c`.

- 0.00–7.00s: S2: Expand F2; inspect file chips, literal HTML and both code fences.
- 7.00–10.00s: I1: Raw path returns0/0.
- 10.00–28.00s: I1: old has nine matches; next/previous.
- 28.00–35.00s: I1: ~solo~ returns0/0. Close/reopen find.
- 35.00–41.00s: I2: ~/hidden/a returns0/0; tail stays collapsed.
- 41.00–52.00s: S2/I3: Manually expand/collapse F2 and automation.
- 52.00–56.50s: I3: ~/auto/a returns0/0.
- 56.50–60.00s: I4: Visible ~/legacy/a returns0/0.
- 60.00–61.30s: I4: Hidden output returns0/0.
- 61.30–64.93s: I4: terminal label returns0/0.

### after-dark-1440

[after-dark-1440.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/f8bb65c3984663a0/after-dark-1440.mp4)

Raw 2880×2000, 230.72s; delivered 2880×2160, 231.63s. Environment `1dabb2cf-3ba8-4f1a-857f-2e4abd91687c`.

- 0.00–13.00s: S2: Expand F2; inspect chips, literal HTML and both fences.
- 13.00–21.50s: I1: Initial query pending. Tool/media clocks differ.
- 21.50–95.00s: I1: old remains Searching...; not a latency benchmark.
- 95.00–156.20s: I1: old settles5/9. Raw-path retry follows.
- 156.20–172.00s: I1: Raw ~/folder-a/skills/x returns1/1, highlighted.
- 172.00–184.50s: I1: Close find, then reopen.
- 184.50–199.00s: I1: old has nine matches; next/previous.
- 199.00–200.80s: I1: ~solo~ returns1/1 with active highlight.
- 200.80–202.30s: I2: Reopen find for collapsed F2 tail.
- 202.30–206.30s: I2: ~/hidden/a returns1/1 and reveals tail.
- 206.30–218.00s: S2/I3: Close find; toggle F2 and automation.
- 218.00–226.20s: I3: Find ~/auto/a; settles1/1 by220.6s. See reveal supplement.
- 226.20–228.47s: I4: ~/legacy/a1/1 at228.258s (brief frame).
- 228.47–230.72s: I4: terminal label0/0. Hidden-output recording is in I4-exclusions supplement.

### before-light-390

[before-light-390.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/361749fc47e2d5cd/before-light-390.mp4)

Raw 780×2000, 18.81s; delivered 780×2240, 18.95s. Environment `1dabb2cf-3ba8-4f1a-857f-2e4abd91687c`.

- 0.00–3.50s: S2: Narrow web. Expand F2; chips and fences.
- 3.50–5.50s: I1: Raw path returns0/0.
- 5.50–9.80s: I1: old has nine matches; next/previous.
- 9.80–11.50s: I1: ~solo~0/0; close/reopen.
- 11.50–15.00s: I2: ~/hidden/a0/0; manual toggle.
- 15.00–18.81s: I3: Toggle automation; ~/auto/a0/0.

### after-light-390

[after-light-390.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8b78f7da91fce1a5/after-light-390.mp4)

Raw 780×2000, 66.04s; delivered 780×2240, 66.35s. Environment `867b12ab-66f3-40ee-95ca-c7bdff7be1ac`.

- 0.00–8.50s: S2: F2 collapsed and F3 folded. Expand F2; chips and both fences.
- 8.50–18.00s: I1: Open find; raw path settles1/1 by12.5s.
- 18.00–24.00s: I1: old has nine matches; next then previous.
- 24.00–32.00s: I1: ~solo~ settles1/1; close/reopen.
- 32.00–44.00s: I2: Find hidden tail;1/1, highlighted and revealed by36.5s.
- 44.00–50.00s: I2/S2: Close find; manually collapse, expand, collapse F2.
- 50.00–58.00s: I3: Manually expand automation then collapse it.
- 58.00–66.04s: I3: ~/auto/a settles1/1, opens automation and highlights path.

