[Harness Watch] Run scheduled comparisons through the existing evaluation workflow

Parent epic: #4627

Goal

Add one thin controller in OpenHands/evaluation. It schedules and coordinates existing eval-job.yml runs; it does not implement another evaluation runner.

This issue absorbs the small versioned-configuration requirement from #4636.

P0

• Commit one configuration containing the model/route, benchmark and fixed instance panel, harnesses, refs, budgets, worker limits, and practical alarm thresholds.
• Resolve refs and model configuration once, then dispatch four independent eval-job.yml runs with identical inputs except agent type.
• Support weekly and manual triggers, an overlap guard, a total concurrency budget, and rotated dispatch order.
• Record child run IDs and result URLs, then invoke #4642 after all cells finish.
• Treat workflow, infrastructure, missing-result, and telemetry failures separately from harness outcomes.
• On a practical threshold breach, rerun only OpenHands and the affected comparator once on the relevant complete blocks with fresh workspaces and keys.
• Update one tracking issue only when the confirmation reproduces the direction with complete telemetry.

Initial alarms may remain simple: a 10-point resolved deficit, or greater than 1.5× spend, fresh input, or proxy requests without a resolved advantage. Thresholds are operational, not significance claims.

P1

Add more sophisticated monitoring, adaptive confirmation, and release-blocking policy only after P0 has stable operating data.

Start by reading the existing OpenHands/evaluation code and the eval-job.yml workflow, then review the related parent and follow-up issues #4627, #4636, and #4642. Trace how workflow inputs, run IDs, result URLs, telemetry, and concurrency are currently handled. Done means a versioned controller can schedule the four runs, guard overlaps, handle retries and failure categories, and update tracking only after complete confirmation telemetry.