PP-QUANT-001 Q2 (gguf/inference CPU): retype the boundary — raw u32 qtype → GgmlType; ratchet 10 files → 0 #3437

Parent: #3421 · Slot: 0.70.0, after M1+M2 (ratcheted: if it does not fit the train, the remainder rides the next tag with a slipped_from: note).
Mechanism: acceleration.rs names 13 ids, matmul*.rs 10, forward/{single,batch,results}.rs 3, traced.rs/ffn_block.rs 1 — the same tensor is accepted by one path and refused by its neighbour.
Fix: parse u32 → GgmlType once at the loader boundary; everything downstream matches exhaustively on the enum. One PR for this directory only.
Falsifier: a non-increasing ratchet on files naming GGUF_TYPE_* or matching a bare integer qtype in this directory (baseline 10, measured on a672677f4); adding a new raw-u32 match → RED. Qwen2.5-Coder tok/s within APR-PERF-GATE-001 on the merge sha, same host.

Ĵ: 24 turns [A] — borrowed; estimate.sh aprender pools 0 of 47 ledger rows today, so there is no measured basis.
Plan + evidence: #3427 · design discussion: #3418

🤖 Generated with Claude Code