No wall-clock speed asserts in tests: one budget helper that is load-proof, and the gate refuses raw "under N ms" expects #4429

Probe 2026-10-21 14:45Z: release 0.7.232 was held because its main check failed GET /settings/cloud/audit names its phases and answers under 50 ms (2 of 5446) - main's next check on 42d30b1a is green; 0.7.230 was held the same way by a timing-dependent test. #3069 fixed /map only; main now has 7 tests titled "answers under N ms" (the Server-Timing series adds one per page) - each a coin flip in a loaded check cell, and they now hold releases.

Done:

1. One helper, e.g. expectBudget(fn, { ms, against: baseline }): measures relative to a same-run baseline (or counts work: queries, rows, bytes) and only asserts absolute ms when an env says the cell is quiet (FORGE_TEST_QUIET=1, set by a nightly perf lane, not the pull gate).
2. Every existing "answers under N ms" test moves to it.
3. scripts/gate.ts (or a test-lint) refuses a new expect(elapsed).toBeLessThan(...)/under N ms pattern outside the helper, with a message pointing at it.
4. Real p95 budgets live in production Server-Timing/RUM numbers on /boards, where they mean something.

Test: the 7 tests pass 20/20 beside 6 busy cells; a raw ms expect in a new test fails the gate.

Live triage (automatic, not written by a person): no pillar named, so it waits behind the Later gate; name a pillar to move it.