Patch the newest models too, so their equivalence tests stop skipping

[AndreSlavescu]

Sixteen cases in the model matrix were skipping, not passing — and a skip reads like a pass in a
summary. Qwen3.5-4B, Qwen3.5-9B, Qwen3.6-27B and Qwen3-VL-4B reported "patches nothing", so their
logits agreed with MLX for the uninteresting reason that meTile was not running.

No kernel was missing — a name check was

The allowlist was keyed on a set of modules plus a hardcoded class name of "MLP". Every
architecture involved computes the same gated MLP — down_proj(swiglu(gate_proj(x), up_proj(x))),
where swiglu(gate, x) is nn.silu(gate) * x, which is exactly what _execute_quantized_mlp
computes — but they don't all call the class MLP:

model	MLP class	block class
Qwen3.5 / 3.6	qwen3_next.Qwen3NextMLP	qwen3_5.DecoderLayer
Qwen3-VL	qwen3.MLP	qwen3.TransformerBlock

Membership is now explicit (module, class) pairs in two registries, because a claim about an
implementation shouldn't be spelled as a claim about a name.

Graph fusion keeps a stricter registry: its replacement reproduces a specific residual structure.
Qwen3.5's DecoderLayer is included only because _attention_module now resolves the attention by
attribute — that layer is hybrid, with a GatedDeltaNet bound to linear_attn on every layer
that isn't a multiple of full_attention_interval and no self_attn at all, so a replacement
calling self.self_attn would have crashed. Resolving per call rather than per class is the whole
adaptation; a block binding neither name falls back instead of guessing.

Skips: 16 → 0. All ten newly active cases pass bit-exact on the first run — both token equality
and decode logits.

Lighting them up surfaced a flake the skips had been hiding

In a test this change doesn't touch: Qwen3-VL attention failed bit-exactness in one of five full
matrix runs, while passing in isolation and passing with its own model's cases run alone. Needing
the whole matrix is the signature of memory pressure — and this file never released a model.
Eighty-two tests each load a checkpoint, up to 15 GB, and letting the Python reference go does not
return device memory, because MLX caches freed buffers. A fixture now drops both after every test.

Six full matrix runs since are clean, and consistently faster — 227–235s against 294–304s —
which is independent support for the diagnosis rather than for the fix merely hiding it.

Not proof. Five clean runs against a one-in-five rate is likely but not decisive. The mechanism,
the timing change, and a recorded precedent in this project all point the same way.

82 pass with no skips; 655 in the fast suite.

🤖 Generated with Claude Code