fix(agent): clear stale failure state when resuming runner iterations - #5995

Summary

Fix successful recovery being reported as a failed run after a late follow-up message, which can suppress the final WebSocket reply.

Root Cause

The model-error and empty-response paths assign stop_reason and error before checking for late follow-up messages.

When a follow-up resumes execution, these values survive into subsequent iterations. A successful response replaces the content but retains the previous failure state.

For model errors, this causes TurnDelivery to skip the final WebSocket reply and publish a failed completion. The runner also invokes on_error despite having recovered successfully.

Changes

• Reset stop_reason and error at the start of each runner iteration.
• Add parameterized regression coverage for recovery after model errors and exhausted empty-response retries.
• Verify final reply delivery, successful completion, and absence of a stale run-level error callback.

Result

Recovered runs return stop_reason="completed" and error=None. The final WebSocket reply is queued, and turn completion reports success.

Failures that terminate execution continue to retain their failure state.

Test

• Before the fix, both new regression cases failed:
  • Model-error recovery produced no final outbound reply.
  • Empty-response recovery retained empty_final_response.
• After the fix: 239 tests passed across runner, injection, hooks, delivery, loop recovery/integration, SDK streaming, and facade tests.
• Ruff passed for both changed files.
• BasedPyright passed for nanobot/agent/runner.py.
• git diff --check passed.