Persist Stirrup assistant reasoning in trajectories

[ShuxinLin]

Description

Persist the optional AssistantMessage.reasoning payload for Stirrup turns so tool-only turns can retain provider-supplied reasoning separately from visible assistant text.

Fix Details

• Add a Stirrup-specific turn record with an optional reasoning field.
• Serialize reasoning as {signature, content} without changing text semantics.
• Leave the shared turn schema and other agent runners unchanged.
• Add coverage for populated and missing reasoning payloads.

Impact on Benchmarking

• No change to baselines: This fix only improves stability/performance.
• Baseline change: Persisted Stirrup trajectories now include reasoning when the provider supplies it. Agent answers, tool calls, token counts, and operational metrics are unchanged, but LLM-judge inputs may change because the evaluator serializes the complete trajectory. No full before-versus-after benchmark was run.

Related Issues

• None.

Verification Steps

1. Run uv run pytest src/agent/stirrup_agent/tests/ -q.
2. Confirm all 33 Stirrup tests pass.
3. Construct a Stirrup AssistantMessage with reasoning and confirm trajectory serialization contains reasoning.signature and reasoning.content.

Checklist

• I have added tests that prove my fix is effective.
• My changes are Ruff-formatted. The repository has pre-existing lint findings under the latest standalone Ruff release in this test file.
• I have signed off my commits (DCO).