feat(bench): cold-start benchmark script (measurement only, no CI guard)

[sebastianbreguel]

Summary

Adds tests/cold-start-benchmark.ts — a measurement-only script that spawns node start.mjs N times, polls each child's MCP readiness sentinel, and reports p50/p95/p99 of spawn-to-ready latency plus skip count.

Measurement only. No vitest assertion, no budget, no CI gate. Goal is to give the project a real cold-start baseline so a follow-up PR can add a regression guard with empirically-grounded thresholds rather than guessed ones (same measure-then-tune sequencing as #356).

This addresses cold-start latency, which #270 explicitly listed as a priority (with a Windows call-out).

What it measures

process spawn → start.mjs self-heal layers → ensure-deps → import server.bundle.mjs → server.connect(transport) → sentinel writeFileSync (src/server.ts:2412-2416).

Detection uses the existing PID-stamped sentinel (context-mode-mcp-ready-<pid> via sentinelPathForPid() from hooks/core/mcp-ready.mjs). No new public APIs, no changes to the readiness contract.

Local baseline (macOS arm64, Node v25.9.0, bundle prebuilt)

Metric	Value (ms)
ok-count	10
skip-count	0
min	272.0
p50	290.5
p95	349.0
p99	349.0
max	349.0

$ npm run bench:cold-start
Context Mode — Cold-Start Benchmark
====================================
Node:        v25.9.0
Platform:    darwin (arm64)
Bundle:      PRESENT
Iterations:  10 (warmup: 1)
Timeout:     30000ms per iteration
Warming up (1 iteration, discarded)...
  warmup 1: 282.6ms
  iteration 1: 349.0ms
  iteration 2: 305.6ms
  ...

Scope (deliberately small)

Included:

• Cross-platform spawn via process.execPath + start.mjs resolved relative to script dir
• Sentinel polling via existsSync at 10ms cadence (deterministic across platforms vs fs.watch)
• 30s per-iter timeout → SKIP semantics (partial data > no data)
• SIGTERM → 500ms grace → SIGKILL cleanup chain (server's own graceful shutdown handler unlinks its sentinel)
• SIGINT cleanup hook on the bench script itself — no orphaned children if the bench is killed mid-run
• Env tunables: ITERATIONS (default 10), WARMUP (1), TIMEOUT_MS (30000), POLL_MS (10)

Explicitly NOT in scope (deferred to follow-up PRs):

• Regression guard test with budget — needs the baseline numbers from this PR first
• First-ctx_search end-to-end latency via MCP stdio roundtrip — needs MCP client SDK plumbing; double-scopes this PR
• Windows-specific cold-start audit — separate PR; this script will be the measurement tool that audit uses

Caveats (in script header + below)

• Requires server.bundle.mjs present for representative numbers. Without the bundle, start.mjs falls into the npx tsc --silent first-build branch (slow). Run npm run bundle first.
• "Cold per spawn" means fresh node process per iteration — not fresh disk state. Disk caches stay warm across iterations. True ice-cold (post-rm -rf node_modules) is environmental and out of scope.
• Wall-clock noise on shared-runner CI is real. p95 (not p99) is the headline number.

Test plan

• Happy path: 10 iterations on macOS arm64 with bundle prebuilt → p95 = 349ms, 0 skips, no orphaned node start.mjs processes after run
• Timeout path: TIMEOUT_MS=50 → all iterations record SKIP (timeout), summary handles all-skip with exit code 1
• Cleanup: pgrep -f start.mjs after each test run shows only pre-existing dev MCP servers, no orphans from the bench
• Typecheck clean (npm run typecheck)
• Linux Node 22 — owner CI matrix
• Windows Node 22 — owner CI matrix (sentinel uses tmpdir() on Windows via existing sentinelDir())

Bench output is paste-ready into a follow-up PR body so the regression-guard PR can quote real numbers.