FOMO Radar — calibrated probability output with a labeler and backtester

From ET-25011 (expire76): replace the 0-100 conviction with a checkable probability ('x% chance of 3x within 72h') via isotonic regression on historical outcomes. Prerequisite: a labeler + backtester over past signals (MemeTrans-style public data, rug detection from the first 5 minutes, graduation survival). Nobody has measured the score->outcome link yet; this builds it. --- Acceptance criteria (from expire76's design note "The State Layer", 13 Sep 2026, §02/§11/§14 — ET-25104/25106): 1. Labeling universe includes every launch, including tokens that died within a minute (no survivorship filter). 2. Label = return under a named exit policy (trailing stop / time exit / partial sells), not the price maximum. 3. Labels are liquidity-adjusted: simulate selling the real size against real pool depth at each future timestamp. 4. Walk-forward only, with an embargo longer than the longest label horizon. 5. L1 gate metric is MCC >= 0.30 (class-balance-invariant), never AUCPRC on an 84%-positive set. 6. Persist window_complete into the point-in-time store beside every feature vector; the labeler excludes or flags restart-biased rows. 7. Test by regime window (graduation-rate break Mar-Jun 2026), and account for own price impact in thin pools. 8. A negative-EV result at every horizon is a valid, publishable outcome of this item — it gates 760/761/763.