Built-in table: adopt the grain the Axion and Emerald Rapids calibrations agreed on, if it holds on the tuned machines #593

What

Two machines the built-in table was never tuned on chose the same grain in SEP 46's verification (#582), in every conclusive calibration: a Google Axion (C4A) and an Intel Xeon Platinum 8581C (Emerald Rapids, C4). Both kept the table's waiting and team and chose medium width 8, a Raman threshold and a copy threshold.

• Axion: tracking set 0.965, 0.964, 0.968 (gains to 15 % on the few-mode and Raman cases).
• Emerald Rapids: 0.989 (ctc 0.87, gnlse_raman_shock 0.94).

Machines without a profile (SOLPROP_CALIBRATION=off, or a first run's quick row) run on the table. If the grain carries to the machines the table was tuned on, it could become the table's default and every run would get it without calibrating.

Plan

1. A/B the shared grain as the table, against the current table, on the machines it was tuned on: the Intel Xeon Gold 6240, an AMD EPYC 7763, an AMD EPYC 9B45 and the Apple M5 Max.
2. Adopt it per OS and architecture where it holds everywhere, beyond the noise, with no case beyond its noise plus 5 %.

🤖 Generated with Claude Code