fix(runtime): honor provider stop reasons

[yanpgwang]

Summary

• classify every successful Messages API response before advancing or completing the agent loop
• continue pause_turn with the exact provider content, replacing the prior continuation payload when pauses repeat
• recover max_tokens with a bounded internal continuation prompt while keeping CMA public events unchanged
• reject missing, unsupported, and contradictory stop reasons instead of reporting a false end_turn
• terminate honestly when the per-turn loop safety bound is exhausted
• protect existing Temporal histories with a version gate

end_turn, stop_sequence, refusal, and model_context_window_exceeded remain valid completed provider responses. tool_use must contain a client tool call; pause_turn and max_tokens must not.

The continuation behavior follows the Anthropic stop reason guidance, while the public terminal state remains within the Managed Agents session event contract.

Verification

• make verify
• make test-service
• focused regressions for repeated pause_turn, raw server-tool content, max_tokens recovery and bounds, and contradictory provider responses