Calibration: why measured and patient FFT plans chosen alone rarely win in the step #595

What

SEP 46 §12 item 10, left open at Final (#582). Under the revised rules the FFT planning stage is compared with the plan policy (rows and planes measured, threaded rows estimated), and nothing is published unless an effort wins. In the verification only one calibration published planning: an AMD EPYC 9B45 under Linux chose rows patient, which was not A/B'd. Under the first rules, stored wisdom from a plan chosen alone ran h2_raman_averaged 1.17 times slower in the step on an EPYC 9B45.

The phase 3 evidence records, for the plans that lost, how long each took alone and in the step. It has not been analysed across machines.

Questions

• Does a measured or patient plan chosen alone lose in the step because of timing noise, the cache state, or interplay with the stage's other threads?
• Should the thorough level drop patient plans, and so save planning time, if they never win?

Done when

The losers' alone-against-in-step timings, from the stores of several machines, are summarised in docs/dev/cpu-performance.md, and the ladder is kept or trimmed accordingly.

🤖 Generated with Claude Code