Rotated combined Pareto: research payoff, deployable

Perf record: 029-rotated-pareto.json. Cohere v3, 1M × 1024, cosine, Granite box. --quant binary --rotate 2 --batch 16 --scan-bits N --rerank C.

What we did

Combine the deployable, popcount-speed pieces of the RaBitQ research — the free rotation (026, recall up at no cost; rescues prefix in 027) — with the existing batch winners (tiling 016, prefix 014) and re-draw the frontier. This is the practical answer to "did the research make it better?": rotated-symmetric keeps full popcount throughput, unlike the slow RaBitQ estimator (028).

Result — the frontier moves out at every tier

Conclusions

1. Rotation dominates the prior frontier — for free. At every operating point it delivers higher recall at equal-or-better QPS. The rotation costs nothing at serve time (baked into codes; a tiny per-query FWHT) and nothing in memory.
2. The standout is the prefix tier: 768-bit/C=2000 goes 0.986 @ 762 → 0.994 @ 815 (+8 pts recall and +7% QPS), because rotation rescued truncation (027). High recall at a truncated scan is now real — the prefix lever finally pays at the 0.99 tier, not just below 0.97.
3. New best operating points vs the whole project:
   • 0.996 @ 844 QPS (full, C=1000) — beats 020's 0.993 @ 851 on recall at ~same QPS.
   • 0.999 @ 730 QPS (full, C=2000).
   • 0.994 @ 815 QPS (768, C=2000) — same recall as the original 009 (0.994) at 815 vs 585 QPS (+39%).

The arc

Original 009: 0.994 @ 585. Post-throughput-work (020): 0.993 @ 851. Post-research (029): 0.996 @ 844 or 0.994 @ 815, i.e. higher recall and higher throughput than either — from a free random rotation grounded in RaBitQ/ITQ.

Caveats

• Rotated-symmetric (popcount-speed). The RaBitQ estimator (028) would push recall-per-C further but needs the fast SIMD kernel first.
• tile=16, reps=5, CV < 0.2%; spot box, compare within table.