# Independent pass-2 rebuild

Reviewed issue [#16923](https://git.cafe/pingdotgg/t3code/issues/16923), production base [b5a226e49c](https://git.cafe/pingdotgg/t3code/commit/b5a226e49c52ecab4ddc2d678a7012d97e651a6a), final head [a3da62d35b](https://git.cafe/pingdotgg/t3code/commit/a3da62d35b554e53cf41bcd5c31fde28b82ece62), production implementation [ca0ff3e90c](https://git.cafe/pingdotgg/t3code/commit/ca0ff3e90c). Source remained unchanged. No server, tests, alternate browser, forge mutation, commit or push was started by this auditor.

After parent released capture ownership, used own T3 Preview tab against the existing isolated server on 7293/15333. Auth reused without printing a token. CSS viewport 390x844, DPR 2, light appearance, terminal height 300, default font 13, frozen fixture Date. Imported committed fixture, reset its thread, observed actual Ghostty methods, renamed fixture directory, opened drawer.

First attempt reattached a prior running recovery PTY: the fresh browser had no persisted terminal IDs to close. This attempt was rejected as missing-folder evidence. Resetting the now-recorded fixture session closed term-3, then opening a fresh term-1 produced the actual restore instruction. No synthetic error was injected.

## Cold checkpoint reproduction

`rebuild-N-L-cold.png` is the whole viewport. Compared without resizing/masking against final `after-N-L-cold.png`: 8604/1316640 pixels differ at any RGB delta (0.6534816%); 7514/1316640 exceed max-channel delta 8/255 (0.5706951%). `rebuild-Diff0.png` has gain 1, threshold 0; `rebuild-DiffT8.png` is the strict threshold map. Terminal region y>=1088 physical pixels has ZERO differing pixels at either threshold. Remaining differences are the vertically displaced empty-chat label and breadcrumb chevron. This is a behavioral reproduction and exact terminal-region match, not whole-app pixel reproducibility.

## Natural mounted-effect proof

On the parent's follow-up request, wrapped the real GhosttyTerminalSurface.prototype.write while preserving its original call; recorded Error().stack only for actual [terminal] writes. Renamed the fixture directory again and clicked New Terminal. One locator capitalization mismatch timed out; refreshed snapshot and clicked its exact accessibility ref.

The genuine missing-folder write stack names `writeSystemMessage`, then `t40` at served ThreadTerminalDrawer.tsx:1259, followed by React passive-effect frames. Fetched served module context confirms t40 is the changed session update effect, separately from the setup effect. Live surface id5 was ready at performance.now 154272.2 ms; actual write at 154292.0 ms (wrapper), traced original write 154292.5 ms. There was no intervening reset/write/selection-clear on that surface.

`mounted-effect-stack.json` preserves the stack, trace, served context, viewport and DPR. `rebuild-N-L-new.png` records the resulting error in Terminal 2. This establishes one natural real-app post-mount delivery from the changed effect. It does not establish deterministic gating, raw attach error-arrival timestamp, directly measured equal output/version counters, a matching baseline stack recording, or four-cell causal coverage. No video was made for this supplemental stack attempt.

Folder restored to its exact original path after each accepted attempt. Parent was notified that server 60622 was released; no further browser actions are needed.

## Final capture checks

Final manifest lists 88 artifacts (80 checkpoint/difference images and 8 recordings). Every listed local SHA256 matched. Independently recomputed all 16 absolute RGB gain-1 images, strict >8 threshold maps and changed-pixel counts: all matched. Sampled each before/after video at four elapsed points into local contact sheets; samples show the real drawer/session progression and shell recovery. Timeline offsets are tool-completion observations, approximate seek positions rather than frame-exact input timing.

Final baseline was a three-production-file base overlay with final fixture tooling retained, not a baseline HEAD checkout. Independently confirmed overlay matched base before capture and source matched clean final HEAD after restoration. Wide sidebar geometry and relative timestamps remain mismatched; final index correctly discloses them. Both narrow recovery pairs are exactly pixel-equal; index prose suggesting N-L recovery cursor differences is stale against its own correct zero metrics.

Core missing proof remains full C2 across cells/counters, C3 transition cases, C4 history/focus, C5 aliases, complete L1-L6 recordings, and H1 supplemental truth-table cases. O1/P1/P2/P3/D1 remain unverified extensions, S1/P4 source exclusions. These gaps are not waived by the successful single mounted-effect observation.