# Project retention evidence refresh

The missing baseline dark project chapters and correctly dark desktop head loaded-return recording are now supplied. **Exhaustive visual attribution is still incomplete:** the narrow focus-ring corner residual and light desktop sidebar seam have measured, inspected regions but no verified exact compositor cause. This is an evidence limitation; production MERGE verdicts remain unchanged.

Issue [#14743](https://github.com/pingdotgg/t3code/issues/14743), reused PR [#22208](https://github.com/pingdotgg/t3code/pull/22208). Baseline [22840ab943](https://github.com/pingdotgg/t3code/commit/22840ab943b8418744b48223411e46ba4fdcef1a); head captures at clean production head [30734756fc](https://github.com/pingdotgg/t3code/commit/30734756fcc3424fc3cb1350235c7a190182f170), with external capture-script changes. Replay fixtures committed afterward as local-only [0642083522](https://github.com/pingdotgg/t3code/commit/0642083522df9989de2165a8707012376a1e5864). No production or reviewed browser-test bytes changed against either reviewed production commit. No push.

## Coverage and reproducibility

Eight selected raw recordings, 72 checkpoints, 24 whole-surface pairs: baseline/head × CSS1440×1000 or390×844 × light/dark. Nine checkpoints per cell: Off, first enable, input-click/Ctrl+A selected8, committed30, reload30, Off before leave, enable before leave, General, loaded Storage return. A single registered disposable project with Custom policy is reused, and each chapter starts from the same authoritative null rules. Environment rules remain null/false. No eligible cleanup worktrees or artifacts are seeded; Delete now is never used.

These are real app routes and real server persistence, not extracted component callbacks. Baseline enable immediately writes and acknowledges8; head enable has zero writes and authoritative null. Both replace input with30; Enter then outside blur commits. The selected suite awaits an actual successful30 acknowledgement and then checks server readback; full reload still reads30. After Off, baseline enable persists8, so General and loaded Storage retain On8. Head's uncommitted8 disappears and loaded Storage is Off/null. Baseline draft checkpoints remain unreachable; none is synthesized.

The previously verified R1/R4/R5/R7/R8/R10/R11 suites were not repeated. Existing valid R6 settled-worktree/0, checkout disposal, other project persistence and earlier R2/R3/R9 evidence retain their declared scope. This refresh covers the specific missing project-I theme/comparison/return chapters; it does not claim a fresh complete R6 rerun or native desktop/mobile verification.

Theme is seeded through localStorage **before full navigation**: t3code:theme=light/dark, t3code:theme-follow-system=false, t3code:theme-appearance-mode=light/dark. No class toggling is used. Every loaded checkpoint asserts persisted preference and applied class; per-document mutation logs are retained. Dark raw frames were scanned at an unobstructed background point: all451 decoded dark frames remain dark (max channel12); desktop contact sheets and loaded final frames were inspected. This pixel probe alone is not a semantic page oracle: final-frame inspection and loaded-route DOM/readback provide that evidence. Loaded return is held across animation frames for at least1.2s. Full navigation splash frames remain in the raw files.

Same project title/path/data, no scrolling, same browser/default zoom, DPR2, en-US, UTC, reduced-motion=false, system font stack. Fonts and actual finite document animations settle before stills; raw recordings preserve transitions. Saved desktop PNGs are1280×889, narrow780×1688; raw desktop video2880×2000. No image alignment, registration, masks, cropping or scaling enters counts. The clock was not frozen; this project surface shows no changing clock text. Narration is textual in the action index, with UTC actions and checkpoint-completion offsets; no audio narration or frame-exact synchronization is claimed.

