fix(lsp): deflake "streams events" test — await turn_done, not a timer

[oratis]

Problem

apps/lsp/src/handler.test.ts > handleMessage — executeCommand > returns a turnId for deepcode.runAgent and streams events is flaky in CI: it failed on the macOS runner while passing on ubuntu for the same commit (#92), and a plain re-run went green — a classic timing flake. A flake here red-fails the whole Typecheck + Lint + Test matrix (ubuntu + macos run pnpm -r test) and blocks merges.

Root cause

handleRunAgent fires the agent run as fire-and-forget (void (async () => {…})()) at apps/lsp/src/handler.ts:131. That async path lazily import('@deepcode/core') several times and resolves credentials before it can emit the turn_done event. The test polled for turn_done with a fixed 50 × 20ms = 1s budget (handler.test.ts:48-55). On a loaded macOS runner those imports blow past 1s, so the poll loop exhausts before the event arrives and the toContain('turn_done') assertion fails — the 5000ms vitest timeout never even came into play.

Fix

Wait on the real completion signal instead of a clock: resolve a promise from the send callback the moment the turn_done agentEvent is observed.

let signalDone!: () => void;
const done = new Promise<void>((resolve) => { signalDone = resolve; });
const send = (m) => {
  out.push(m);
  if (m.method === 'deepcode/agentEvent' && m.params.kind === 'turn_done') signalDone();
};
// …
await done;

• Finishes as soon as turn_done fires (~250ms locally), no longer racing a hardcoded 1s cap.
• Per-test timeout raised 5000→15000ms purely as a safety net for a wedged CI runner — it is not what the test waits on (no blind timeout bump to mask the race).
• Error path covered: with no DEEPSEEK_API_KEY, the handler's catch still emits turn_done (handler.ts:195-199), so done always resolves.

Test-only change — no production code touched.

Verification

20/20 clean runs of pnpm --filter @deepcode/lsp test, each ~260ms, 8 tests passing.

🤖 Generated with Claude Code