Benchmark: OpenClaw builtin memory-core comparison

[bm-clawd]

Part of #10

Add the OpenClaw builtin memory search as a comparison provider in the benchmark suite.

Requirements

• Index the same test corpus with OpenClaw's builtin memory-core provider
• Run the same queries through the builtin memory_search tool
• Score with the same metrics (Recall@K, MRR, Precision@K)
• Side-by-side comparison output: BM vs builtin per category

Implementation options

1. Preferred: Start an OpenClaw instance with builtin memory pointing at the benchmark corpus, query via the memory tools
2. Alternative: Replicate the builtin's chunking + embedding + hybrid search logic in a standalone script (more work, but no OpenClaw dependency for CI)

Key comparisons

• Semantic queries: both use vector search, should be similar
• Relational queries: BM has knowledge graph, builtin has only text
• Exact fact queries: builtin has BM25 hybrid, BM has FTS — compare
• Task queries: BM composited search scans tasks, builtin does not
• Context efficiency: BM returns structured observations, builtin returns raw chunks

Depends on

• Benchmark: build realistic test corpus from agent memory patterns #11 (corpus)
• Benchmark: eval harness for Basic Memory search #12 (eval harness)