fix: surface agent timeout as state['error'] in CliAgentEnv

[rasdani]

Problem

CliAgentEnv.wait_for_completion catches asyncio.TimeoutError but only sets state["agent_timed_out"] = True, leaving state["error"] unset. When the agent times out without producing any trajectory, the rollout comes back with trajectory=[], error=None, and the orchestrator scheduler reschedules it as a generic "Empty trajectory" with no diagnostic — indistinguishable from a genuinely silent empty rollout.

Fix

In the asyncio.TimeoutError branch, also set state["error"] = AgentError(f"Agent timed out after {self.timeout_seconds}s"), mirroring the style of the except Exception branch directly below. agent_timed_out=True is preserved for existing downstream checks elsewhere in the file.

Context

Third in a series closing silent-failure holes in CliAgentEnv:

• Fix CliAgentEnv stuck on dead tunnel #1127 — non-zero exit + empty trajectory
• Fix AgentError double-wrapping in poll_job_completion #1130 — AgentError double-wrapping
• this PR — timeout path

A companion fix in prime-rl reorders the scheduler's if empty / elif error so that the error branch is checked first (otherwise this surfaced error would still be hidden behind the empty-trajectory branch). See companion PR in prime-rl.

Tests

No regression test was added. The existing tests/test_cli_agent_env.py covers timeout_reached (the stop-condition) but does not exercise wait_for_completion's asyncio.TimeoutError path. Adding such a test would require mocking asyncio.wait_for and the background-job polling loop, which felt out of scope for a 3-line fix — flagged here for a follow-up if desired.

Checks

• uv run ruff check verifiers/envs/experimental/cli_agent_env.py — passed
• uv run ruff format --check verifiers/envs/experimental/cli_agent_env.py — already formatted

---

Note

Low Risk
Small, localized change to timeout handling that only affects how errors are surfaced in rollout state; minimal behavioral impact beyond improved diagnostics and potential downstream branching on state["error"].

Overview
CliAgentEnv.wait_for_completion now records timeouts as a real failure by setting state["error"] to an AgentError when asyncio.wait_for hits TimeoutError, in addition to the existing state["agent_timed_out"] = True flag.

This ensures timeout rollouts surface a diagnostic error instead of looking like a silent empty trajectory to downstream schedulers/consumers.

^{Reviewed by Cursor Bugbot for commit e88fd94. Bugbot is set up for automated code reviews on this repo. Configure here.}