Before the first board after 2026 week 3 (latest_rows D-362 serves 2026 rows once 2026 has three weeks), which tables joined by share.feature_sql have no 2026 rows, and which served features lose their source on every row at once?

Owner lane: construction

Question

Before the first board after 2026 week 3 (latest_rows D-362 serves 2026 rows once 2026 has three weeks), which tables joined by share.feature_sql have no 2026 rows, and which served features lose their source on every row at once?

Prediction

Registered on H-239 at 2026-09-23T0325Z before any run: >= 5 joined tables with zero 2026 rows and no live-path writer; >= 6 features flagged (2026 minus 2025 NULL share > 50pp on rows_for weeks 1-2), all traceable to those tables, 0 traceable to a live-path table. Control 2025 vs 2024 flags <= 1. Positive control: an empty TEMP rel_coach flags the coach pair. REFUTED if <= 1 feature flags.

Finding and verdict

CONFIRMED IN MECHANISM, PREDICTION FAILED ON COUNTS. Three joined tables (rel_coach, regime_gate, kalman_share) have zero 2026 rows and no writer on any live refresh path. At the first board after week 3 five served features lose their source: coach_pos_tgt / coach_pos_car become a literal 0.0, which is BELOW EVERY training value for WR, TE and RB (so H-067's 'served NULL, takes the missing branch' is wrong in the mechanism: it is an out-of-support value, not a NULL), and the three _l3_est shares fall back to raw l3 on the ~96% of rows that carried the Kalman estimate. Separately, drive_volume has 2 of 32 teams for 2026 week 2 because only the stopped `pw refresh` writes it.

Reasoning

Run 1 (NULL share) FAILED its positive control; run 2 (join coverage) passed both controls. The count prediction was built on the assumption that the rel_*/pbp_team one-offs and the build-only modules were all in the join; rapm_theta, uplift_cate, ngs, pfr and pbp are written in-season.

Next action

Time-bound: re-run h239_join_coverage.py and h074_served_vs_trained.py on the first board after 2026 week 3 settles (not before 2026-09-29). The price (served 0 vs a 2025-carried coach value, and ks vs raw l3) belongs to the forecast lane's walk-forward; the writer gap is a build item: regime.build/build_kalman and a rel_coach writer on the live refresh path, and drives.build_volume on the reingest path.