[TimelordUK]

Stacked on #485. GitHub retargets this to master when #485 is merged.

npm run record-corpus -- <tester.js>

A corpus scenario can now be a script (test/contract/scripts/*.js): an async function that gets one connection and helpers that log in the corpus format. It's for what a single-query scenario can't say: several cases in sequence, or a repro from an issue.

module.exports = async t => {
  t.section('A running cancel')
  t.query('A', BIG, null, { at: { row: { 100: ['cancel'] } } })
  await t.settled('A')
}

The command:

1. copies the tester into scripts/, so its source is kept beside its transcript;
2. runs it 3 times, each in its own process, and reports whether it is deterministic;
3. writes corpus/scripts/<name>.txt and its native recording.

From then on the real-server contract test, the replay and the jittered-timing replay all cover it. The helpers are documented in test/contract/scripts/README.md.

The issue repros

dhensby's repros for #478 – #481 are recorded this way as scripts/issue-478 … -481. Whether an issue is fixed is now a diff on its own transcript. They pin what happens today (unreviewed):

#480	fixed: 51 + 51 rows at pause rows 50, 51 and 52
#481	fixed: strict order on every error path; B2's cancel works
#479	points 4 and 5 fixed. A (done before the cancel callback), B (an extra done on a late cancel) and C (a cancel after a PRINT cancels nothing) are not
#478	not fixed: info 3621 before error 2627 in A – C

As the fixes land, each transcript gets edited to the agreed sequence and marked agreed.

Also

• The fake sql module inherits the real one's helpers (PollingQuery and the rest).
• capture.js --exact <name>.
• The jittered-timing replay in CI includes scripts/.

Testing (Windows, localdb)

• Contract plus replay suites: 209 passing.
• Replay alone, including the 4 scripts plain and over two seeds: 126 passing.
• Explorer over the scripts: 4 of 4 held over 4 timings.

🤖 Generated with Claude Code