On the row the board serves, CP-H49 refreshes the target/carry share slots through the man's final game while targets_l3, carries_l3 and snap_pct_l3_est still stop one game earlier. Does that inconsistency cost close start/sit decisions, and is the cost tier-dependent?

Owner lane: decision

Question

On the row the board serves, CP-H49 refreshes the target/carry share slots through the man's final game while targets_l3, carries_l3 and snap_pct_l3_est still stop one game earlier. Does that inconsistency cost close start/sit decisions, and is the cost tier-dependent?

Prediction

Frozen in the runner docstring and committed (466c3006) before the run. P1 primary: SVN - S product accuracy on SEL_C +0.2 to +0.8pp, regret -0.02 to -0.10; Bonf-4 may span 0. P2: S0 - S within +/-0.5pp spanning 0; S0 MAE below S (H-095's +0.031 reproduces in sign). P3: MAE SVN below S by 0.02-0.08. P4: SVN - S gain larger on XX than LL pairs. REFUTED if SVN - S accuracy within +/-0.3pp with a 95% interval containing 0 AND the MAE contrast interval contains 0.

Finding and verdict

TESTED -- H-158's cost claim REFUTED by its registered rule. The premise holds: on the live board's row, share describes the final three games and volume/snap the three before (fixture: shipped volume matches the Wednesday window on 17.6% of rows, snap on 5.6%). Aligning them buys nothing on the decision: SVN - S -0.10pp [Bonf-4 -0.40, +0.20], regret +0.014 [-0.019, +0.050] on 89,991 consensus-selected close pairs. MAE +0.005 [-0.000, +0.011], snap alone +0.006 worse. Reverting the share refresh (S0) is also nil (+0.05pp). No tier shows a gain. The served row's -1.27pp decision cost is untouched by the repair and sits three times heavier on pairs involving a non-leader (secondary, unpredicted).

Reasoning

H-089 and H-222 name the served-row staleness as the start/sit cost, and CP-H49 fixed six share columns of it. H-158 asked whether the half-fix creates its own harm by mis-aligning a man's role. The decision lane owns the consumer of that row, and nobody had separated inconsistency from staleness.

Next action

(1) H-222's repair must refresh the WHOLE history block; a column subset does not recover the decision cost -- SVN - T is as bad as S - T. The block-wise decomposition (which of pb/fm/ng/pf/rp/uc/_season/fp carries S - T) is the next measurement on this harness: one leave-one-block-fresh arm per block, the same pair sets. (2) The tier gradient of S - T (LL -0.60pp vs XX -2.06pp) is a secondary finding; confirm it on 2026 served rows once 4+ weeks settle, pre-registering the XX - LL contrast. (3) SN's MAE harm says raw snap l3 is worse than snap_pct_l3_est; CP-H49 already writes RAW shares into the _est slots. Whether that part of CP-H49 costs anything is a separate, cheap arm (write the refreshed value only into _l3_raw).