[kvlonge]

Why

Three CI runs of the fresh-process cache guards with identical code varied by about 35% in absolute time. For example, 3000-model warm compiles measured 5.30s, 5.60s and 7.12s. Absolute limits near 1.2x therefore fail on slow runners; the 3000-model warm limit on main would have failed the slowest run. Warm-to-cold and edit-to-cold ratios stayed within a few percent across all three runs, so they are the reliable tight guard.

Changes

• The warm-to-cold ratio guard tightens from 0.67 to 0.65. The observed maximum is 0.585.
• New edit-to-cold ratio guard of 0.60 for leaf and macro edits. The observed range is 0.48–0.52.
• Absolute limits become about 1.15x the slowest of the three runs, as a backstop against gross regressions:

Models	Cold	Warm	Edit
3000	14.5s	8.25s	7.25s
5000	23.5s	13.5s	12s
10000	42.5s	24.5s	22s

Slowest CI times:

• 3000: cold 12.65s, warm 7.12s, edit 6.36s.
• 5000: cold/config 19.35s, warm 11.14s, edit 9.84s.
• 10000: cold 37.15s, warm 21.28s, edit 19.01s.

Verification

• ruff, ty and fensu pass on the changed files.
• This PR's cache-compile CI jobs exercise the new limits and ratio guards.