Review-depth pilot: measure audit token burn and shallow-sample quality

Follow-up to MARS-687, not a retroactive verdict project. Measure audit token burn against a stated comparable baseline using actual per-agent token records, not estimates; record window, denominator, extraction method, and limits. Measure shallow-review quality only after real shallow verdicts/samples exist: report sample count, findings, what-was-checked evidence, and false-negative limits. The 2026-10-06 bounded Mars window was 11/11 DEEP, so shallow sample denominator is zero and quality is currently unmeasured. Do not reinterpret historical closed WIs, backfill audit events, or turn 27/581 into a rate. Done when both metrics have defensible actual measurements or an explicit PM/Elazar decision records why one cannot be measured.