# Runtime evidence: PR [#23751](https://git.cafe/pingdotgg/t3code/pulls/23751)

BEFORE [4eae8ae](https://github.com/pingdotgg/t3code/commit/4eae8ae011e2e56ca20159c13f3fb8a2ac49a991); AFTER [19f6543](https://github.com/pingdotgg/t3code/commit/19f654326598cc1f7411a56dbb07e651b6b5f254). Clean shared checkout, no edits/commits/pushes/PR edits. All databases synthetic, created only in this evidence directory. No real T3 database was accessed.

Correctness passes on the documented synthetic matrix. The performance claim is mixed: the thread SELECT improves, but grouped counts are slower and the complete snapshot/subscription does not improve in this capture. Historic author timings are not reproduced proof.

| Boundary | Before ms | After ms | Interpretation |
| --- | ---: | ---: | --- |
| Actual selectShellThreadRows, warm median | 136.16 | 44.72 | Query plus client row materialization; source timing wrapper |
| Actual getShellSnapshot active, warm median | 208.59 | 219.65 | Full store read + decode, 998 active shells |
| Actual subscribeShell first snapshot, warm median | 212.82 | 223.01 | In-process handler boundary; no transport |
| Isolated matching/missing interrupt expressions | 72.96 | 1.69 | Earlier 903,000-item fixture; predicates use synthetic latest IDs |
| Isolated correlated/grouped counts, warm median | 87.98 | 111.39 | Grouped slower; earlier 903,000-item fixture |

Migration 068: actual runMigrations on 903,000 existing synthetic items 220.31 ms; returns only [68, ShellStopRequestIndex]. Restart 4.42 ms returns []. Turn-item SHA256 unchanged before/after. Index definition and two covering-index stop_request searches verified in complete production SELECT plan. Grouped query uses covering thread/run searches plus outer GROUP BY temp B-tree. No universal speedup claim.

Initial getShellSnapshot first-in-process BEFORE/AFTER 315.08/304.24 ms. Initial subscription 259.20/255.37 ms. All repetitions, source hashes, exact queries, numbered logs and deterministic drivers are included. Timings are wall clock on a shared machine, not statistically controlled benchmarks.

| Proof row | Result at recorded SHAs | Evidence |
| --- | --- | --- |
| P01/P02 latest+blocking match/no-match | PASS actual SELECT predicates for latest-only/blocking-only/both/neither and full before/after shell equality | verify.raw.txt; plans-migrate.raw.txt |
| P03 no runs | PASS empty count0 and runless-only count1; snapshot equals getThreadShell at both refs | verify.raw.txt |
| P04 counts | PASS completed/running/rolled-back/missing/runless; whole/scoped/empty result equality | plans-migrate.raw.txt; verify.raw.txt |
| P05 forks | PASS inherited runless prefix and boundary (1303), nested fork (1704), archived external-project source, missing/deleted source, cycle; individual and whole shell equality | augmentation.json; verify.raw.txt |
| P06 many active | PASS correctness, 1000 total/998 active shells, 9002 runs, 903003 items, 650 running-run threads; measured full runtime (performance improvement NOT established) | compare.raw.txt; timings.json |
| P07 existing-data migration | PASS real runner68 timing, ledger apply-once/restart, unchanged item hash, exact partial index and two searches | plans-migrate.raw.txt |
| P08 preview | PASS preview53 470.47ms, preview54 no-cleanup 460.77ms, cleanup 439.54ms; rollback/retry391.57ms; unknown later history rejected, imported progress/created_at/data preserved, rerun empty, two index lookups | preview.raw.txt |
| P09 subscription | PASS actual in-process initial handler + following committed/published event; before/after stream items identical and delta equals individual shell. Transport UNVERIFIED | *subscribe-matrix.raw.txt; delta-compare.raw.txt |
| P10 provenance/resources | PASS full driver SHA256 manifest; selected-source hashes; numbered SHA-qualified lines; limits obeyed | SHA256SUMS; sources.json; transcript-numbered.txt; resources-before-cleanup.json |
| P11 two-lookup assertion | PASS production plan has exactly two new-index searches; forcing either SQL branch onto old index makes the equivalent assertion reject | mutation.raw.txt |

Exact remaining unverified scope: P02 has no deliberately cross-thread item carrying another thread's valid run ID (same-thread unrelated requests and ordinary matching-run items are tested). P05 does not newly exercise a fork whose selected source run is itself rolled back (ordinary rollback count exclusion is tested). P06 does not recreate the historical ~5000-active-thread database, provider-session/error/secret/pending-work skew, real history payload decoding, or production reader worker. P09 real WebSocket/RPC serialization/network/ACK, repository enrichment cost, remote/relay and browser clients remain unverified. No repeat of the previously reported 41 scoped tests is claimed; this capture uses standalone runtime drivers, not a new suite run. Mutation proof operates on SQL plans rather than editing/rerunning the checked-in test. Preview timing is a synthetic 1000-item upgrade per lineage.

Resources before cleanup: own DB files 309.30 MiB; entire own directory 319.05 MiB; all evidence 720.34 MiB; free 17.16 GiB (df reports18G). Initial inventory ~379MiB other reviewer material; never removed or modified it. Max own runtime fixture ~304MiB, under500MiB target; total below1.5GiB cap and free always >=6GiB. Resources-after-cleanup.json will prove owned DB deletion after successful uploads; text/drivers retained.

FRICTION: local main was stale (invalid first comparison preserved as stale-local-main.raw.txt); corrected to origin/main merge-base. One initial HEAD capture lacked index and is labeled head-without-index.raw.txt, excluded from final timings. SQL extractor initially passed top-level return into TypeScript stripping; fixed. Direct Node strict assertion exposed null-prototype SQL rows; scalar/plain-row assertions fixed without changing app code. Delta schema initially received JSON text instead of parsed object; fixed before successful captures. These harness failures are not product findings. Shared-machine and OS-cache noise limit timing conclusions. Full transport was deliberately not launched; actual handler stream is the reported boundary. No browser/server or live-state access, installs, tsgo or npx tsc.