fix: lag ratio guards only engage when churn latency is actually slow

[arzafran]

What this does

Stops the tests-build-and-lag job failing on runs whose typing latency is objectively excellent. Six unrelated PRs today failed on ratio blowups where churn p95 was 13–28ms against a 35ms cap — fast baselines make the ratio explode on a few ms of runner jitter.

Summary

• Ratio checks (p95 1.70x / avg 1.70x) now also require the churn result to be slow enough for a ratio to be meaningful: churn p95 ≥ 20ms / avg ≥ 14ms (env-tunable, PROGRAMA_LAG_MIN_CHURN_*_MS_FOR_RATIO)
• The delta caps (+20ms/+12ms) and absolute cap (35ms) are untouched — a genuine regression on a fast baseline still fails via deltas (3x from a 7ms baseline = +14ms p95… a 4x = +21ms trips it; and any churn above 20ms re-enables the ratio check)
• Of today's six flakes, five would have been skipped by this gate; the sixth (churn 27.6ms) still enforces — correctly, as the only genuinely slow run

Test Plan

• python syntax
• tests-build-and-lag on this PR