Fix vLLM inference test failures for reasoning models

[robballantyne]

Reasoning models (DeepSeek-R1, Qwen3, etc.) often consume all tokens on internal chain-of-thought, leaving content empty. This caused false test failures even though the model was serving correctly.

• Handle finish_reason=length as a warning instead of failure
• Check thinking_content field in addition to reasoning_content
• Only fail inference when ALL prompts produce no output at all

[Copilot]

Pull request overview

Updates the vLLM serving pipeline inference test to avoid false failures for “reasoning” models that may return empty content while consuming tokens for internal reasoning.

Changes:

• Treat finish_reason=length with empty output as a warning (not a hard failure).
• Consider thinking_content as an alternative to reasoning_content when determining whether the model produced output.
• Only fail the inference check when no prompts produce any output signal (no passes and no warnings).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The log label ("reasoning_content") can be inaccurate now that thinking_content is also treated as reasoning. When thinking_content is present (and reasoning_content is not), is_reasoning becomes true but the output still claims reasoning_content, which can mislead debugging. Consider tracking which field was used (reasoning_content vs thinking_content) and printing the correct label (or a generic reasoning/thinking_content).