No proof contradictions found in the twelve delivered Linux rows for PR [#25970](https://git.cafe/pingdotgg/t3code/pull/25970). W remains `cannot-verify windows`, the explicit limitation accepted by the brief. The sufficient verdict below covers the requested Linux EPERM proof and does not claim native Windows execution.

Reviewed head [bfd52848f9](https://git.cafe/pingdotgg/t3code/commit/bfd52848f9c02add53ab87e4a7dff31aae12b745) against its direct parent and PR observed base [e9236274a3](https://git.cafe/pingdotgg/t3code/commit/e9236274a3a6e0c731a7df4092db1e686f039068). The forge metadata still names this head, branch fix/issue-19681, and target main. The supplied worktree stayed clean at the requested head.

I read all twelve complete raw transcripts, all three evidence sheets, the index, matrix, capture.py, summarize.py and preload. I independently checked every row's printed SHA, owning-package cwd, command, NODE_OPTIONS, preload hash, runtime, named outcomes, summary and exit status. The standalone transcripts, scripts and sheets also match their bundle members byte for byte; no bundle extraction or downloaded Python execution was needed.

All seven parent EPERM failures originate in the injected fs.symlink during fixture setup, before the tested behavior/assertions. The server scanner's two pre-existing alias cases skip on both EPERM revisions. Comparing all 118 test names across revisions shows exactly seven F-to-S changes under EPERM and zero changes normally. No other skips, failures, added/missing names or unhandled errors appear. These comparisons agree with the sheets' Diff0 and what-else-changed claims, the pass-1 expectations and the PR body's focused-test Verification table. The separate check:changed/knip claim was not independently re-audited by this proof pass.

The complete transcript recordings are the authorized substitute for visual captures. Their fixture descriptions match the checked-in test bodies and observed named results. Viewport, theme, image thresholds, pixel deltas and narrated UI steps are N/A for this test-only change.

P/F/S = passed/failed/skipped. Parent EPERM exit 1 is the expected reproduction.

| Row | Observed P/F/S | Exit | Status |
|---|---:|---:|---|
| [S-P-N](/tmp/lead-25970-r1/dl/S-P-N.txt) | 100/0/0 | 0 | SUFFICIENT |
| [S-P-E](/tmp/lead-25970-r1/dl/S-P-E.txt) | 95/3/2 | 1 | SUFFICIENT |
| [S-H-N](/tmp/lead-25970-r1/dl/S-H-N.txt) | 100/0/0 | 0 | SUFFICIENT |
| [S-H-E](/tmp/lead-25970-r1/dl/S-H-E.txt) | 95/0/5 | 0 | SUFFICIENT |
| [O-P-N](/tmp/lead-25970-r1/dl/O-P-N.txt) | 5/0/0 | 0 | SUFFICIENT |
| [O-P-E](/tmp/lead-25970-r1/dl/O-P-E.txt) | 3/2/0 | 1 | SUFFICIENT |
| [O-H-N](/tmp/lead-25970-r1/dl/O-H-N.txt) | 5/0/0 | 0 | SUFFICIENT |
| [O-H-E](/tmp/lead-25970-r1/dl/O-H-E.txt) | 3/0/2 | 0 | SUFFICIENT |
| [A-P-N](/tmp/lead-25970-r1/dl/A-P-N.txt) | 13/0/0 | 0 | SUFFICIENT |
| [A-P-E](/tmp/lead-25970-r1/dl/A-P-E.txt) | 11/2/0 | 1 | SUFFICIENT |
| [A-H-N](/tmp/lead-25970-r1/dl/A-H-N.txt) | 13/0/0 | 0 | SUFFICIENT |
| [A-H-E](/tmp/lead-25970-r1/dl/A-H-E.txt) | 11/0/2 | 0 | SUFFICIENT |
| W | Not run | — | cannot-verify windows (expected limitation) |

Independent rebuilds used the exact delivered O-H-E and A-H-E test commands from the owning packages, changing only NODE_OPTIONS to my copied preload. Its verified SHA-256 is `e364864529ef79a816183b75d210f52b7fc134bc220b318de3dcaddf93cecdde`. The environment matches: Linux fs-35/kernel 6.1.102/x86_64, Node v26.12.1, local Vite+ 1.4.0, Vitest 5.0.5 and the same rendering-N/A line. Both exited zero, with **0 changed test names or statuses** against the delivered captures; timing differences were ignored.

- [O-H-E rerun](/tmp/lead-25970-r1/proof-audit-3/rerun-20261027T143700.023867Z/O-H-E.txt): 3 passed, 2 skipped.
- [A-H-E rerun](/tmp/lead-25970-r1/proof-audit-3/rerun-20261027T143700.023867Z/A-H-E.txt): 11 passed, 2 skipped.
- [Independent validation output](/tmp/lead-25970-r1/proof-audit-3/audit-validation.json) and [validator](/tmp/lead-25970-r1/proof-audit-3/verify-evidence.py).

No server, browser, emulator or lease was started. Evidence, worktree and reruns remain for the lead, as instructed. Available disk space remained 19 GB.

FRICTION:
infra | No native Windows runner; W remains the expected unverified platform row. | A Windows runner with symlinks denied would verify W.
tool | One combined evidence read was truncated; all affected material was reread in smaller batches. | Bound combined reads to the tool output budget.

VERDICT: SUFFICIENT