## Pixel accounting

Both RGB and RGBA counts are supplied. diff0 means any channel changes; diff16 means maximum absolute channel delta>16. Alpha is separately measured, avoiding the false conclusion that an RGB-identical transparent seam is fully identical. Every changed pixel is partitioned into inactive control, lower rows or the single sidebar-boundary column in pixels.json/CSV. Diff0, diff16, RGBA maps and8× amplified maps retain the entire surface.

| Pair | RGB diff0 / >16 | RGBA diff0 / >16 |
|---|---:|---:|
| 1440-light-Off | 34 / 0 | 34 / 0 |
| 1440-light-enable | 2362 / 1514 | 2422 / 1645 |
| 1440-light-enable-selected | 386 / 0 | 390 / 0 |
| 1440-light-committed30 | 225 / 0 | 347 / 0 |
| 1440-light-reload30 | 255 / 0 | 255 / 255 |
| 1440-light-loaded-return | 3732 / 947 | 3732 / 947 |
| 1440-dark-Off | 0 / 0 | 0 / 0 |
| 1440-dark-enable | 1654 / 1138 | 1654 / 1138 |
| 1440-dark-enable-selected | 0 / 0 | 0 / 0 |
| 1440-dark-committed30 | 0 / 0 | 0 / 0 |
| 1440-dark-reload30 | 0 / 0 | 0 / 0 |
| 1440-dark-loaded-return | 3726 / 908 | 3726 / 908 |
| 390-light-Off | 0 / 0 | 0 / 0 |
| 390-light-enable | 7288 / 5110 | 7288 / 5110 |
| 390-light-enable-selected | 28 / 3 | 28 / 3 |
| 390-light-committed30 | 0 / 0 | 0 / 0 |
| 390-light-reload30 | 0 / 0 | 0 / 0 |
| 390-light-loaded-return | 79951 / 51802 | 79951 / 51802 |
| 390-dark-Off | 0 / 0 | 0 / 0 |
| 390-dark-enable | 7218 / 5037 | 7218 / 5037 |
| 390-dark-enable-selected | 29 / 3 | 29 / 3 |
| 390-dark-committed30 | 0 / 0 | 0 / 0 |
| 390-dark-reload30 | 0 / 0 | 0 / 0 |
| 390-dark-loaded-return | 80379 / 51322 | 80379 / 51322 |

Concrete regions:

- **Desktop dark:** Off, selected-focus8, saved30 and reload30 are exactly identical across the entire RGBA surface. First-enable difference is confined to x959..1115/y311..338: input focus border/ring versus unfocused baseline field, plus icon/switch edge pixels. The8/days glyph interior matches exactly. An actual click/Ctrl+A on both inputs removes every difference (0/0); this intervention establishes the focus-state explanation for this representative desktop comparison, rather than assuming AA. Loaded return's only changed region is that same control: baseline rendered saved8/On; head Off/null. All other desktop-dark pixels match.
- **Narrow:** Off, saved30 and reload30 are whole-surface identical. First-enable differences are confined to the inactive control and its autofocus ring/caret. Matching actual focus and selection leaves only28 light /29 dark pixels at x341..354/y1145..1198, 3 pixels above16 each, at the right curved focus-ring corners. The inspection crop shows the difference; its exact capture/raster cause remains unverified. It is not claimed harmless AA or a production regression.
- **Narrow loaded return:** baseline8 has the full number group, head has Off and an off switch. That changes the inactive control and reduces the row height. Every pixel of the lower visible rows matches after an exact20-image-pixel diagnostic translation (10 CSSpx): baseline y1248..1687 equals head y1228..1667 in RGBA, in both themes. This accounts for settled/merged row text, switches, rules and separators moving upward, and the different bottom viewport cut. Translation is used only for inspection; published full-surface counts are unaligned. All pixels above the inactive control match.
- **Desktop light:** all extra pixels outside the inactive control lie on x226, the sidebar boundary. Some change RGB and some alpha; notably reload30 has255 alpha differences above16 while the field/content match. Off has34 RGB-only subthreshold pixels; selected-focus has390 raw changed seam pixels, all subthreshold. This seam is isolated and measured, but its precise source/compositor cause remains unverified. There are no unexplained fine text-glyph regions elsewhere in these selected pairs.

