Review-depth pilot close: reconciliation is unverifiable by construction - Elazar decides keep classifier / drop pilot

The review-depth pilot (MARS, Elazar-authorized 2026-08-12, window opened at 5f1da49) closed around 2026-08-26 and its close reconciliation never ran. It cannot run now. WHY IT IS UNVERIFIABLE BY CONSTRUCTION, not merely late: the pilot's primary output was the SILENT-SKIP COUNT - pushes that went out with no review brief at all. Measuring it requires matching every SHA in git log 5f1da49..HEAD against the reviewer's verdict record. That verdict record lived with audit-mars-ca, which left the Mars roster on 2026-09-11 without producing it. git log enumerates SHAs PUSHED; it cannot enumerate briefs SENT. Those two sets differ by exactly the quantity being measured, so reconstructing the reconciliation from git alone does not approximate the answer - it returns zero skips by construction, which is the most favorable possible result and the least supported. DO NOT let anyone produce that number and present it as the pilot's measurement. Record the pilot as inconclusive, not as passed. WHAT SURVIVES REGARDLESS, and should not be dropped with the pilot: scripts/review-depth.sh is in the repo, works, and is the only thing on Mars that decides review depth from the actual diff rather than from a human calling a tier by eye. It stays in force on every Mars push independent of this decision. WHAT ALSO CHANGED UNDER IT: the pilot assumed a cross-VENDOR reviewer (Codex vs Claude). Mars no longer has one - review is now coder02-mars-cc (Sonnet 5) against coder-mars-cc (Opus 5), cross-model, same vendor. So even a clean reconciliation would not have measured the arrangement Mars runs today. ELAZAR DECIDES: keep the depth classifier and formally drop the pilot as inconclusive, or re-open a measurement window under the current two-agent reciprocal review. Audit's own PASS:5f1da49 was explicitly conditional on push-path enforcement (wiring the depth artifact into gitpush.sh so a missing brief becomes impossible rather than merely counted) before any rollout - that condition was never met either, and it is the thing that would make a second window actually measurable.