[kvlonge]

Why

The fresh-process cache guards fail on runner speed. Unchanged code missed the 9.0 s and 33.6 s cold budgets at 9.06 s and 33.78 s, each within 6% of its own cache-disabled control. Across about 22 recent jobs per size, cold wall and CPU spread 1.6x to 1.9x between runner classes, while cold over cache-disabled stayed at 0.94 to 1.15 wall and 0.99 to 1.05 CPU.

No absolute ceiling that fits that spread catches a 20% slower compiler, and a slower shared cold path slows the cache-disabled control too. The same-runner base/head jobs are what catch it, but their cold limits of 1.42x and 1.49x, and 2.93x on model analysis CPU, were sized against main before native analysis. Main now has it on both sides: twelve cold readings since then measured 0.95 to 1.03 wall and 0.96 to 1.01 CPU, and six model analysis readings 0.96 to 1.01.

Changes

• The fresh-process cache guard scales its wall budgets by measure_runner_calibration().factor, as the dense guards do, and prints them. Its ratio and fingerprint checks are unchanged.
• The dense and fresh base/head cold comparisons and the model analysis CPU gate return to the 1.10x limit. The dense edit limits stay as approved.
• The 1.42 and 1.49 multipliers on the absolute fresh and dense cold budgets stay. Their comments now say to drop them when the native cold path is optimised, not when the ratio guards return.
• COMPILE_BENCHMARKS.md says why absolute budgets cannot catch a slower compiler.

Test-only, so no DuckShop case.

Verification

• Same machine, fresh 3,000 models, cold, three alternating runs: identical builds gave 1.018 wall and 0.979 CPU and passed. Injecting CPU work equal to 20% of cold wall into every head compile gave 1.234 wall and 1.073 CPU and failed the 1.10x limit; the previous 1.42x and 1.71x limits would have passed it.
• Both 3,000-model cache guards pass locally (factor 1.000).
• make check-ci passes, except rust-lint: no Rust changed.

[kvlonge]

Folded into #1413 so the fixes share one CI run. The branch is kept.