Make live retrieval diagnostics non-reportable and fail closed

For a current-head retrieval diagnostic, mark the run non-reportable and stop the batch on the first service failure. Persist a started marker before each request and the complete raw tool object immediately after it so completed, failed, and never-attempted cases remain distinguishable without relying on truncated interface output. Grade returned passages separately from request execution. If opening a cited source fails, leave every source-dependent claim unverified.

This protocol applies to live diagnostic batches. It does not create a fixed corpus snapshot, an independent quality review, or evidence about candidates outside the returned set. Avoid automatic retries when a retry would erase the first-failure boundary the diagnostic is meant to preserve.