feat(query): make the hybrid semantic weight configurable (default unchanged)

[salishforge]

What changed

The 1.5× multiplier on the semantic arm's RRF contribution was a hardcoded constant with a large, undocumented effect: with RRF K=60, it means semantic ranks 1–31 all outscore a keyword-only rank-1 hit — lexical matches the vector arm missed are close to invisible. Now HYBRID_SEMANTIC_WEIGHT, documented in the config reference.

The default is deliberately unchanged, and that is the finding.

The experiment, and why the answer is "leave it alone"

Pre-registered before running: sweep on a stratified 42-question sample (7 per category), then validate the winner on a disjoint 42-question sample selected the same way. Adopt only if it beats the 1.5 baseline on held-out R@5 without regressing R@1 by more than 2pp.

weight	train R@5	test R@5	train R@1	test R@1
0.75	95.2%	88.1%	76.2%	76.2%
1.50 (current)	92.9%	88.1%	69.0%	81.0%

On train, 0.75 looked clearly better across every k — the monotonic trend was tidy enough to be convincing, and any single-sample tuning would have adopted it. It did not replicate. On held-out data R@5 is identical and R@1 is 4.8pp worse.

Every metric flips direction between the two samples. That's the signature of sampling noise, not signal: at n=42 one question moves a metric 2.4pp, and the observed gaps are one to two questions.

So the pre-registered rule fails on both conditions, and the weight stays at 1.5.

Why this is still worth merging

• The constant is now measurable per deployment. The optimum plausibly depends on the embedding model — this was measured with local bge-small, and a stronger embedder would likely shift it.
• The K=60 interaction is documented where someone changing it will see it.
• Resolving the question properly needs the full 500-question run, and the knob is what makes that a config change rather than a code change.

Gates: type-check, lint, integration 24, http 43, retrieval-determinism 3, causal-graph 39 — all green.

🤖 Generated with Claude Code

https://claude.ai/code/session_011xCqQo49d3CEbn6oEvb3Ru