perf: fix keyword search returning ≤1 result and Redis-absent stalls

[salishforge]

Found by running the corrected benchmark. Both are real defects, not tuning knobs — and neither trades quality, security, or functionality for a number.

1. Keyword search matched almost nothing

queryKeyword used plainto_tsquery, which ANDs every lexeme. "what kind of exercise did I say I do" becomes 'kind' & 'exercis' & 'say', and a single conversation session rarely contains all of them.

Measured on the real corpus: AND matched 0 rows where OR matched 16. Keyword mode returned a median of one result for any natural-language question — and since hybrid fuses the two arms, hybrid was effectively running semantic-only.

Rewriting the operators to OR restores graded matching, with ts_rank_cd discriminating: rows matching more query lexemes, closer together, outrank weak single-term hits. Precision comes from ranking + LIMIT rather than from refusing to match. The rewrite runs through plainto_tsquery's own output, so stemming, stop-words and escaping stay Postgres's job — no user input reaches query syntax.

Recall@k (sessions)	before	after
keyword R@3	60.0%	83.3%
keyword R@5	60.0%	92.9%
keyword R@10	60.0%	95.2%

Latency cost: p50 5ms → 9ms.

2. An absent Redis stalled every cached read

A failed connect costs ~7.5s (connectTimeout plus the reconnect ladder), and the attempt left no memory of itself — so /query and /stats re-paid it on every request, while the docs advertise Redis as optional with graceful degradation.

Circuit breaker: after a failure, skip Redis until a cooldown expires, then allow one probe. Consecutive queries measured 7.6s → 0.016s → 0.007s. Behaviour with Redis present is unchanged.

Benchmark methodology fixes found along the way

• Limited runs were category-biased. LongMemEval-S stores instances in contiguous blocks by question type, so BENCHMARK_LIMIT=50 sampled 50 single-session-user questions and nothing else. Tuning retrieval against that would have optimised for whichever category sat at the offset. Adds deterministic stratified sampling (default for partial runs; BENCHMARK_SAMPLE=sequential restores the old slice).
• Partial runs could masquerade as official. The generator overwrites RESULTS.md in place, so a 20-question smoke test silently replaced the published page in the official format. Partial runs now carry a not-an-official-result banner naming sample size and strategy.

Stratified baseline (n=42, 7 per category)

	keyword	hybrid
R@1	69.0%	69.0%
R@5	92.9%	92.9%
R@10	95.2%	92.9%
p50	9ms	28ms

Deliberately not done

Two things I measured but did not tune, because n=42 is too small to justify moving a constant and doing so would be fitting to the benchmark:

• Hybrid R@10 (92.9%) is below keyword alone (95.2%). Semantic's 1.5× RRF weight can crowd out correct keyword hits. That weight was calibrated when the keyword arm returned ~1 result, so it likely wants revisiting — but on a full run, not this sample.
• single-session-preference is the weakest category (71.4% keyword / 85.7% hybrid), independently confirming WB-03 in NEXTGEN-RECOMMENDATIONS.md.

Also noted: tests/memory-namespaces.test.ts hangs — verified pre-existing on main, and it is absent from the npm test list, which is why it went unnoticed.

Full suite green: integration 24, http 43, security 29, epistemic 33, causal 39, abstractions 39, bootstrap 22, contested 9, metrics 17, cache-degradation 5.

🤖 Generated with Claude Code

https://claude.ai/code/session_011xCqQo49d3CEbn6oEvb3Ru