daily-fix: measured n_train + regime-change pilot basis

Workflow gap

Two same-day sizing failures trace to plans quoting inferred numbers where a measured artifact-side read was available:

1. Inferred n_train invalidated a plan post-smoke. #2061's §7 n_train table was "inferred, not measured; a registered cell is n<d" — sharpened to "§7 used the WRONG DENOMINATOR; likely 5 of 7 combos degenerate" (corpus n_built vs realized turnstore rows). Cost: an epm:strategy-pivot + plans v11→v13 (~2.6 h re-planning) — caught before the ~70 GPU-h production spend, but only post-smoke. The #1887 n<d refusal makes an inferred count plan-invalidating by construction.
2. A decoding-regime change invalidated the parent's wall-time as sizing basis. #1491's greedy rerun was quoted at 40–50 GPU-h off the parent Repro footer's 5.4 h wall (which "swept in setup and two failed launches"); greedy decoding's repetition loops (cap-hit 18.47% vs parent 6.66% at 0.5B) pushed realized cost to ~78–82 GPU-h (~1.7–2.0× the figure Thomas approved). Disclosed proactively + recorded as epm:compute-deviation v1 with measured causes.

verified-at-filing: incident 1 is the probed pivot-marker chain on #2061 (session 5c878aa4 rows 1061/1140/1147, fact-check reproduced all 55 store counts); incident 2 is #1491's own compute-deviation marker (2026-08-06T00:58:41Z, probed). grep -n 'n_train' .claude/agents/planner.md | head -3 and grep -n 'MEASURED' .claude/rules/plan-compute-sizing.md | head -5 run at compose time — the measured-pilot rule exists for wall-time; neither surface names these two specific bases.

Proposed change

• .claude/agents/planner.md §7: any registered per-fold/per-cell n_train must cite a MEASURED artifact-side row count (sidecar/manifest read via a named command), never a corpus-name inference.
• .claude/rules/plan-compute-sizing.md: name the regime-change trap — a decoding/ sampling regime change (sampling→greedy, temperature change, cap change) invalidates the parent run's wall-time as a sizing basis (completion-length distribution shifts); require a 1-rung measured pilot before quoting a user-facing cost for a regime-changed rerun.

Provenance

• fingerprint: 56904e1c9622

• workflow_fix_target: .claude/agents/planner.md, .claude/rules/plan-compute-sizing.md

• origin: /daily 2026-08-05 problem sweep — miner 5 P8, miner 1 P7.