Benchmark: eval harness for Basic Memory search

[bm-clawd]

Part of #10

Build the eval harness that runs queries against real Basic Memory and scores results.

Requirements

• TypeScript, runs via bun
• Uses real bm search CLI (not mocked)
• Indexes the test corpus into a BM project
• Runs all queries from queries.json
• Scores: Recall@5, Recall@10, MRR, Precision@5
• Groups results by query category
• Outputs results as JSON + human-readable table
• just benchmark in the justfile

Implementation notes

• Set up a temporary BM project for the corpus
• Run bm search via CLI subprocess for each query
• Parse JSON output, compare against ground truth
• Also test bm context for relational queries (unique to BM)
• Measure latency per query (secondary metric)
• Clean up temp project after run

Stretch

• Compare composited memory_search (MEMORY.md grep + BM search + task scan) vs BM search alone