feat(provider): add provider-specific policy refusal markers

[bukeeastrey]

Summary

When upstream LLM providers (such as Azure OpenAI, OpenAI, Gemini, or Anthropic) filter or reject a prompt or response due to content safety or moderation policies, they return specific error codes or message phrases (e.g. content_filter, responsible_ai_policy, content management policy, flagged by content filter, blocked by safety).

Previously, _is_policy_refusal() only matched a small set of generic phrases, causing these upstream safety blocks to fall through and be misclassified as BAD_REQUEST or UNKNOWN.

This PR adds common provider safety and moderation markers to _is_policy_refusal() to ensure all such errors consistently classify as ProviderFailureKind.POLICY_REFUSAL.

Changes

• src/agentos/provider/failures.py: Added the following markers to _is_policy_refusal():
  • "content_filter" (Azure OpenAI / OpenAI error code & finish reason)
  • "content filter" (e.g. "The prompt was flagged by content filter")
  • "responsible_ai_policy" (Azure / cloud policy violation code)
  • "content management policy" (Azure OpenAI canonical error phrasing)
  • "flagged by content filter"
  • "blocked by safety" (Google Gemini block reason)
• tests/test_provider_failure_classification.py: Added unit tests covering policy refusal error shapes across Azure OpenAI, OpenAI, Gemini, and Anthropic.

Verification

• uv run ruff check src/agentos/provider/failures.py tests/test_provider_failure_classification.py (Passed)
• uv run mypy src/agentos/provider/failures.py --show-error-codes (Passed)
• uv run pytest tests/test_provider_failure_classification.py -q (93/93 tests passed)

closes #629

[andreapn]

Closing as superseded. Reviewed alongside #779 on issue #629; #779 was picked.

Same markers, same place — the difference is the test coverage. #779 adds negative cases pinning that the widened list does not capture invalid_request_error, rate_limit_exceeded or context_length_exceeded, which is the actual risk when you loosen a substring matcher. It also drops "flagged by content filter" as redundant with "content filter".

This PR currently also conflicts with main on tests/test_provider_failure_classification.py.

Thanks for the contribution.