Calibration: try the medium width and full sizes together, so the few-mode cases keep a narrower team's gain #592

What

From SEP 46's verification report (#582). On 16 cores of a Google Axion a process-wide team of 8 ran the few-mode cases faster (h2_raman_averaged 0.90, argon_blueshift_envelope 0.80) and the production-sized radial problem 1.17–1.18 times slower, so the per-problem rule refused it. The grain the calibration then admitted (medium width 8 with Raman and copy thresholds) keeps only part of that gain: h2_raman_averaged 1.00, argon_blueshift_envelope 0.91.

Those cases' operations sit above the full sizes, and the grain stage admits a move only when it wins alone, so a medium width with a higher full size is never tried as one candidate.

Leads

• Search the medium width and the full sizes together in the grain stage, as one candidate per pair, within the stage's four.
• Fill the team with more work per region instead: a source's forward transforms as one batch (Native CPU: a stage's forward transforms as one batched region (#560 item 4) #584).

Done when

The Axion's few-mode cases keep their team-of-8 gain without slowing a production-sized problem beyond the noise plus 5 %, or the measurement says why not.

🤖 Generated with Claude Code