Harbor upstream: agree live-site oracle evidence and parity acceptance protocol

Decision needed for upstream acceptance

Parent #331 proposes a documented no-op oracle because ClawBench uses live websites. Harbor's current adapter contribution guide requires solvability/oracle validation and parity with the same agents, models, and settings; it recommends discussing difficult tasks case by case. Our solve_script() only prints a message and exits successfully; it provides no intercepted request or successful judged outcome. A shell exit of 0 is not oracle evidence.

Source: https://github.com/harbor-framework/harbor/blob/fd0049126a060f930372f9b0b0aa330caf1f19e5/docs/content/docs/datasets/adapters.mdx (oracle validation, parity planning, structured reporting).

Acceptance criteria

• Obtain and record an upstream-agreed approach for live-site solvability evidence (real reference trajectories or an explicit exception); do not label the no-op script an oracle pass.
• Decide whether upstream adapter submission uses the conventional root layout or accepts our supported steps/run layout. Harbor's loader accepts steps; contribution acceptance is a separate question.
• Specify a representative task subset, exact corpus/agent/model/judge revisions, browser/tool configuration, timeouts, retries, run count, and rate limits before experiments.
• Report Stage 1, lenient/strict Stage 2, judge-inconclusive and infrastructure counts separately, with denominators and mean ± sample SEM across independent runs.
• Produce actual adapter_metadata.json, parity_experiment.json, reference YAML, and reproducible evidence before claiming official integration.
• Keep upstream docs-only PR harbor-framework/harbor#2495 distinct from an adapter or registered dataset submission.

This issue narrows #331 §4 rather than reopening the overall integration request. Honor Perry2004's decision to defer image publication and CI; neither is made a prerequisite here. No parity results have been fabricated or experiments launched by this issue.

@Perry2004 please review the live-site oracle approach and parity protocol and coordinate the upstream acceptance decision.

Start with docs/content/docs/datasets/adapters.mdx and the current solve_script() behavior described in the issue. Resolve the upstream live-site oracle and layout decisions, then define the requested corpus, agent, model, judge, browser, and run parameters. Done means producing adapter_metadata.json, parity_experiment.json, reference YAML, and reproducible evidence without claiming integration prematurely.