fix: warmup cooldown timing + rising cost threshold + accurate TTL pricing

[BYK]

Summary

Fixes cache warming economics — warmup was costing $65.89 but only saving $4.43 (net -$59.50). The dominant issue was a cooldown timing bug causing every continuation warmup to be a cold cache write ($1.25) instead of a cheap refresh ($0.10).

Root Cause

The normal-path cooldown was set to ttlMs (300s for 5m TTL) instead of ttlMs - warmupMarginMs (255s). Since warmups fire in the margin window (4:15–5:00), a 300s cooldown means the next warmup fires right as the previous cache expires — every warmup is a cold write. The forced/tool-call path already used the correct tighter cooldown.

Changes

Fix 1: Cooldown timing (the $65 → $6.50 fix)

• Uses ttlMs - warmupMarginMs as cooldown for all paths, not just forced/tool-call
• Each continuation warmup now fires while the previous cache is still alive (cache read ~$0.10 vs cold write ~$1.25)
• Also fixes cooldownFor() (used by dashboard diagnostics) to match

Fix 2: Rising cumulative cost threshold

• Adds cumulativeCostThreshold(k, read, write) — after k warmup cycles, break-even requires higher P(returns)
• Formula: P > k × read / (write - read) — rises linearly with cycles spent
• For Opus 5m: k=1 needs 8.7%, k=3 needs 26%, k=5 needs 43%, k=6 needs 52%
• At k=1 matches costThreshold() exactly; clamped to 1.0
• Applied in Phase B continuation to replace the flat 8.7% threshold

Fix 3: TTL-aware cost recording

• recordWarmupCost() and recordWarmupHit() accept optional ttl parameter
• For 1h TTL sessions, doubles cache_write to match Anthropic's actual pricing
• Previously, displayed costs and savings were undercounted by up to 50% for 1h sessions
• recordTTLSavings intentionally uses base rate (counterfactual is against 5m, not 1h)

Fix 4: Session-level ROI guard

• After 10+ total warmups, if session hit rate < 20%, warming stops entirely
• Applies to both normal path AND tool-call fast path
• Prevents unlimited spending across many short breaks where per-break checks pass but aggregate ROI is negative
• Dashboard notWarmingReason updated to show ROI guard and rising threshold rejections

Tests

• Updated cooldown test to verify the fix (260s > 255s allowed, 250s < 255s blocked)
• Added cumulativeCostThreshold() tests: formula values, k=1 matches costThreshold(), clamping, edge cases
• Added session ROI guard tests: low hit rate rejected, high hit rate allowed, skipped below 10 warmups, applies to tool-call path

Impact Estimate

For the session that triggered investigation (410 turns, 65 warmups, 12 hits):

• Before: $65.89 warmup cost, $4.43 savings → net -$61.46
• After (cooldown fix alone): ~$6.50 warmup cost → net -$2.07
• After (all fixes): Rising threshold stops warming around cycle 5-6, ROI guard prevents chronically unprofitable sessions