### after-dark-1440-I3-reveal

[after-dark-1440-I3-reveal.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a6d0a9d2ccd1d658/after-dark-1440-I3-reveal.mp4)

Raw 2880×2000, 13.54s; delivered 2880×2160, 14.40s. Environment `1dabb2cf-3ba8-4f1a-857f-2e4abd91687c`.

- 0.00–6.10s: I3: Automation initially folded. Open find.
- 6.10–9.70s: Type ~/auto/a; wait for settled search.
- 9.70–13.54s: I3:1/1. Automation opens; literal path highlighted.

### after-dark-1440-I4-exclusions

[after-dark-1440-I4-exclusions.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0a57bce3792121fd/after-dark-1440-I4-exclusions.mp4)

Raw 2880×2000, 80.33s; delivered 2880×2160, 80.33s. Environment `867b12ab-66f3-40ee-95ca-c7bdff7be1ac`.

- 0.00–52.00s: I4 supplement: Find open. Typing transport retries; skip to52s.
- 52.00–65.00s: I4: Visible ~/legacy/a settles1/1 with active highlight.
- 65.00–73.50s: I4: Find hidden output; settles0/0 by67.3s, then held.
- 73.50–80.33s: I4: Find terminal label; settles0/0 by75.7s, then held.

## Exact app RPC observations

[Before exact responses](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6e2e4fdb744781ee/before.json) · [After exact responses](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/87d16862e935d93c/after.json)

App-generated `threadFind` query and `threadFindProgressive` subscription atoms invoke orchestration.searchThread/searchThreadStream through the real authenticated client. Logs retain exact success values, including repeated stream observations; these are app atom observations, not a claim of raw transport frame capture. Auth fields omitted. Query inputs and all response fields retained. Both methods give the same totals:

| Query | Before | After |
|---|---:|---:|
| ~/folder-a/skills/x |0|1|
| old |9|9|
| ~solo~ |0|1|
| ~/hidden/a |0|1|
| ~/auto/a |0|1|
| ~/legacy/a |0|1|
| hidden output |0|0|
| terminal label |0|0|
| hidden/b).terminal label |0|0|

Reproduce in authenticated browser: run scripts/browser-rpc-setup.js, then scripts/browser-rpc.js with the destination environment ID adjusted. No provider sends or mutations. Hidden structured context is excluded from search, but a valid generated visual chip remains unverified as above.

## Independent rebuild

Proof2 auditor independently rebuilt fresh isolated projection state from committed fixture at final SHA and reproduced the canonical after-dark1440 capture. Full1280×889 pixels: zero changes at threshold0 and>16, no mask/alignment. Reported by independent auditor; copied exact metrics and screenshot:

[Independent metrics](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/dc3ba68d5230c6c7/metrics.json)

![rebuilt-after-dark-1440.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/cde90773613c7e5c/rebuilt-after-dark-1440.png)

## Scope, limitations and friction

User single-tilde behavior changes across prose and ~solo~. Assistant/default markdown and double-tilde behavior unchanged. No CSS, dependencies, Electron-shell or native-mobile files changed. Fixture/seeder source and focused tests are the remaining diff. Parent reports scoped shared12/12 and browser10/10; evidence agent did not rerun those tests or launch servers. No PR, no live DB writes, no screenshots/assets committed. Local/Chrome verification only; no relay/tunnel/provider-adapter compatibility claim.

Explicit unverified: native Electron, native mobile (untouched follow-up), generated structured terminal-context visual chip. All other required matrix web rows have still/recorded evidence. Scheduled automation was projection-only, with no active scheduler started.

Friction by category in friction.json: infra browser disconnect/reconnect; tool typing transport failure, native output resizing and tool/media clock mismatch; brief fixture table correction; runbook freeze-before-remount and responsive focus settlement; model initial composer-state capture error corrected. Superseded pre-table, initial expanded-F3 narrow recording, old desktop-button screenshots and old uploaded URLs are excluded from this index/bundle. Final filename stability preserved.

Machine-readable artifact index: manifest.json; per-file SHA256: checksums.sha256. Public uploads are world-readable synthetic evidence only. uploads.md retains latest applicable ghdrop -m lines verbatim. Bundle excludes private pairing URLs, credentials, uncommitted DBs and superseded artifacts.