For the last 6 player-side windows H-021 never screened (snap_pct_season, target_share_season, carry_share_season [lags.sql season_td mean of weekly shares], td_oe_l8 [finishing.py ratio of sums, 8 rows], hurry_delta_l8 [tempo.py ratio-of-sums difference, 8 rows, gated 20/60], own_scramble_rate_l4 [tempo.py team scramble/dropback, 4 team-rows]), does reweighting the SAME rows geometrically (half-life 2 over 8 lags, D-209's winner and the family of the 09-25T1221Z screen; and half-life 4 over 16 lags) (a) estimate the realised next-row quantity better model-free and (b) carry anything the shipped line does not already contain?

Owner lane: premise

Question

For the last 6 player-side windows H-021 never screened (snap_pct_season, target_share_season, carry_share_season [lags.sql season_td mean of weekly shares], td_oe_l8 [finishing.py ratio of sums, 8 rows], hurry_delta_l8 [tempo.py ratio-of-sums difference, 8 rows, gated 20/60], own_scramble_rate_l4 [tempo.py team scramble/dropback, 4 team-rows]), does reweighting the SAME rows geometrically (half-life 2 over 8 lags, D-209's winner and the family of the 09-25T1221Z screen; and half-life 4 over 16 lags) (a) estimate the realised next-row quantity better model-free and (b) carry anything the shipped line does not already contain?

Prediction

Registered in protocol.json (sha b8cfc792) at 13:17Z before any computation: (a) season shares EWMA beats flat model-free for all 3 at both hl, gains 5-25% of flat error; td_oe hl2 NOT better; hurry and scramble within +-5%; (b) every |E2 corr| < 0.03, none outside Bonf-36; (c) no E3 below its Bonf-12 permutation quantile; overall MEASURED NOTHING for all 6. REFUTED if a candidate breaches E2 AND E3 on the same candidate.

Finding and verdict

MEASURED NOTHING end to end on H-021's last 6 player windows (24 of 35 now nil at this resolution); ONE STABLE RESIDUAL LEAD: season-share recency is under-read on the targets/carries endpoints (rho 0.03-0.05, 3/3 seasons for target share), worth <= 0.0013 targets/row linearly, below any refit's +-0.01 resolution

Reasoning

Completes the player side of H-021 with the no-refit screen on the exact-replica baseline, and finds the first window in 24 whose recomposition is NOT already absorbed by the line on some endpoint, while bounding how little it can be worth.

Next action

H-021 player side done: 24 of 35 windows nil end to end at +-0.01. Remaining 11 are team windows (points_l3, points_trend, points_allowed_trend, plays_l3, plays_trend, pass_rate_l3, pass_rate_trend, pass_rate_season, opp_plays_l3, opp_pass_rate_l3, opp_points_allowed_l3) and need MATCHED_BOTH_BALL because they feed the opponent copy too; H-079 already covered the teamtd points block. The season-share lead is not worth a refit arm at its linear bound; a referee could test whether the tree uses x nonlinearly (quantile or per-role split), which a Pearson screen cannot exclude.