fix(server): flush reasoning parser at stream end (supersedes #653)

[janhilgard]

Supersedes #653 with the CI failure resolved.

Background

@Thump604 opened #653 on 2026-07-28 to flush parser-buffered terminal text (e.g. Gemma 4's truncated <|channel> markers) before emitting the terminal chunk of a stream. The fix is correct and needed — Gemma 4's finalize_stream() already implements the fallback, stream_chat_completion just wasn't calling it.

CI on that PR broke four existing streaming tests:

tests/test_server.py::TestStreamChatCompletion::test_reasoning_stream_emits_structured_tool_calls
tests/test_server.py::TestStreamChatCompletion::test_reasoning_stream_redirects_gemma4_tool_marker
tests/test_server.py::TestStreamChatCompletion::test_reasoning_stream_skips_tool_parser_until_markup_appears
tests/test_server.py::TestStreamChatCompletion::test_response_format_stream_promotes_reasoning_json_to_content

All four failed with AttributeError: 'FakeReasoningParser'/'ReasoningOnlyParser' object has no attribute 'finalize_stream'.

Root cause

Those tests declare inline test fixtures (FakeReasoningParser, ReasoningOnlyParser) that duck-type extract_reasoning_streaming and reset_state but don't inherit from ReasoningParser. The base class provides finalize_stream() as a documented no-op, so any parser that does inherit gets the default automatically. #653 called reasoning_parser.finalize_stream() unconditionally, which broke the duck-typed fixtures.

The alternatives are:

1. Add hasattr / getattr guard in _resolve_reasoning_stream_delta.
2. Add finalize_stream = lambda self: None stubs to every fake fixture.

I picked (1). The base class's own docstring establishes finalize_stream as an optional extension point ("Default implementation is a no-op"), so the calling code should tolerate parsers that don't implement it — that keeps the interface contract loose for third-party parsers as well, not only tests.

Change

One-line change in _resolve_reasoning_stream_delta (vllm_mlx/server.py):

# Before (from #653):
final_delta = reasoning_parser.finalize_stream()

# After:
finalize = getattr(reasoning_parser, "finalize_stream", None)
final_delta = finalize() if callable(finalize) else None

Everything else is @Thump604's original diff, unchanged. Authorship on this commit is preserved as Thump604 <...>; I'm listed as Co-Authored-By for the guard change.

Validation

Ran the 5 relevant tests locally (Python 3.14, macOS ARM64, uv-installed dev extras against fresh clone):

tests/test_server.py::TestStreamChatCompletion::test_reasoning_stream_emits_structured_tool_calls PASSED
tests/test_server.py::TestStreamChatCompletion::test_reasoning_stream_redirects_gemma4_tool_marker PASSED
tests/test_server.py::TestStreamChatCompletion::test_reasoning_stream_skips_tool_parser_until_markup_appears PASSED
tests/test_server.py::TestStreamChatCompletion::test_response_format_stream_promotes_reasoning_json_to_content PASSED
tests/test_server.py::TestStreamChatCompletion::test_gemma_stream_flushes_pending_content_before_length_finish PASSED
5 passed in 1.72s

The four previously-failing tests now pass without modification — the guard tolerates duck-typed fixtures. The new test (@Thump604's Gemma flush regression) still passes as designed.

Not claimed

Same non-claims as #653:

• Preserves template selection, sampling, request limits, scheduler behavior.
• Does not change Jobs integration, model configuration, constrained decoding.
• Does not claim every length-bound Gemma generation produces final content — only preserves parser-buffered terminal content when it exists.

Followup

Please close #653 after merging this. The intent is identical; this is just a rebased single-commit variant with the CI regression fixed and the base-class no-op contract respected.

Closes #653.

[janhilgard]

Closing in favor of #653 — I pushed the same fix directly to the PR branch (amended into the original commit, preserving @Thump604 as author with co-author credit for the guard change). See #653 for the updated PR.