qwen38-mtp-head-q2q4-cons

Qwen3.8-27B MTP q2-q4-rerank head, conservative-coded 4-bit stage

Rebuild of amal-david/qwen38-mtp-head-q2-q4-rerank-v1@ae62827 for the mlxfast Qwen-MTP challenge. The affine-2 shortlist tensors, precision-island tensors, and all norms are byte-identical to that artifact. The eight 4-bit/g64 decoder triples (fc, q/k/v/o_proj, gate/up/down_proj) — verified byte-identical in the original to a naive mx.quantize(w, 64, 4) of the organizer-pinned bf16 head — are replaced with a per-group best-of-two coding of the SAME bf16 source: each group stores whichever of {classic coding, bf16-consistent coding} has the smaller true as-stored reconstruction error. No range shrink; min/max endpoints and outlier anchoring are pointwise identical to the original; every group is at least as faithful by construction (56% of groups improve, mean reconstruction error −5.1–7.4% per tensor).