Keep assistant tool_calls and tool messages on the MLLM chat path

[waybarrios]

Summary

Fixes #608. On the MLLM path, _build_mllm_chat_messages rebuilt every message as {role, content} and dropped any message without renderable text. An OpenAI assistant message with content: null + tool_calls was removed entirely, and role:"tool" messages lost their tool_call_id. The rendered continuation prompt came out byte-identical to the fresh prompt, so agent clients looped on the same tool call forever.

What changed

All in _build_mllm_chat_messages (shared by chat() and stream_chat(), so both routes are covered with no call-site changes):

• Assistant messages carrying tool_calls survive even with empty text. Arguments are normalized str -> mapping via the same normalize_messages_for_chat_template helper the batched engine path wraps, so both paths render identically. reasoning_content is forwarded when present.
• role:"tool" messages keep their tool_call_id. Empty tool results (legit for action tools) survive with content: "" instead of being dropped, so template forward-scans can still pair calls to responses.
• Everything else is byte-identical to before: image/audio/video part ordering, plain-text messages, and the drop rule for genuinely empty messages (verified pre/post on the same inputs).

SUPPORTS_NATIVE_TOOL_FORMAT on Gemma4ToolParser (issue suggestion 3) is intentionally left out of scope.

Tests

5 new regression tests in tests/test_mllm_message_ordering.py (written first, 3 failed on main for the exact bug reasons):

• assistant {content: null, tool_calls} survives with normalized arguments
• tool messages keep tool_call_id + content
• empty tool results survive with their anchor
• a 4-message tool exchange rebuilds strictly more messages than the fresh 3-message turn
• genuinely empty messages without tool_calls are still dropped

pytest tests/test_mllm_message_ordering.py -q   -> 8 passed
pytest tests/test_mllm*.py tests/test_server.py -> 256 passed
pytest tests/ -k "not slow"                     -> 2185 passed, 0 failed
ruff check                                       -> clean

Live verification (issue repro)

Served mlx-community/Qwen3-VL-8B-Instruct-4bit with --mllm --enable-auto-tool-choice --tool-call-parser hermes, M4 Max:

• Probe A (fresh: system + user + 1 tool): tool_call emitted, prompt_tokens = 166
• Probe B (continuation: + assistant{content:null, tool_calls} + tool{result}): prompt_tokens = 213 > 166 (history now contributes tokens), no repeated tool_call, final answer uses the tool result: "The weather in Paris is currently 19°C with sunny conditions."

On main, probe B renders the same prompt as probe A and re-issues the identical tool call.

[waybarrios]

Full test logs from the branch:

Regression tests (verbose) — 8 passed

$ pytest tests/test_mllm_message_ordering.py -v
tests/test_mllm_message_ordering.py::test_mllm_chat_messages_preserve_text_only_shape PASSED [ 12%]
tests/test_mllm_message_ordering.py::test_mllm_chat_messages_preserve_image_text_order PASSED [ 25%]
tests/test_mllm_message_ordering.py::test_mllm_chat_messages_preserve_audio_text_order PASSED [ 37%]
tests/test_mllm_message_ordering.py::test_mllm_chat_messages_keep_assistant_tool_calls_without_text PASSED [ 50%]
tests/test_mllm_message_ordering.py::test_mllm_chat_messages_keep_tool_call_id_on_tool_messages PASSED [ 62%]
tests/test_mllm_message_ordering.py::test_mllm_chat_messages_keep_empty_tool_results PASSED [ 75%]
tests/test_mllm_message_ordering.py::test_mllm_chat_messages_keep_tool_exchange PASSED [ 87%]
tests/test_mllm_message_ordering.py::test_mllm_chat_messages_still_drop_empty_messages_without_tool_calls PASSED [100%]
============================== 8 passed in 0.25s ===============================

MLLM + server suites — 250 passed

$ pytest tests/test_mllm_cache.py tests/test_mllm_continuous_batching.py \
    tests/test_mllm_message_ordering.py tests/test_mllm_mtp_routing.py \
    tests/test_mllm.py tests/test_server.py -q
====================== 250 passed, 13 deselected in 5.18s ======================

Full fast suite — 2185 passed, 0 failed

$ pytest tests/ -q -k "not slow"
=============== 2185 passed, 11 skipped, 23 deselected in 30.95s ===============

$ ruff check vllm_mlx/models/mllm.py tests/test_mllm_message_ordering.py
All checks passed!

Live repro verification (Qwen3-VL-8B-Instruct-4bit, M4 Max)

$ vllm-mlx serve mlx-community/Qwen3-VL-8B-Instruct-4bit --mllm \
    --enable-auto-tool-choice --tool-call-parser hermes --port 8077

# Probe A — fresh turn (system + user + 1 tool), temperature 0
prompt_tokens: 166
tool_calls: [{"id": "call_0a25c249", "type": "function",
              "function": {"name": "get_weather", "arguments": "{\"city\": \"Paris\"}"}}]
content: ''

# Probe B — continuation (+ assistant{content:null, tool_calls} + tool{result})
prompt_tokens: 213        # > 166: history now contributes tokens
tool_calls: null          # no repeated call — loop broken
content: 'The weather in Paris is currently 19°C with sunny conditions.'

On main, probe B renders byte-identical to probe A (166 tokens) and re-issues the same tool call.