fix(ci): unblock main build (stop_reason refs + omni-test CPU collection)

[krishung5]

Summary

1. Drop dead stop_reason refs (f804b56255)

Main is currently failing the dynamo-llm build with error[E0609]: no field stop_reason on type &mut ChatChoiceStream at lib/llm/src/preprocessor.rs:1581. ChatChoiceStream (lib/protocols/src/types/chat.rs:813) only has index, delta, finish_reason, logprobs — no stop_reason.

PR #9058 (f6117b82ec, landed earlier today) introduced two dead assignments to a field that was already removed in PR #8119.

2. Skip omni test collection on runners without vllm_omni (9e086c21c0)

sample-runtime / Unified Test cuda12.9, amd64 runs pytest -m "pre_merge and gpu_0 and unified" on a CPU-only runner with no path filter, so pytest walks the whole repo and imports every test_*.py before applying marker filters. components/src/dynamo/vllm/tests/omni/test_*.py transitively imports vllm_omni at module-load time; on platforms vllm_omni doesn't support, that raises something other than ImportError, so the per-file try/except ImportError: pytest.skip(...) doesn't catch it.

Extends the existing pytest_ignore_collect hook in components/src/dynamo/vllm/tests/conftest.py (which already skips test_vllm_*.py when vllm is unavailable) to also skip files under an omni/ directory when vllm_omni is unavailable.

Test plan

• vllm-runtime / Build multi-arch cuda12.9 and dynamo-runtime / image / Build multi-arch cuda12.9 go green
• sample-runtime / Unified Test cuda12.9, amd64 goes green
• No change in behavior on runners that DO have vllm_omni installed — omni tests still get collected and run

🤖 Generated with Claude Code

[coderabbitai]

Walkthrough

This PR adjusts streaming choice data handling during the end-of-stream flush operation. The preprocessor now clears reasoning-related delta fields instead of stop_reason, and the test helper is updated to align with this field-handling change.

Changes

Stream Flush Behavior for Reasoning Fields

Layer / File(s)	Summary
Flush path refusal and reasoning field clearing lib/llm/src/preprocessor.rs	In the end-of-stream flush branch of strip_leading_reasoning_start_from_stream, the emitted buffered choice now explicitly clears choice.delta.refusal and choice.delta.reasoning_content instead of clearing choice.stop_reason.
Test helper alignment lib/llm/tests/postprocessor_parsing_stream.rs	The mock_multi_choice_content_chunk helper no longer initializes the stop_reason field on ChatChoiceStream to reflect the change in how stop_reason is handled during flush operations.

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~3 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Title check	⚠️ Warning	The title mentions 'omni-test CPU collection' but the raw summary shows only stop_reason fixes; the second change (pytest collection) is not reflected in changed files.	Update title to accurately reflect only the stop_reason fix, or provide context for the omni-test change mentioned in description.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description check	✅ Passed	The PR description is comprehensive and well-structured, clearly explaining the problem, root cause, and proposed fix.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[Aphoh]

LGTM