Is the 09-24 replica failure fit non-repeatability, second-fit-in-process state, TEMP shadowing per se, value recomputation, or physical row order reaching a fit, and which fitted component carries it?

Owner lane: Not recorded

Question

Is the 09-24 replica failure fit non-repeatability, second-fit-in-process state, TEMP shadowing per se, value recomputation, or physical row order reaching a fit, and which fitted component carries it?

Prediction

Stage 1 (registered before any fit): S1a, S1b, S2 bit-identical; COPY bit-identical to S1a; ORDER moves forecasts at the scale REPLICA did, so the cause is physical order of player_week reaching a training-side read. Stage 2 (registered after stage 1 and the swap): with efficiency.py:182 ANY_VALUE(position) replaced in memory by arg_max(position, season*100+week), S1a_P, ORDER_P and REPLICA_P are bit-identical on all 6,824 forecasts.

Finding and verdict

CAUSE IDENTIFIED for the H-021 replica failure: physical row order of player_week reaching efficiency.fit through ANY_VALUE(position) at efficiency.py:182. Not fit non-repeatability, not TEMP shadowing, not the window recompute. Both preregistered predictions held. MEASURED NOTHING about EWMA accuracy. H-021 stays OPEN, blocked on WORK (a one-line order-free position rule), now checked by invariant H8