Core agent structured outputs via Responses API JSON Schema (stacked on #131)

[rkirkendall]

Modernizes agent action generation to use strict structured outputs.

Summary

• Uses Responses API JSON Schema in responses mode; falls back to Pydantic parse in legacy mode.
• Surfaces explicit refusal; prefers typed parsed payload over post-parse JSON.

Why

• Consistent, typed outputs across models and modes.
• Unlocks strict schema validation and better failure handling.
• Backward compatible with legacy path.

Scope

• Only core agent path + client plumbing. No notebooks or examples included.

Changes

• tinytroupe/agent/action_generator.py: Send JSON Schema with strict=true (derived from CognitiveActionModel / CognitiveActionModelWithReasoning) when API_MODE=responses; prefer parsed; handle refusal.
• tinytroupe/openai_utils.py: API_MODE=responses|legacy; map to responses.create/parse (max_output_tokens, reasoning.effort); unified extractor returning role/content/parsed/refusal.
• tests/unit/test_action_generator_structured_outputs.py: Tests for parsed-preference and refusal handling.

Backward compatibility

• Legacy path unchanged; schema enforced only when requested.

Testing

• Unit tests pass locally.

Notes

• Stacked on/blocked by Enhance OpenAIClient to support Responses API mode #131: Enhance OpenAIClient to support Responses API mode #131

Follow-ups

• PR3: Enrichment structured outputs.
• PR4: Extraction/Normalizer structured outputs.