cc-turboquant-index Sidecar Report

This sidecar moves TurboQuant from the Python validation path into Rust packed-code candidate generation. The Python prototype validated distortion and recall, but it pre-dequantized the corpus into fp16 for search. The sidecar keeps rows bit-packed and estimates inner products directly from packed codes. | vectors | dim | bits | queries | recall@10 | per query ms | build ms | ratio fp32 | packed bytes | |---:|---:|---:|---:|---:|---:|---:|---:|---:| | 4096 | 768 | 4 | 32 | 0.878125 | 4.802 | 21.643 | 5.953x | 2113536 | | 4096 | 768 | 8 | 32 | 0.981250 | 4.321 | 22.658 | 2.988x | 4210688 | | 32768 | 768 | 4 | 16 | 0.837500 | 37.417 | 168.610 | 5.953x | 16908288 | Sidecar v0 proves the right implementation direction: packed-code Rust scanning is faster than the Python/Numpy prototype on the same small synthetic shape while preserving the same quality regime. It also exposes the remaining performance wall. A scalar packed scan at `32768 x 768` is already `37.417ms/query`, so the final RAG++ target at hundreds of thousands of vectors requires blocked scanning, memory-mapped snapshots, exact rerank, and Apple Silicon SIMD. This does not directly improve N'Ko ASR CER. It improves the AGP control plane around ASR by making retrieval candidates, semantic state packets, and transfer bottlenecks cheaper to move and search. CER improvement still comes from better ASR checkpoints plus bounded AGP correction decisions admitted by the Rust/Graph Kernel gate.