docs: document mlx-lm backend for Apple Silicon and known gotchas

[EmreCelenli]

Summary

Adds an mlx-lm section to docs/local-openai-compatible-providers.md, plus two troubleshooting rows for
real issues hit during setup: bare model names 404 against Hugging Face (needs full HF repo ID), and a
known unfixed mlx-lm bug (ml-explore/mlx-lm#973) where <|im_end|> leaks into output.

Anti-slop triage

• Classification: docs-only
• Evidence: verified end-to-end on a 16GB M1 Mac running mlx_lm.server + claw.
• Non-destructive review result: merge candidate

Verification

• All commands in the new section run end-to-end as written.
• git diff --check passes.
• No secrets, tokens, or unrelated churn included.

Resolution gate

• Doesn't resolve a tracked issue, documents a gap found via direct testing.
• Intended to merge.
• No automation-lane merges/closes without owner approval.

[1716775457damn]

Nice docs-only addition! Documenting the bare-model-name 404 gotcha and the <|im_end|> leakage bug (mlx-lm#973) saves others significant debugging time. Verified end-to-end on a real device makes this trustworthy. LGTM for merge.

[1716775457damn]

Solid documentation addition — the gotchas section (tool calling limitations, batch size quirks) will save users hours of debugging. Thanks @EmreCelenli!

[EmreCelenli]

Thanks for the review, @1716775457damn! I’m looking into issues and fixes from the MLX repo.

[1716775457damn]

Great to see this merged! If you're tracking MLX upstream fixes, the <|im_end|> leakage bug (mlx-lm#973) is the most impactful one to watch — once that lands upstream, the tool calling section of this doc could be updated to remove the stop-token workaround note. Keep us posted on what you find.

[1716775457damn]

One more gotcha worth documenting in a follow-up: mlx-lm's default chat template varies between model families (e.g., Qwen vs Llama 3), and using the wrong template silently degrades multi-turn coherence even when single-turn responses look fine. The --chat-template flag in mlx_lm.server can address this, but the mapping from model ID to correct template isn't always obvious. Might be worth a note in the gotchas section alongside the stop-token issue.