PASS — fresh read-only evidence proof audit, pass 2

PR: [#24264](https://git.cafe/pingdotgg/t3code/pulls/24264)
Workspace: /home/ubuntu/wt/pr-24264
Base: 22dc1dabc6d3b1616d0fa5017268c086a77887bc
Head: 0350d612d92ccad3c10e114fe87aeb09fa8168a7

No concrete fixes required. This is an evidence audit, not an additional full code review.

Independent execution
- Executed the supplied script unchanged: node /tmp/to-24264-proof2/benchmark.mjs /home/ubuntu/wt/pr-24264 /tmp/to-24264-proof2
- Exit 0; results.json and run.log retained alongside this audit.
- Confirmed checkout HEAD and merge-base match the supplied revisions, and checkout remained clean.
- Script uses git show of both exact commits and imports their real client-runtime source modules, not copied reducer/equality algorithms. Workspace dependencies are shared by both; the revision diff changes only the two runtime files and their tests.
- Downloaded immutable published script/results using curl. Both are byte-identical to the supplied local originals. Script SHA-256: a57b4d5f8bd2a627bafde4eb57042402945b7b7dfe3fc94e582b4e6d09ab8ee1. Original results SHA-256: 0507f822be8fd0df337ab6f1468b3e82341a5ecdca8f1d38708670f36d142e8f.

Timing proof
- 10,000 items, 500 newest-item replacements per batch.
- One untimed warmup per revision/mode, five measured batches, alternating revision order. Timed runs precede accessor instrumentation; fixture/event construction, module load and equivalence traversal are outside the timer. Final output contributes to a consumed sink (240240). Combined mode also asserts equality decisions inside its loop, equally for both revisions.
- Rerun median reducer-only: 71.317083 ms base -> 16.381704 ms head (4.35x).
- Rerun median reducer-plus-equality: 133.803002 ms base -> 33.795145 ms head (3.96x).
- Original reported medians 73.785165 -> 15.269366 ms and 126.627947 -> 27.260831 ms agree with the original raw results. Independent timings vary, but reproduce the improvement; no exact-duration reproducibility claim is warranted.

Read-count proof
- Exact original counts independently reproduced.
- Reducer item.ordinal reads: 19,998,000 -> 1,000.
- Reducer row.position reads: 4,999,500 -> 500.
- Equality numeric slots: previous turnItems, next turnItems, previous visibleTurnItems, next visibleTurnItems each 5,000,000 -> 500; total 20,000,000 -> 2,000.
- Ordering accessors count fields separately and restore their counters across equality. Slot readers are installed on the original arrays only after reduction and enabled only during equality. No Proxy/slice replacement destroys the WeakMap hint identities; cleanup restores data properties before the next event.

Correctness proof
- Deep equality between actual base/head reducers after all 500 streaming events and each of the nine sequential adversarial events, plus all stored snapshot/sibling/resume histories.
- 14 snapshots per revision; 392 all-pairs comparisons total. Every hinted decision equals the same revision's full comparison with hint-free copied arrays, and base/head decision sequences agree.
- Histories cover consecutive updates, skipped/batched observations, reasoning/type and status changes, ordinal reorder, structural title change, rollback visibility filtering, inherited-to-local replacement, copied snapshots, sibling branches and resumed updates.
- Actual AtomRegistry quietThreadAtom test on both revisions: consecutive text updates retain identity, null yields null and clears state, resumed snapshot establishes a new structure, and subsequent text update retains that new identity.

Scope and limits
- Diff contains no render/layout changes; visualMatrix: [] is appropriate. No browser, device, screenshots or live-data access was needed or used.
- These are host synthetic microbenchmarks, not device, frame-time, rendering, transport or end-to-end measurements.
- O(N) array copies remain. Item/visible-row searches remain worst-case O(N); newest-item replacements favor the existing reverse searches.
- Mobile uses the shared reducer and predominantly non-quiet thread reads. Web uses quietThreadAtom through shared detail state; desktop wraps web. The equality improvement must not be advertised as measured mobile rendering improvement.
- Benchmark fixtures are deliberately synthetic, and minimal run status records exercise the visibility helper fields rather than transport schema validation.
- This pass did not rerun the repository unit suite or independently validate the two prior code-review verdicts.

FRICTION
- No blocking friction. One exploratory search used a nonexistent visibility-module path; corrected by following the import to packages/shared/src/orchestrationV2Timeline.ts. No effect on evidence or execution.
- The accessor-count phase is much slower than timing batches; it completed normally in the same single run.
- Free space checked by the script at extraction/batch boundaries and every 50 counted events: start 5,608,906,752 bytes, end before cleanup 5,602,140,160 bytes. No below-5-GB condition encountered.
- Both disposable extract directories were removed by the successful script. Audit, exact benchmark copy, downloaded originals, raw rerun results and log remain in /tmp/to-24264-proof2 for the parent.
- No commits, PR edits, pushes, forge reads, browser work, live-data access or delegation performed.