perf(open): Cut the fixed delays before the harness starts

[4ndreello]

codedeck open spent about 950 ms before the harness started, and most of it was fixed waiting inside CodeDeck, not Claude Code:

• the boot animation slept 19 x 40 ms before the spawn;
• ensureDaemonStarted slept 200 ms before its first socket check, while the daemon accepts in 60 to 80 ms;
• every open ran claude twice before the launch: which plus claude --version in resolveBinary (78 ms, version unused) and claude --append-system-prompt-file in assertSupport (176 ms).

Changes:

• playBoot derives its frame delay from a 200 ms budget (190 ms total). Same frames, same final banner, non-TTY path untouched.
• ensureDaemonStarted polls every 25 ms until a 6 s deadline. Same error, same budget.
• resolveBinary scans PATH in process and keeps detectBinary as the fallback. assertSupport records support in <base>/claude-support.json, keyed by realpath, size and mtime, only when the probe prints the missing-argument error. Unknown option, ENOENT and other failures are never recorded.

Spec: .specs/features/open-boot-latency/spec.md (R1 to R10).

Known limitation: when claude is a version-manager wrapper (mise, asdf, volta), realpath is the wrapper, so upgrading the real binary does not invalidate the record. Accepted because the probe only guards a long-standing flag and the real launch still reports the error. Documented at the lookup.

Measurements

Fake claude on PATH, isolated RUN_AGENT_DIR, pty via script(1), open general --effort high --no-worktree, 7 runs per build, base 252485b vs this branch:

Scenario	Before (median)	After (median)	Saved
warm daemon, default theme	955 ms	356 ms	599 ms
warm daemon, --no-theme	222 ms	155 ms	67 ms
cold daemon, --no-theme	714 ms	533 ms	181 ms

Fake claude invocations per warm open: 3 before (version, probe, launch), 1 after (launch).

Tests

Scoped batches on the pre-rebase HEAD:

• npx vitest run tests/open-boot.test.ts tests/ipc-daemon-start.test.ts tests/open-claude-probes.test.ts: 11 passed; the size-key test added afterwards passed in npx vitest run tests/open-claude-probes.test.ts (6 passed)
• npx vitest run tests/open-: 281 passed
• npx vitest run tests/ps.test.ts tests/claims-cli.test.ts: 44 passed
• npx tsc --noEmit -p .: clean; npm run build: ok
• R4 real-time test under 16 CPU burners on 12 cores: 20/20, p95 28 ms against the 50 ms bound
• Mutation probes caught: final boot frame skipped, poll at 60 ms, poll at 100 ms, 5 s deadline, early return removed, detectBinary called first, mtime or size dropped from the key, support recorded on ENOENT, on unknown option, or on any probe failure

Not covered: Claude Code's own time to input-ready with and without the plugin (needs the authenticated account), scripts/pty-gate.sh and the theme and rename gates, and the codex and opencode launchers, which keep the same probe pattern.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud