fix: fall back to openai-harmony rendering for gpt-oss models without chat_template

[FaisalFehad]

Summary

gpt-oss models whose packaged tokenizer does not ship with a chat_template — several community MLX quants of the context-1 family, for example — fail every chat-style endpoint (/v1/chat/completions, /v1/messages, /v1/responses) with:

Chat template error: Cannot use chat template functions because
tokenizer.chat_template is not set and no template argument was passed! ...

even though oMLX already has full Harmony infrastructure (adapter, streaming parser, anthropic→harmony converter, structured tool-call extraction). The only endpoint that responds at all is /v1/completions — and that path drops the parser's structured tool_calls on the floor, which makes tool use on these models effectively unusable.

Fix

When model_type == "gpt_oss" and the tokenizer has no chat_template, render the prompt directly via openai-harmony (already a dependency) instead of falling into tokenizer.apply_chat_template. Models that ship a template continue to use it — nothing changes on the happy path.

Files

• omlx/adapter/harmony.py — new render_harmony_prompt(). Translates OpenAI-style messages + tools into a Harmony prompt string. Handles:
  • content blocks ([{"type":"text","text":"..."}])
  • assistant tool_calls → commentary channel with to=functions.NAME
  • role=tool responses, resolving the function name via tool_call_id lookup when the message omits name (orphan tool messages are skipped with a warning rather than given a fabricated name)
  • reasoning_effort and conversation_start_date via chat_template_kwargs
  • Encoding cached at module scope (one load per process, not per call)
• omlx/engine/batched.py::_apply_chat_template — branches to the new renderer only when the fallback is needed. Single guarded if at the top of the method; existing path untouched.
• tests/test_harmony_render.py — 19 unit tests covering basic rendering, content-as-list, tool rendering, tool-call roundtrip (incl. name resolution via tool_call_id and orphan-skip), template kwargs, and malformed input.

Verification

• 19/19 new tests pass.
• Full harmony test suite (67/67) passes unchanged.
• End-to-end against a local gpt-oss model without chat_template:
  • /v1/chat/completions — returns structured tool_calls.
  • /v1/messages — returns structured tool_use blocks.
• Regression-checked against gemma-4-e4b-it-8bit (ships a chat_template) — unchanged behaviour, new branch not taken.

Test plan

• pytest tests/test_harmony_render.py — 19 passed
• pytest tests/test_harmony.py tests/test_harmony_parser.py tests/test_harmony_render.py — 67 passed
• Live /v1/chat/completions call on gpt-oss-no-template model → structured tool_calls
• Live /v1/messages call on gpt-oss-no-template model → structured tool_use
• Live /v1/chat/completions call on model with template → unchanged