Structured output falls back to free text ~1x per run on DeepSeek thinking models

Problem

On DeepSeek thinking models, the free-text fallback in invoke_structured_or_freetext is not a rare edge case — it fires once or twice in every run.

The comment at structured.py:77 already anticipates the shape ("a thinking model can answer in plain text instead of calling the tool"). What this adds is the frequency, which seems high enough to justify a retry policy rather than just a logged warning.

Measurement

v0.5.0, deep_think_llm=deepseek-v4-pro, quick_think_llm=deepseek-v4-flash, four analysts, full research and risk debates, ~21 LLM calls per run. Over 10 consecutive runs:

structured-output invocation failed (structured output returned no parsed result);
retrying once as free text

  Trader             6
  Research Manager   4
  Portfolio Manager  1
  ------------------------
  total             11 across 10 runs

tool_choice was correctly suppressed throughout (both model IDs have capability rows), and "does not support this tool_choice" occurred 0 times. So this is distinct from #1388 / #1389 and is not fixed by closing that naming gap.

Impact

1. Schema enforcement is effectively not guaranteed — on average one agent per run produces unvalidated output.
2. It is only a logger.warning, invisible in the finished run summary, so it goes unnoticed unless you are watching stderr.
3. Ratings survive only because they are recovered from the **Rating**: / **Recommendation**: / **Action**: labels rather than from the schema. Numeric schema fields (entry price, stop loss) have no such safety net.

Possible directions

• Retry the structured call once before falling back. The miss is a sampling artifact, so a retry will often succeed, at far lower cost than losing the schema.
• Use json_mode for known thinking models (supports_json_mode already exists in the table).
• At minimum, surface a fallback count in the run summary so users know how many agents skipped the schema.

Happy to send a PR for whichever direction you prefer.