[kvlonge]

Why

The dense 5,000-model guard scales its budgets by the runner calibration's fastest repetition over a reference. The 0.55 s reference sat inside the EPYC 7763's own readings, so 7763 runs got factor 1.0 or slightly more depending on noise. Their warm (about 9.2 s) and edit (about 9.6 s) medians sit right at the unscaled 9.0 s and 9.5 s budgets, so unchanged code passed or failed at random.

Changes

• The calibration reference drops from 0.55 s to 0.50 s, below every 7763 reading so far.
• Budgets, repetitions, the workload and the same-runner ratio jobs are unchanged. Test-only, so there is no DuckShop case.

CPU	Calibration (s)	Factor at 0.55	Factor at 0.50
EPYC 9V45	0.31-0.39	1.00	1.00
EPYC 9V74	0.46-0.52	1.00	1.00-1.04
Xeon 8573C	0.54	1.00	1.08
EPYC 7763	0.540-0.62	1.00-1.13	1.08-1.24
Xeon 8370C	0.67-0.68	1.22-1.24	1.34-1.36

The reference is below 7763's minimum (0.540 s), so every 7763 run now gets at least 1.08: warm 9.72 s and edit 10.26 s against observed medians of 9.18-9.22 s and 9.57-9.60 s. Two 7763 jobs on unchanged compiler code read 0.540-0.587 s, and one failed at factor 1.000 with a warm median of 9.22 s.

The factor over-reads the compile slowdown between classes, which leaves slack and never undershoots. On 8370C the absolute guard could miss a regression of up to about 45%; the same-runner ratio jobs still catch it.

Verification

• make check-ci passes.
• The calibration helper runs locally and reports the 0.50 s reference.
• The factor ranges above come from recorded calibration readings on hosted runners.