Preserve Pydantic AI context across validator retries

[hertznsk]

Preserve Pydantic AI context across validator retries

Summary

• preserve the completed Pydantic AI message history when a semantic validator rejects Claude or OpenAI output
• send validation feedback as the next user turn instead of rebuilding the original prompt from scratch
• keep existing stateless validator retry behavior for providers that do not expose continuation state

Why

Validator retries previously called the primary agent again with a reconstructed prompt. For the Pydantic AI providers, that discarded the model response and any tool-call/tool-result exchanges from the first attempt. The correction now uses Pydantic AI message history so the model can revise its existing work with the validator feedback in context.

Testing

• make check
• uv run pytest tests/test_engine/test_validator_integration.py tests/test_providers/test_pydantic_ai_interrupt.py tests/test_providers/test_pydantic_ai_provider.py tests/test_providers/test_openai.py -q
• make test (8610 passed, 32 skipped; 4 unrelated environment/performance failures: two POSIX unreadable-path assertions and the existing Claude performance thresholds)

Notes

• continuation state is provider-opaque and in-memory only; workflow checkpoints and serialized outputs are unchanged
• Pydantic AI receives the prior all_messages() as message_history and the validator feedback as a new user prompt
• external providers are not passed the new keyword unless continuation state exists
• AGENTS.md validator sections updated to describe the continuation behavior; a fail-open regression test covers a failing continuation re-run