fix(#360): command handlers honour ctx.cwd, not process.cwd()

[randomm]

Closes #360.

What

/work (commands.ts:191) and /work-status (work-status.ts:362) resolved the repo root from process.cwd() instead of the ctx.cwd Pi hands the handler. ExtensionContext.cwd is first-class API and resolveRepoRoot(cwd) already took the parameter — the plumbing existed, it was just fed the wrong value.

Why it mattered

Two ways:

1. Production: whenever Pi's process cwd differs from the session cwd, /work writes its state file into the wrong repo.

2. Observable today: because production ignored the context it was given, test-command-flow.ts had no seam to isolate through. Every run of the §1 pre-push gate wrote fixtures into the live state dir:

   .pi/work-state/547.json 548.json 549.json 551.json 561.json 789.json
   

   These sat at "status":"running" with an empty eventLog, so a real /work 547 would have collided with a fixture. 561.json carried "issues":[561,562,563] — the exact multi-issue shape work-driver-context.ts:139 cites as the historical false-MERGED incident. (Correction to the issue text: .pi/work-state/ is gitignored, so these were written-by-accident, never committed.)

Test changes beyond passing a temp cwd

Passing a temp dir alone turned the suite red, for two reasons worth recording:

• The temp dir is git init-ed. resolveRepoRoot falls back to the raw cwd outside a repo, which makes the driver throw early and emit its crash-report sendUserMessage — perturbing the message counts asserted downstream.
• Count assertions now read a promptMessages() view that filters the driver's own pi-ensemble: status lines. The driver is fire-and-forget and shares the sendUserMessage transport with prompt bodies, so every hardcoded count in this file was a latent race on how fast gh issue view failed. Against a real repo gh was slow enough to hide it; against a local temp repo it fails instantly and the race surfaced.

Verification

• Full §1 gate green: bun run build, tsc --noEmit, bun run check, 60/60 offline smoke tests.
• test-command-flow.ts run 3× consecutively — stable, no residue.
• The new guard was proven to fail. With ctx.cwd reverted to process.cwd(), the closing assertion reports ✗ /work cases honour ctx.cwd — repo .pi/work-state/ is unchanged by the suite and the file exits 1. A gate nobody has watched fail is not a gate.
• Lint diagnostics in the touched smoke test: 24 before, 24 after (all pre-existing; bun run check scopes to src/, which is clean).

Out of scope

spawn.ts:162 (spec.cwd ?? process.cwd() — correct fallback, no ctx), and permission-*.ts / session-autosave.ts, which run outside command-handler context where the process cwd is genuinely right.