daily-fix: null-statistic gates need measured calibration

Workflow gap

#1491's approved plan pre-registered Gate-1 as abs(r2_null) < 0.05, but the shuffle-REFIT null it thresholds has expected R² ≈ −1 — "unsatisfiable by construction." All 8 shards of the 0.5B rung FAILed the mid-shard gate and aborted (GPUs at 0% on an 8×H200 pod at ~$44/h); the first fix replaced it with an ASSERTED fixed floor (null_floor=-3.0) which then killed the 1.5B rung too (r2_null=-3.40…-3.80; "My prediction was wrong"). Only the third change (advisory floor) held — implementer measurement showed the null depth is non-monotone in n/h (double-descent peak at n≈h), so ANY constant floor was wrong. The defect survived plan approval + multiple code-review rounds because no review calibrated the gate against a measured null draw.

verified-at-filing: the gate FAIL values, both fix commits (ccc650f42e, 6d5c675a95), and the double-descent measurement are the recovery miner's probed transcript reads (session 8d7f8b25 rows 2383–2638, incl. the epm:failure v3 diagnosis "unsatisfiable by construction").

Proposed change

Add to .claude/rules/selection-symmetric-nulls.md (and cross-reference from planner.md §7): any pre-registered numeric gate that thresholds a NULL statistic requires a MEASURED calibration basis — a 1-cell pilot of the null at production n/d shape — before production; asserted constants on an unmeasured null distribution are the #1491 Gate-1 class. Prefer advisory logging over hard-abort for null-side conditions (the null is diagnostic, not a kill criterion, unless the plan argues otherwise).

Provenance

• fingerprint: 46643e3b4037

• workflow_fix_target: .claude/rules/selection-symmetric-nulls.md

• origin: /daily 2026-08-04 recovery sweep — miner 2 P2 (probed rows).