fix(parsec): wait for simulator quiescence before seeding, and raise the agent ceiling

[bcarmeli]

Two failure modes found while running the h4 held-out suite. Both produce results that look valid, which is why they went unnoticed for several arms.

1. Seeding races an in-flight tool call

PUT /api/v1/simulation/database is refused with HTTP 409 while a service still has a call in flight. A trial ends when the agent stops — which does not wait for the simulators it stopped talking to — so an abandoned call outlives the trial that made it, usually on a service the next task never touches.

Arms that seeded only each task's declared services never hit this: the service the current agent used is quiet by construction, because the agent waited for its answer. Seeding every service exposes it, and it cost 69 errored trials in one sweep.

_await_quiescent_sims() polls every service in MCP_PORTS to queue_depth == 0 first. It lives in the adapter, not in install_seeds.py, because that script runs under the timeout=120 subprocess cap — a drain budget inside it races its own caller, which is how an earlier attempt turned a recoverable wait into a dead trial. An unreadable service counts as busy, never idle: seeding into a service whose state is unknown is the thing being prevented.

2. Trials outlive the task's agent timeout

Degraded gateway latency pushed 10-18 call tasks past the task's own [agent] timeout_sec (600s in h4). Harbor cancels the agent, the trial still writes a reward.json, the verifier's completion gate multiplies it by zero, and it lands in the results as a 0.0 indistinguishable from a wrong answer — with errored_trials still reading 0, because the trial completed as far as the harness can tell.

15% of one arm's trials died that way, against 0.7% in its neighbours.

Both knobs are needed: the multiplier raises the ceiling Harbor enforces, and TRIAL_TIMEOUT_SEC must stay above it or this subprocess cap preempts the ceiling it exists to contain. A test asserts that ordering, because getting it backwards is the same nesting mistake as the drain budget in §1.

These are tuning values, not permanent: revert to 20 * 60 / 1.0 once the gateway slowdown is understood. A raised timeout makes a failing trial cost 2.5x more wall clock, and it did not rescue the one task that thrashes on a deliberately unmodelled tool.

Notes for review

• Ports come from parsec_paths.MCP_PORTS rather than a second table; service names are derived from its keys.
• 5 new tests cover: all-idle, naming the busy service, unreadable-is-not-idle, draining mid-wait, and the timeout ordering invariant.
• test_adapter.py has two pre-existing failures on origin/main (TASK_ID assertions left behind when that support was removed). Untouched here; they fail identically before and after. Happy to fix them in a separate PR if wanted.

[skillberry-bot]

🏷️ Automatic Labeling

I've analyzed this pull request and added the following labels:

• bug - honesty - bug - honesty - priority-p1

These labels were selected based on the PR title, description, and changed files. If you believe any labels are incorrect or missing, feel free to adjust them manually.