The earlier unequal-seed/mislabeled R6 project pairs and light-ending desktop-dark recording are **superseded/nonproof for these comparisons**. Earlier component whole-surface outputs and unresolved old glyph/stripe/transition comparisons are not rehabilitated; these real-app pairs replace them only for the representative states declared here. Original184 assets and valid behavioral evidence are retained, not rewritten. The prior311-pixel replay switch discrepancy is not reclassified as solved by waiting: these new settled captures have a narrower declared scope.

## Downloads

- [project-proof.zip](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e309d3081f99b998/project-proof.zip)
- [after-desktop-dark.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/553b9d4972350edf/after-desktop-dark.mp4)
- [before-desktop-dark.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d5c65f314168f743/before-desktop-dark.mp4)
- [pixels.csv](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d2ba69d7dee21153/pixels.csv)
- [actions.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b9f91df0ebd6cbeb/actions.md)
- [All dark-frame background scans](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/61557a565fb335b9/dark-frame-scan.json)

![After: actual final raw frame is dark, loaded Storage with inactive retention Off](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/70c1fc40f1d1f98c/after-desktop-dark-final.png)

![Before left and after right: dark desktop inputs with actual matching focus/selection, whole-surface diff0=0](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/348584ab3ab07b24/desktop-dark-focused-before-after.png)

![Before left and after right: narrow dark loaded return, baseline saved8 versus head Off](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/429f59be563f4d8a/narrow-dark-return-before-after.png)

![Baseline left, head middle, amplified difference right: remaining narrow focus-ring corner pixels](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/18392e8fbd575527/narrow-ring-inspection.png)

## Validation and FRICTION

378 focused offline artifact/state assertions passed, including72 theme/preference/scroll checkpoints, actual readbacks, enable journal bounds, successful30 acknowledgement and exhaustive changed-region partitioning. These are artifact/state assertions, not378 app tests. Fixture JS syntax checks, disposable seed execution and targeted formatter passed; no repo-wide checks ran. Production-equivalence hashes are in the bundle.

- Initial raw WebSocket attempts omitted orchestrationProtocol=2; corrected before selected captures. No failed connection is evidence.
- Baseline Enter alone left30 unsaved; retained the actual outside blur commit. A later immediate read raced the acknowledged save; the selected suite waits on the matching actual ack event, without sleeps/polling. Abandoned segments are excluded.
- Desktop provider-update availability differed in the first batch. Selected captures disable update checks in both fixtures and move the pointer to the breadcrumb before stills. The abandoned batch is not the selected comparison set.
- New incognito tab lost authentication; hash navigation did not establish a session. The existing startup token entered through the actual token form recovered it. No credential or pairing URL is in the published evidence.
- A multi-operation same-file apply_patch was rejected; wrote a fresh external results filename. A contact-sheet read preceded generation in the selected directory; corrected the output root and inspected completed sheets. No missing sheet is claimed inspected.
- Used the [test-t3-app skill](/home/ubuntu/t3code/.agents/skills/test-t3-app/SKILL.md) and uploaded only sanitized artifacts with the [ghdrop skill](/home/ubuntu/.codex/skills/ghdrop/SKILL.md). App-owned PR link is registered, watching=false. No forge writes, comments or watcher.
- Stopped every owned retained server session; closed only owned preview tabs. Restored fix/issue-14743, committed only three replay fixture scripts, left the branch clean. Owned disposable homes and generated modules were removed; ports7763/15803 have no listeners. Other worktrees and live T3 state were not edited. App-managed evidence attachments are retained by the app; external evidence remains under /tmp/issue-14743-fresh-proof.

Verdict: targeted R6 missing chapters supplied; exhaustive visual attribution remains incomplete for the explicitly measured narrow ring corners and light sidebar seam. No change to production review verdicts.