List failure signatures visible to the caller (org + system tier).

Seat scorecard for every dispatch in a date range, grouped by project × dispatch kind × card revision at dispatch time × model × profile × harness × effort. Per bucket: seat count, status mix, the completed seats that delivered over budget (a stored runner budget breach - apart from failures), the review seats' typed verdict mix (with seats missing a verdict), the GitHub review mix, dollars and tokens per seat (every cost row), p50/p90 duration, eval pass rate, first-pass merge and rework rounds. Review accuracy, profile, effort, time to merge and cache share are not reported until their sources land; `unavailableDimensions` names the issue for each, and for seats with no stored launch revision or typed verdict. The window is half-open: dispatchedAfter is inclusive (default 30 days before dispatchedBefore or now) and dispatchedBefore is exclusive. The scan reads the newest dispatches first and sets `truncated` when the limit cut it. A projectId the caller cannot access answers 404, exactly like a missing project. Missing links land in an "unattributed" bucket; open PRs count as pending. Each source reports its latest row and fresh/stale status; every metric derived from a stale source is null, with the reason in `unavailable`.