Add metile.compile(model): one call, structural detection, verified before it is kept

[AndreSlavescu]

The integration worked but nobody could reach it. Accelerating a model meant importing from
metile.integrations.mlx_lm, choosing among eleven keyword flags, and knowing which combination was safe
for the architecture in hand. This is the Liger-shaped entry point:

import metile
print(metile.compile(model))

meTile on qwen2
  accelerating: attention, rms_norm, graph_fusion, quantized_mlp
  surfaces replaced: mlp, input_layernorm, post_attention_layernorm, block
  verification: logits match MLX exactly

Structural detection is what makes unlisted models work today

The patcher gated on a list of module and class names, which only ever covers what someone remembered to
add — Qwen3.5, Qwen3.6 and Qwen3-VL were all excluded by it, and their equivalence tests reported skips
that read like passes. A class carrying gate_proj/up_proj/down_proj is now a candidate whether or
not it has been seen. gemma2.MLP, never listed, is admitted; llama.Attention is not.

Verification is what makes that safe

Structure is a weaker claim than a name: a class can have the parts of a gated MLP and still scale the
product or use a different activation, presenting identically. So compile runs the model before and
after, compares decode-step logits, and keeps only what reproduces MLX.

Decode steps rather than prefill, because attention only engages at query length one — a prefill
comparison reports agreement while never running the kernel. When the full set disagrees it bisects per
feature rather than reverting everything, which is what keeps three of four features on Llama-3.2-1B
instead of none:

meTile on llama
  accelerating: attention, rms_norm, graph_fusion
  verification: logits match MLX exactly
  declined quantized_mlp: changed the logits by 0.0352, 2.3e-03 relative
                          -- reduction-order scale, raise tolerance to keep it

Exact by default, and the report says what that costs

That 0.0352 is a summation-order difference where meTile measured as the more accurate side — 4.10
against MLX's 18.05 versus a float32 reference. Declining it by default is the conservative call, and the
report names the feature, the absolute and relative size, and that tolerance=5e-3 would keep it. The
trade is visible before it's taken rather than discovered later.

The report is falsy when nothing was replaced

That's the point of it. The dangerous outcome here is not a crash but a silent no-op, and sixteen tests in
this repository reported "skipped" for models meTile wasn't touching for as long as it took someone to read
the skip list. if not metile.compile(model) is now a real check.

Verified on three real checkpoints: Qwen2.5-1.5B and Qwen3.5-4B take all four features with exactly
matching logits; Llama-3.2-1B takes three with the fourth declined and explained.

693 pass. Lint and vulture clean.

🤖 Generated with Claude Code