[Harness Watch] Generate deterministic paired comparison reports

Parent epic: #4627

Goal

Compare the four existing evaluation result archives deterministically. An LLM must not calculate metrics or decide whether a threshold was crossed.

P0

• Validate matching instances, model/route, refs, budgets, harness identity/version, and complete telemetry.
• Join cells by instance and report missing or invalid cells explicitly.
• Compare OpenHands separately with OpenCode, Pi, and Hermes.
• Report resolution difference and wins/losses/ties, with an exact paired binary test.
• Report paired spend, proxy requests, input/cache/output tokens, and wall time using mean, median, P90, and one reproducible paired-bootstrap interval.
• Report attempts, timeouts, infrastructure failures, telemetry coverage, and complete-pair count.
• Emit versioned JSON and concise Markdown with the largest outcome/resource differences and links to trajectories and telemetry.
• Test ties, missing cells, zero costs or successes, heavy tails, and bootstrap reproducibility.

Operational alarms use practical thresholds plus a confirmation run. A non-significant result is not proof of parity.

P1

Add multi-model and repeated-run models, cost-effectiveness views, dashboard/index publication, partial-quality signals, and historical analysis.

Start with parent epic #4627 and inspect the four existing evaluation result archives and their harness metadata. Define the deterministic validation, pairing, comparison, bootstrap, and JSON/Markdown reporting requirements in the P0 list. Done means reproducible reports explicitly cover invalid or missing cells, ties, resource metrics, failures, telemetry, and the named edge cases.