docs(adr): ADR 0039's change reaches Metal too, through QuantizedMatmul

[glennneuber]

This adds a dated amendment to ADR 0039: its change reaches Metal too, through QuantizedMatmul. Docs only, one file.

The sentence it corrects. ADR 0039's consequences say "Metal is unaffected in behaviour: its GatherQMM still receives m × 2688". They add that Metal's 26b and 31b encoder numbers move only for mlx#3912, and that "this change does not interact with that". The Metal host pointed out in #414 that the claim rests on GatherQMM alone. The ADR's own wrapper table says otherwise for the dense path.

Checked at 0.34.0 (8a7ba949):

• The loader stored f32(m) × 2688, and scaleAndCast divided by 2688 in float32.
• Only GatherQMM takes a Metal branch. Among the scale wrappers, the only MetalIsAvailable branch is GatherQMM's (x/mlxrunner/mlx/ops_extra.go:137). The loader has none.
• The other MetalIsAvailable hits are a benchmark, the runner and gemma4 model code. The vision tower's only makes the attention output contiguous, so its dense linears take the same path on every platform.
• So QuantizedMatmul, which carries every dense nvfp4 linear, took the round trip on Metal too, and ADR 0039 removed it there.

It was measured. ADR 0037, which is about Metal, records 31b's golden delta going 0.1406 → 0.1094 with mlx#3912. With the round trip removed it reached 0.0898, which is MLX-CUDA's value exactly. Metal's goldens on the v0.34.4 release read 0.0898 (#404).

What changes:

• The original sentence stays.
• A bolded, dated amendment goes under it, as ADR 0012 and ADR 0043 do.
• The status line gets a one-line note, as ADR 0033's does.
• No other document repeats the claim.

Checks:

• check_source_paths.py --changed-since origin/main is clean.
• The name scan is clean.
• The relative link to ADR 0037 resolves.

ai-server/mlx-cuda

🤖 Generated with Claude Code