feat(v1): expose RLM execution timeout

[hallerite]

Summary

• expose exec_timeout on the v1 RLM harness and pass it through the ACP runtime policy
• pin nano-RLM commit bc3ab4d033d1978ad942f032700577eda8950163
• define the setting as active IPython execution time: direct sub-agent waits and gathers made only of sub-agent calls are excluded; skills and other work remain charged

The timeout mechanism lives in nano-RLM’s trusted supervisor broker. Verifiers only owns the public harness setting, ACP transport, and exact compatible pin.

Companion: PrimeIntellect-ai/nano-rlm#171

Validation

• uv run pytest tests/ — 84 passed; 76 credentialed integration cases skipped locally
• uv run pre-commit run --all-files
• pre-push ty CI-parity check
• signed commits verified as SSH signatures
• old nano-RLM regression: exec_timeout=1 interrupted a plain two-child gather after one second, despite both child inference calls completing
• exact pinned commit on Prime: 3/3 DeepSeek V4 Flash rollouts completed the same plain gather; each waited 4.31–5.50 seconds, both children overlapped, no timeout was emitted, and all semantic call/return/continuation edges were present