docs(spike): S3 query-aware fusion — validate as already-shipped + neutral

[dongwhee]

S3's feature (query-type weighted RRF: _classify_query +
_fusion_weights, identifier -> bm25 up-weight) was already implemented,
unit-tested, and default-on — shipped in b81d57b (2026-04-27, before
the roadmap was written) and live since hybrid default-on (#58). Same
stale-roadmap pattern as R1/R2/R4/C7. The missing piece was the
isolated A/B the gate calls for.

Adds benchmarks/query_aware_fusion_ab.py: per reader_comparison case
(3 identifier / 3 concept — real identifier coverage, so not vacuous),
fetch end-to-end twice flipping only _fusion_weights between the
shipped query-aware values and equal weights via monkeypatch (no
production toggle); graded metric = top-1 identity + first-fact rank,
measured rerank-on and rerank-off.

Result: end-to-end neutral (net facts delta +0, top1 changed 1/6
marginal). At the pure-fusion layer the identifier cases — the
weighting's target — are identical between weighted and equal RRF
(rankers concur at top, so the 5x swing can't reorder; the lever does
fire elsewhere, confirming this is real invariance not a dead patch).
Gate (coding net >= +1) not met.

Decision: keep as-is. No toggle (unrequested config, no rollback need),
no simplification (touches the load-bearing retrieval path for zero
end-to-end benefit; tested + harmless), no weight tuning (the
structure, not the magnitude, makes it inert here). S3 closed as
already-implemented + validated-neutral; retrieval.py unchanged. Design

• outcome + cited-paper reconciliation in
  docs/superpowers/specs/2026-06-06-query-aware-fusion-validation.md.

🤖 Generated with Claude Code