[kvlonge]

Why

The fresh-process cache benchmark limits allowed 1.5–1.7x the times CI now records. A regression could have undone most of the recent warm-compile gains and still passed.

Changes

These limits are about 1.25x the slowest time CI recorded for each phase on #779:

Models	Cold	Warm	Edit
3000	18s → 14s	10s → 8s	11s → 7s
5000	28s → 23s	15s → 12s	16s → 11s
10000	50s → 42.5s	28s → 24s	30s → 20.5s

The CI maxima behind them are:

• 3000: cold 11.11s, warm 6.03s, edit 5.35s.
• 5000: cold 18.09s, warm 9.58s, edit 8.48s.
• 10000: cold 33.89s, warm 18.85s, edit 16.18s.

The 0.67 warm-to-cold ratio guard is unchanged; CI measured 0.54–0.55. The dense and cold-compile limits were already within 1.04–1.29x of the CI times, so they are unchanged.

Verification

• ruff format --check and ruff check pass on the changed file.
• This PR's cache-compile CI jobs exercise the new limits.