bench: add distilled-knowledge retrieval corpus (30 docs, 30 QA pairs)

[LavaDMan]

Summary

All existing benchmarks test episodic memory — conversational recall
(LongMemEval, LoCoMo, MemBench, ConvoMem). This PR adds a benchmark for a
different memory category that currently has no coverage:

Semantic / procedural memory — retrieving engineering constraints,
architectural decisions, and system invariants from prose documents. The kind of
knowledge engineers distill into runbooks, architecture decision records, and
internal wikis.

What's in this PR

• benchmarks/distilled_knowledge_bench.py — self-contained benchmark script
  (no external data download; corpus embedded in the file)
• benchmarks/distilled_knowledge/README.md — methodology and results
• benchmarks/BENCHMARKS.md — updated to document the new benchmark

Corpus: 30 prose documents covering cross-cutting engineering constraints:
concurrency / locking, database driver gotchas, package security gates, process
isolation, lifecycle state management, and error handling edge cases.

QA pairs: 30 paraphrased queries. Question wording deliberately avoids the
exact vocabulary in the target document to test semantic similarity, not keyword
matching. Each pair includes a paraphrase_note documenting the specific
vocabulary shift.

Results on raw ChromaDB baseline

Measured with MemPal v3.3.x, ChromaDB ephemeral client, default
all-MiniLM-L6-v2 embeddings:

Metric	Score	Count
Recall@1	76.7%	23/30
Recall@5	100%	30/30

Recall@5 is perfect — every correct document appears in the top-5. The ~23pp
gap to Recall@1 is consistent with the pattern seen on LongMemEval and LoCoMo;
hybrid retrieval or reranking would likely push R@1 into the 90s.

How to run

# No data download needed
python benchmarks/distilled_knowledge_bench.py
python benchmarks/distilled_knowledge_bench.py --top-k 1 --verbose

Relation to existing benchmarks

This is additive, not critical of the existing suite:

Benchmark	Memory type	Query style
LongMemEval / LoCoMo / MemBench / ConvoMem	Episodic	"What did X say?"
This PR	Semantic	"How does X work?"

Real production systems accumulate both types. A memory system that handles
episodic recall perfectly but cannot retrieve "why do we use file locks instead
of asyncio.Lock?" is incomplete for agentic use cases.

🤖 Generated with Claude Code