fix: real benchmarks, Cypher multi-row fix, and honest README

[aepod]

Summary

• Fix Cypher Issue CypherEngine: MATCH RETURN produces single row — context.bind() overwrites previous bindings #269: MATCH queries now return all matching rows instead of collapsing to 1
• Real benchmarks: Independent comparison against hnswlib (C++) and numpy brute-force with ground-truth recall
• Fix misleading README claims: "100% recall" and "simulated Python baseline" corrected with real numbers

Cypher Fix (Issue #269)

The execute_match() function collapsed all match results into a single HashMap — each context.bind() call overwrote the previous value. Three matches → only the last survived.

Fix: Implemented proper ResultSet pipeline (Vec<ExecutionContext>) threaded through all statement executors. Cross-product expansion for multiple patterns in one MATCH.

Test results: 30 passed, 0 failed, 2 ignored

Key tests:

• test_match_returns_multiple_rows — 3 nodes → 3 rows (was returning 1)
• test_match_return_properties — Alice + Bob both returned correctly
• test_match_where_filter — WHERE correctly filters multi-row results
• test_match_many_nodes — 100 nodes → 100 rows

Also fixed column name generation for property expressions (n.name instead of ?column?).

Real Benchmark Results

All measurements are real — recall measured against brute-force ground truth, no simulated competitors.

10,000 Vectors (128d, M=32, ef=200)

Engine	QPS	Recall@10	Build (s)	p50 (ms)
hnswlib (C++)	1152.6	0.9895	7.5	0.73
ruvector-core	443.1	0.9830	44.0	1.98
numpy brute-force	134.8	1.0000	0.003	3.26

100,000 Vectors (128d, M=32, ef=200)

Engine	QPS	Recall@10	Build (s)	p50 (ms)	Memory
hnswlib (C++)	249.5	0.7427	395.3	2.57	—
ruvector-core	85.7	0.8675	855.6	10.14	~523MB
numpy brute-force	69.2	1.0000	0.016	10.20	48.8MB

Key Findings

• Recall is competitive: 98.3% at 10K (within 0.65% of hnswlib), 86.75% at 100K (higher than hnswlib's 74.27%)
• QPS is 2.6-2.9x slower than hnswlib — expected since hnswlib has SIMD optimizations
• Build time is 2.2-5.9x slower — the largest gap, narrows at scale
• Previous benchmarks had hardcoded recall=1.0 and memory=0.0 — now corrected

README Corrections

• Replaced "100% recall at every configuration" with actual measured recall ranges
• Replaced "15.7x faster than Python" (simulated) with real hnswlib comparison
• Added independent benchmark section with link to full methodology

Test plan

• All 30 Cypher tests pass (0 failures, 2 ignored)
• Multi-row MATCH returns correct number of rows
• Property expressions generate correct column names
• WHERE filtering works on multi-row results
• No regressions in existing test suite
• Benchmarks reproducible with included scripts and raw JSON data

🤖 Generated with claude-flow