# Final source-freshness/evidence audit — PASS

**MERGE** recommendation for [#24263](https://git.cafe/pingdotgg/t3code/pulls/24263), head [6fc097e](https://github.com/pingdotgg/t3code/commit/6fc097e420b550849c74c780e746b6c0624b6a56) (`6fc097e420b550849c74c780e746b6c0624b6a56`), merge base/current main [26c3625](https://github.com/pingdotgg/t3code/commit/26c3625f484eadd4d5b6ca5a3dd1f8f8bd5533b4) (`26c3625f484eadd4d5b6ca5a3dd1f8f8bd5533b4`). Full final diff and relevant callers reviewed afresh; no confirmed new blocker. Source checkout remained clean/read-only. Remote CI/review state not refreshed in this source-only audit.

**40,225 exact comparisons, zero differences** (4,574 capture + 6,176 extended prefixes + 29,475 generated integration comparisons). **29 focused tests pass**. New anchor/table/entity/fence fixtures use current-main HTML normalization; pipe/malformed delimiter fixtures remain explicit. Full fallback fixtures retain **2 → 2 parses**. Historical failures use their old baseline, separately labeled.

Whole HTML implementation matches current main, including reverse splice edits; production file regex unchanged. Native patch/archive/both harnesses match [db2b6e1](https://github.com/pingdotgg/t3code/commit/db2b6e143b2ea8b80efe9fea221500d658e500fe) byte for byte, so prior actual narrow Linux C++ and HOST MODEL ONLY evidence is reused with hashes, not claimed newly executed.

Fresh host stream: **27 updates, 81 → 54 parses, 2896.152 → 1632.552 ms median (43.63% lower)**. Two warmups/five samples; 1024 UTF-16 units/update; 27,100-byte final input, 29,100-byte max output; virtual 50ms cadence. Full documents use two warmups/eight samples. Actual production regex, parse counts, raw timing samples, ~3,000 malformed lexical-fragment measurements and scaling are recorded. No device-latency or blanket complexity claim.

JS distinct-stage contents **108,200 → 79,200 logical bytes**; caller pair **56,200** both sides; not heap/RSS/liveness. Reused narrow C++ cache: **269,246,464 → 8,388,288 accounted bytes, 512 → 8 entries**; not native total memory. Reused queue **HOST MODEL ONLY**: **134,479,872 → 524,288 modeled source bytes**, 512 → 1 burst renders, both apply final 512. Stalled main queue completed captures still accumulate (134,479,872 modeled bytes); no global two-revision bound or native speed claim.

[REPORT.md](/tmp/takeover-24263/evidence/REPORT.md) contains methods, full measurements, reproduction and limits. [fixture-matrix.json](/tmp/takeover-24263/evidence/fixture-matrix.json), [audit-check-results.json](/tmp/takeover-24263/evidence/audit-check-results.json), [integration-source-checks.json](/tmp/takeover-24263/evidence/integration-source-checks.json), [SHA256SUMS](/tmp/takeover-24263/evidence/SHA256SUMS) provide machine-checkable outcomes/provenance. Report generator uses fresh data and regenerates this file. Historical evidence is preserved outside the active evidence directory.

**UNVERIFIED:** iOS native cache, ObjC harness/runtime and device rendering/layout/native memory (Linux lacks Foundation/device); Android rendering (emulator unavailable/prohibited). No source edits, subagents, uploads, comments, commits, pushes, merges, browser, server, emulator, remote machines or userdata access.

FRICTION: stale premerge evidence and hardcoded claims required regeneration; historical checks needed their original oracle; initial overlapping timing run replaced with solo audit capture. Dependency-patch whitespace warnings are unified-diff context; focused lint exits 0 with the inherited HTML reverse warning. No unresolved scoped-host evidence discrepancy. PASS applies to final integration/source freshness and scoped host proof, not native release validation.