Deferred from PR #6 review (Qodo): TILE_FLOOR=128 deterministically idles workers on wide pools at mid-size batches. Intended tradeoff for stream-sharing (each chunk streams the full sidecar once; measured 7.7k→10.2k q/s at 2048-query batches, 1.26M×1024). A configurable or bandwidth-adaptive floor is worth designing if rerank-bound wide-pool workloads materialize. Related: multi-pass tiling fallback for extreme nq×m in ordvec's streamed core (deferred from ordvec#278).
Deferred from PR #6 review (Qodo): TILE_FLOOR=128 deterministically idles workers on wide pools at mid-size batches. Intended tradeoff for stream-sharing (each chunk streams the full sidecar once; measured 7.7k→10.2k q/s at 2048-query batches, 1.26M×1024). A configurable or bandwidth-adaptive floor is worth designing if rerank-bound wide-pool workloads materialize. Related: multi-pass tiling fallback for extreme nq×m in ordvec's streamed core (deferred from ordvec#278).