Cap transformers <5.13 to keep mlx-lm importable on macOS (v0.3.19)

[codelion]

Problem

On Apple silicon, import optillm crashes with:

AttributeError: 'str' object has no attribute '__module__'
  at transformers/models/auto/auto_factory.py:680 in register()

Root cause: transformers 5.13.0 tightened AutoTokenizer.register() to require a class (it does key.__module__). mlx-lm 0.31.3 (latest) registers a custom tokenizer by string name:

# mlx_lm/tokenizer_utils.py:505
AutoTokenizer.register("NewlineTokenizer", fast_tokenizer_class=NewlineTokenizer)

transformers ≤5.12.1 tolerated the string; 5.13.0 raises. Since optillm declares transformers>=5.0.0 with no ceiling, a fresh install pulls 5.13.0.

macOS-only: optillm installs mlx-lm only under platform_machine=="arm64" and sys_platform=="darwin". Linux CI never installs mlx-lm, so it silently ran 5.13.0 fine — which is why CI stayed green while local Mac installs broke.

Fix

Pin transformers>=5.0.0,<5.13.0 (resolves to 5.12.1). Verified: mlx-lm then imports cleanly.

The transformers 5.13.0 HF path itself is fine — I confirmed dhara-250m loads and generates on 5.13.0. The incompatibility is purely mlx-lm's; lift the cap once mlx-lm ships a 5.13-compatible release.

Testing

• Reproduced the crash with transformers==5.13.0 + mlx-lm in a clean venv.
• Verified transformers>=5.0.0,<5.13.0 resolves to 5.12.1 and import mlx_lm succeeds.
• Confirmed the dhara HF path (tokenizer + generate) works on both 5.12.1 and 5.13.0.