chore(bench): add analyzer benchmark suite and PR regression CI

[veksen]

Summary

• Adds src/remote/optimizer.bench.ts — a vitest bench suite that measures QueryOptimizer.start() on small (3t/5q), medium (20t/100q), and large (300t/1000q) shapes against a single testcontainer Postgres.
• Adds .github/workflows/benchmark.yaml — runs the bench on both PR HEAD and the merge-base, uses scripts/compare-bench.mjs to diff the two JSON reports, posts/updates a single PR comment with a regression table, and fails the check when any bench's mean regresses by >20%.
• Wires up npm run bench and a benchmark.include pattern in vitest.config.ts.

Test plan

• First PR run emits a "no baseline" comment (base commit predates the bench file) — confirms graceful fallback.
• Subsequent PRs show a diff table with 🔴/🟢/⚪/🆕 verdicts and RME deltas.
• npm run bench works locally.

🤖 Generated with Claude Code

[github-actions]

Query Doctor Analysis

View full run details

83 queries analyzed

2 pre-existing issues

• SELECT "guests"."id", "guests"."session_id", "guests"."username", "guests"."avatar_path", "guests"."color", "guests"."side", "guests"."audio_recording_path", "guests"."audio_recording_public", "gue...
  index assets(event_id, inserted_at desc)
  cost 31,003,449 → 1,493 (100% reduction)
• SELECT * FROM guest_ip_addresses WHERE ip_address = '127.0.0.1';
  index guest_ip_addresses(ip_address)
  cost 154,402 → 8 (100% reduction)

_{Using assumed statistics (10000000 rows/table). For better results, sync production stats.}

[github-actions]

Benchmark comparison

Threshold: ±20% on mean. 🔴 regression · 🟢 improvement · ⚪ within noise · 🆕 new/removed.

	Benchmark	Base mean	PR mean	Δ	RME (base → PR)
⚪	src/remote/optimizer.bench.ts > query optimizer > large (300 tables, 1000 queries)	163,677ms	163,558ms	-0.1%	±2.9% → ±2.5%
⚪	src/remote/optimizer.bench.ts > query optimizer > medium (20 tables, 100 queries)	2,575ms	2,584ms	+0.4%	±5.0% → ±7.0%
⚪	src/remote/optimizer.bench.ts > query optimizer > small (3 tables, 5 queries)	89ms	93ms	+4.6%	±10.2% → ±16.9%

Benchmarks use testcontainers + wall-time; some noise is expected. Treat single-digit deltas as not-significant.
Base commit: ``