Verify + land #542 PQ/IVFADC within-list pruning (BET5, ADR-206) — scale-gated WIN finding

[ruvnet]

Verify #542 — PQ/IVFADC within-list pruning (BET5, scale-gated WIN, ADR-206)

Assessment: like #537, #542 arrived as a complete, pre-registered research artifact — not a nightly PoC needing a 6-step build-out.

• Implement — self-contained benchmark crate ruvector-bet4-ivf-bench (data/kernel/pq/pca/oracle + IVF/PQ pruning sweeps). By design it does not change shipping code — it establishes the finding.
• Test — pre-registered, frozen WIN gates: tests/pq_gate.rs (3) + tests/oracle_gate.rs (3).
• Benchmark / finding — ADR-206: PQ/IVFADC within-list pruning is a scale-gated WIN over tuned plain IVF nprobe (wins at scale; the cluster-level bound was a NO-GO in ADR-205/SepRAG BET 4 (finding): IVF cluster-pruning is structurally redundant with tuned nprobe — NO-GO #540, redundant with nprobe).
• Rigor — docs/plans/bet5-ivf-pq/PRE-REGISTRATION.md frozen before any run.

Loop verification (this issue)

• Merged PR BET 5 (SepRAG #534): PQ/IVFADC within-list pruning vs tuned IVF nprobe — scale-gated WIN (ADR-206) #542 onto current main (resolved a Cargo.toml members conflict).
• cargo test -p ruvector-bet4-ivf-bench → 6/6 gate tests pass.

Recommendation

Land #542 as the documented finding (benchmark crate + ADR-206). The genuine productionization follow-up — wiring PQ within-list pruning into the production IVF index with a scale-gated runtime switch — is a substantial separate effort the finding sets up but does not itself do (correctly, per its scope). Not a 5-minute step; flagged as the real next work.

Source: #542 · ADR-206 · SepRAG #534. Loop job 05bbd4fb.