[Download the sanitized evidence archive](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/475e7be7c096a7c7/issue-24660-runtime-evidence.tar.gz). All evidence paths below are relative to its root.

Fresh runtime evidence for issue 24660 confirms that native lookup acquisition can proceed while a shell reader transaction is still open on [37df35aff6](https://git.cafe/pingdotgg/t3code/commit/37df35aff6b2da81ab99b2219365e45860c7a700), compared with baseline [c79ef8c454](https://git.cafe/pingdotgg/t3code/commit/c79ef8c454e1268ff79da5a9dcb74572e86f7861). Cancellation and subsequent health completed in this workload, including a separately proven four-reader saturation case. This does not establish a universal latency bound or deterministically exercise the narrow scheduler-yield race.

The branch was `fix/issue-24660`, clean at start and finish. Fresh measured requests ran 2026-10-25 06:19:20–06:19:42 UTC. The only diff is in shared SQLite implementation/tests; visual matrix: **EMPTY**. No browser, screenshot, implementation/test change, test run, commit, push, forge action, merge, or watcher action was performed. The supplied 18 passing shared tests were not rerun.

The isolated retained `.t3` fixture was read-only inspected before startup: 4,122 active threads; 1,000,000 completed, runless items; 121 native V2 target creation events; zero legacy-import rows. DB size after shutdown: 270,667,776 bytes. Synthetic count-only projections test query load, not coherent event history. This is much smaller than the reporter's 16–25 GB database. No live userdata was used. All paths below name files inside this evidence archive.

**Before/after observations.** Baseline three shell HTTP requests returned 200 in 962.786/905.991/919.791 ms; native snapshots in 19.769/17.335/11.922 ms. With a shell issued 30 ms earlier, native snapshot+synchronized took 358.669/280.409/319.851 ms. Actual baseline cheap lookup acquisition waits were 219.579/185.269/181.246 ms; overlapping reader transaction holds were approximately 203–241 ms. Baseline span `ebba7491dc5c812a` waited 219.579 ms for the reader held by shell transaction `cad1b7e75b1e64f9` and acquired only after its end. See `before/contention-exact-spans.json`, `before/focused-spans.ndjson`, and `before/load-results.ndjson`.

Fresh head's first measured shell on the new process returned 200 in 1139.930 ms; subsequent warmed serial shells took 839.467/804.752 ms. Native snapshots took 18.503/13.109/12.286 ms. The first overlap, which can include another worker's first use, produced snapshot+synchronized in 25.604 ms; warmed overlaps took 9.361/9.615 ms. Shell responses for those overlaps took 820.793/823.195/850.341 ms. First-use observations are kept separate from warm observations; individual worker startup cost was not independently isolated. These are measurements, not pass thresholds.

The decisive SQL evidence is stronger than HTTP ordering. Native lookup `326d64699df57794` acquired a reader with `db.lock_wait_ms=0.193` and completed **170.634 ms before** shell reader transaction `8c64dc0648534889` emitted its commit marker. Warm lookup `22c7531deb9171aa` completed 185.681 ms before commit of `058b9bc3881b840a`; `fdfe41bc0342c2d8` completed 165.935 ms before commit of `30d331540c0d48ed`. Each lookup begins after the overlapping shell's read body starts, so that shell already owns its reader. See `final/final-overlap-spans.json`, `final/focused-spans.ndjson`, and recomputable `final/final-analysis.json`.

**Cancellation and immediate health.** The two replay batches interrupted 120 subscriptions each after intended 80 ms (actual 89.521/88.919 ms). All 240 returned interrupted terminal exits with `socketReadyState=1`, by 641.982/473.775 ms from batch issue. Snapshot chunks can race cancellation: batches ultimately saw 4/13 chunks. Health began 1/0 ms after the aggregate terminal-exit log, before awaiting any outstanding shell response, without restart/shutdown. Each health sequence returned shell/native HTTP 200, then subscription and shell-overlapped subscription snapshot+synchronized plus interrupted terminal exits on the same open socket. See `final/load-results.ndjson` and `final/rpc-wire.ndjson`.

The separate four-shell run now has a retrospective ownership proof, not a timer asserted as a barrier. All four shell read bodies had begun by nanosecond epoch `1792909178407723659`; the first reader commit was `1792909178731370094`, a 323.646 ms interval holding all four readers. All **120** native lookup spans began and ended interrupted entirely inside that interval and had no acquisition annotation. This demonstrates queued cancellation under four occupied readers. All 120 RPCs returned interrupted terminal exits on the still-open socket by 1391.582 ms; zero chunks arrived. Four shells returned HTTP 200 in 2874.295–2962.264 ms. Health began in the same logged millisecond as the final exits, while shell responses were still outstanding, and all shell/native/subscription/overlap probes succeeded. See `final/saturation-spans.json` and `final/saturation-results.ndjson`.

Across the fresh trace, 401 native legacy lookups comprise 57 successful and 344 interrupted spans; all 344 interrupted lookups lack acquisition annotations. Reader transactions recorded 50 successes and three interruptions; the latter have explicit rollback markers under threadSnapshotThenLive. All 21 shell reader transactions committed successfully. `final/rollback-spans.json` and focused spans preserve the correlation. Health probes show recovery for this run; they cannot prove the absence of every permit leak or specifically force cancellation at the scheduler-yield boundary fixed in this head.

The negative legacy lookup still repeats. Cache absence is outside this diff and belongs to the existing contributor [#24707](https://github.com/pingdotgg/t3code/pull/24707); no forge state was read. Four occupied slots still queue. No permanent-hang claim is supported: baseline already terminated all 240 requests, and this head terminated all 360 intentional load/saturation cancellations.

**Integrity, publication, cleanup, and friction.** Original baseline and provisional directories were SHA-256 checked unchanged. There were zero dev watcher restarts. A post-run mtime audit of 2,571 regular files in packages (including regular cache files), server src, and server package.json found no writes during the server lifetime; git remained clean. Symlink targets were not recursively audited. See `final/stability.json` and `final/server-lifecycle.log`.

Stopped only captured process group 2063223 with SIGTERM at 06:19:54.770 UTC; server 13860 and web 5820 were confirmed free. Retained isolated `.t3` remains for inspection. Public files exclude all credentials, tickets, request headers, databases, and private snapshot bodies. Script copies read auth from the environment. Syntax checks passed for all published drivers; public focused-span analysis reproduced the private raw-trace results. See README.md for every seed/reproduction driver and commands; SHA256SUMS.json covers archive integrity.

FRICTION: First client attempt preceded the listening marker and received ECONNREFUSED; it issued no measured request, and the successful replay began after startup. No restart occurred. The prior provisional live trace barrier failed and issued no cancellations; this run used retrospective read-body/commit overlap instead. Prior health exit-duration logging was fixed only in copied evidence drivers. Public sanitization initially overmatched a JavaScript variable named ticket; this was corrected and every published driver syntax-checked. No evidence-only file was committed.

FRICTION: Python urllib received HTTP 403 when checking the public archive. curl subsequently received HTTP 200 and the downloaded archive matched the local SHA-256 exactly.