[flexdinesh]

• Behavior owner and resulting change:
  • cmd/diffx/local.go (acquireLocalInput): local mode rejects a checkout path that is not an existing directory before serving. Since fix: start local server before collecting workspaces #76/feat: stream local collection and render available snapshots #79 moved collection after startup, a typo such as diffx change started a foreground server that never exited, which hung TestCollectorOptionsRejectAmbiguousInputsBeforeStartup for Go's 10m timeout (or passed by accident when a developer's diffx was running). The test now isolates runtime/config/state and runs each case under a 10s deadline; TestLocalInputRejectsMissingCheckoutPathBeforeStartup covers missing, file and repeated paths. TestLocalInputDefaultsToCurrentDirectory now creates its --path fixture directories.
  • Faster checks:
    • Browser tests: global setup builds the web app and CLI once; each Playwright worker owns an in-memory fixture server behind a Vite preview (SPA fallback, API proxied), so specs run fullyParallel at 33% of cores with no shared review state and no fixed port (DIFFX_TEST_PORT removed). The 4000px full-render inline-comments test is marked test.slow().
    • mise.toml: suites after check:ci (contracts, JS, release, Go) run concurrently; check:push runs the Go suite once, with the race detector, instead of plain then race.
    • Go tests time out after 3m (5m race) so hangs fail fast.
  • TestRemoteRunShutdownClosesActiveEventStream flaked under load (3/40 on origin/main with -race): a keep-alive client could leave a spare, never-used connection that http.Server.Shutdown waits on for 5s. The test client now disables keep-alives; 80/80 pass under load.
• Contracts/invariants affected; local/remote implications: local startup input validation only; later Sync of a removed path still reports a discovery diagnostic. Test/tooling changes otherwise; pnpm dev still starts its own fixture via go run.
• Validation results; remaining risks:
  • Pre-push mise run check:push passed in 3m56s (previously ~12m), with 88/88 browser tests in 2.9m, while another project's Playwright suite loaded the machine (load avg ~30-50). A run at lower load with warm caches took 96s.
  • Risk: concurrent suites raise peak CPU; timing-sensitive tests may surface on slower machines. Worker count is capped at 33% for headroom.
• Intentional architecture/design changes and rationale: docs/system.md documents the path rule; docs/development.md documents per-worker browser servers, concurrent suites, race-only Go run in pre-push and Go timeouts.
• UI evidence: n/a.