Speed up web vitest suite (~51s -> ~32s): cut per-file import/jsdom overhead

[Aztec03hub]

Summary

The web unit-test suite's wall-clock was dominated by per-file overhead, not test logic. The same module graph (App, mqtt-store, components, bits-ui, the lucide-svelte icon barrel) was re-imported and re-executed, and a fresh jsdom environment created, for every one of the 94 spec files. Actual test execution was already cheap (~16s).

This PR switches the bulk of the suite to Vitest's vmThreads pool, which runs each spec file in its own fresh V8 VM context (full per-file isolation, no cross-file state leakage) while reusing the worker's compiled-module cache across files. That collapses the dominant import/environment cost. No test assertions or test code changed; the only file touched is web/vitest.config.js.

Levers evaluated

Lever	Result
isolate: false (threads)	Rejected: 111 failures / 14 files. ESM bindings are captured at first import, so the shared vi.mock('api.js') of the mqtt-store leaks across files. Exactly the cross-file coupling the brief warns about.
deps.optimizer.web (lucide / multi-dep)	No measurable gain (within noise); multi-dep made it worse.
pool: 'threads' (whole suite)	Slower than the forks baseline for this jsdom-heavy workload.
pool: 'vmThreads' (applied)	~halves wall-clock with full isolation.

Under vmThreads the VM's window / location globals are non-configurable, so the 3 specs that replace the whole window/location object (api-base-derivation, connection-status, slash-commands) are routed to a standard threads project via test.projects. Specs that only define a property on window (matchMedia, innerWidth, etc.) run fine under vmThreads. A future spec that reassigns the whole global fails loudly under vmThreads, never silently.

Before -> after (matched-load A/B, 3 runs each)

Metric	Before	After	Delta
wall-clock	~51s	~32s	-37%
duration	~47s	~30s	-36%
import	~240s	~174s	-28%
environment	~46s	~5.4s	-88%
transform	~71s	~64s	-10%
tests	~16s	~12s	-25%

The dominant saving is the per-file jsdom environment cost (-88%) plus reduced module re-import.

Stability gate (5 consecutive runs)

All 5 runs: exit 0, 94 files / 1292 tests passed, 0 unhandled errors, 0 stderr lines.

stab1: wall 32.89s  94 files / 1292 tests  stderr 0  unhandled 0  exit 0
stab2: wall 32.24s  94 files / 1292 tests  stderr 0  unhandled 0  exit 0
stab3: wall 32.69s  94 files / 1292 tests  stderr 0  unhandled 0  exit 0
stab4: wall 32.42s  94 files / 1292 tests  stderr 0  unhandled 0  exit 0
stab5: wall 32.56s  94 files / 1292 tests  stderr 0  unhandled 0  exit 0

The bits-ui scroll-lock drain in tests/setup.js (PR #46) and the clean 0-unhandled / 0-stderr output (PR #48) are preserved. pnpm --dir web build stays green. No skips/excludes added; same 1292 assertions.

🤖 Generated with Claude Code