feat: smriti bench — search quality benchmarking

[ashu17706]

Summary

• Adds smriti bench command: a ground-truth quality benchmark inspired by Karpathy's autoresearch pattern
• Produces one combined score (0.0–1.0) to track search quality over time and gate upstream QMD upgrades
• 30 tests, all passing

What's in it

Corpus — 12 realistic sessions across 3 projects (webapp, ml-pipeline, infra), 2 agents, 6 categories. Session s04 has sidecar content (thinking, artifact, attachment) to exercise sidecar search paths.

Query suite — 36 ground-truth queries across 3 tiers:

• Easy (12): exact keywords → tests FTS path
• Medium (12): paraphrased → tests recall path
• Hard (12): conceptual + filters → tests filtered search + cross-session recall

Metrics — MRR, Hit@3, Hit@5 per tier. Combined score formula:

combined = 0.15 * easy.mrr + 0.35 * medium.mrr + 0.35 * hard.mrr + 0.15 * filterBonus

CI thresholds — Easy MRR ≥ 0.70, Medium ≥ 0.25, Hard ≥ 0.12, Combined ≥ 0.30

CLI

smriti bench                                    # full run
smriti bench --profile ci-small --no-perf       # quality only
smriti bench --json --out bench/results/ci-small.json
smriti bench --compare bench/results/ci-small.json

Current scores (ci-small, FTS-only — no embeddings in bench)

Tier	Hit@3	MRR
Easy	91.7%	0.917
Medium	8.3%	0.083
Hard	41.7%	0.417
Combined	—	0.362

Medium is low by design — recall queries need vector embeddings which aren't available in the in-memory bench. As QMD's embedding/reranker models improve (autoresearch loop), this score will rise.

Test plan

• bun test test/bench.test.ts — 30/30 pass
• bun src/index.ts bench --profile ci-small — runs end-to-end
• bun src/index.ts bench --json --out /tmp/bench.json — JSON output works
• bun src/index.ts bench --compare /tmp/bench.json — comparison works

[github-actions]

Benchmark Scorecard (ci-small)

Bench Scorecard (ci-small)

threshold: 20.00%

metric	baseline	current (median)	delta	status
ingest_throughput_msgs_per_sec	1735.800	447.820	-74.20%	WARN
ingest_p95_ms_per_session	6.960	25.482	+266.12%	WARN
fts_p95_ms	0.410	0.639	+55.85%	WARN
recall_p95_ms	0.436	0.777	+78.21%	WARN

Summary: WARN (4 metrics)