test: de-flake the search stress benchmark under parallel suite load

[resure]

What changed

The "keeps a 4000-note keystroke burst responsive" test in src/search.stress.test.ts asserted indexed <= onTheFly * 1.1 + 5 from a single timing sample of each path. Under a parallel full-suite vitest run, another worker preempting one burst mid-run inverted the comparison (observed: on-the-fly 37.0ms vs indexed 66.8ms → fail; 2 of 3 full-suite runs on 2026-07-05). It always passed in isolation.

The test now:

• Interleaves best-of-5 runs of each path and keeps the minimum — the min discards preempted samples, and interleaving keeps the comparison fair under drifting load. On-the-fly still runs first in each pair so the indexed path can't be flattered by a cold JIT.
• Widens the slack to × 3 + 10ms — absorbs the observed ~1.8× scheduler-noise inversion, but a real order-of-magnitude regression in the indexed path (the thing the 2026-06 perf-audit guard, merged via afa46c2, exists to catch) still fails loudly.
• Keeps the absolute < 1500ms responsiveness budget unchanged, and renames the assertion wording from "never slower" to "within noise of on-the-fly" to match.

Why

The test's intent is a regression guard on the precomputed-lowercase index (searchNotes's lowerById), not a strict microbenchmark race — the old bound made it a coin flip under machine load.

Reviewer notes

• Verified: passes in isolation and in two consecutive full-suite runs (868/868), where the best-of-5 numbers show the index winning comfortably (1.16× and 1.25× faster). Lint + typecheck clean.
• The extra runs add ~300ms of test time.
• The third test in the file (cappedMs < uncappedMs) is the same class of single-sample comparison but with a ~7–9× real margin, so it's far less exposed — left untouched to keep this scoped to the reported flake.

🤖 Generated with Claude Code