test(#100): fix e2e flakiness — condition-based polling instead of fixed sleeps

[lis186]

Closes #100.

Root cause (5 Whys)

The proxy writes logs from a spawned child process. Several e2e tests read index.ndjson / log files after a fixed await setTimeout(500) and then assert. Under the parallel full-suite run (node --test test/*.test.js), many e2e files each spawn a real server, saturating CPU/IO, so the child's async write sometimes lands after the guessed delay → the read finds nothing → expected OpenAI index entry and friends. Each test passes in isolation; the suite failed a different subtest run-to-run.

The codebase already had the right pattern (waitForPort polls /_api/health; waitForIndexEntry polls the index) — it just wasn't applied to these sleep-then-read sites.

Change

• Add poll-until helpers: waitFor + readIndex (startup.test.js), waitForIndexEntries (websocket-proxy.test.js), modeled on the existing pollers.
• Convert the sleep-then-read sites to wait for the actual condition (index entry present, file count reached, sysprompt API / SSE snapshot state) instead of a fixed delay.
• Bonus: faster fast-path — an early write returns immediately rather than always sleeping 500ms.

Scope: only the proven read-gating waits. A few unrelated fixed delays remain (error-recovery health checks, inter-turn WS pacing, teardown settle) — they aren't the observed flaky class and the terminal condition-poll now absorbs upstream timing.

Verification

CCXRAY_HOME=$(mktemp -d) node --test test/*.test.js — 6 consecutive runs, 947 pass / 0 fail, zero flaky, ~48s each. Files also pass in isolation (startup 62/62, ws 17/17).

Note: while verifying I hit an unrelated artifact — running the suite back-to-back through the npm/shell-pipe wrapper occasionally failed to return even though node --test itself completed and printed its summary. Running node --test directly exits cleanly every time. Not caused by this change; flagging in case it recurs in CI.

🤖 Generated with Claude Code

[lis186]

Codex 二審：no blocker. One should-fix addressed in f46cbea.

• should-fix — waitFor awaited fn() before checking the deadline, so an async predicate that never settles (e.g. httpGet to a dead proxy) could hang the test instead of timing out. Fixed by racing fn() against the remaining budget (with clearTimeout so the loser timer doesn't linger), plus a 5s socket timeout on httpGet so a silent no-response rejects and frees the socket.

Re-verified: 9 consecutive full-suite runs via node --test test/*.test.js (empty CCXRAY_HOME), all 947/0, zero flaky.