fix(routing): price InputInflation tokens as cache_read on cache branch

[SakuraPuare]

Summary

Upstream-injected inflation tokens (system prompts, tool schemas) are stable per session and share the provider cache lifecycle with the reusable prefix. The current code puts them in suffix alongside the newly appended user turn and prices them at pricing.InputPerToken (the full input rate). For any cache-supporting upstream this over-estimates the cache branch cost by roughly (InputPerToken - CacheReadPerToken) * inflation_tokens * n per window — about 10× the correct read price for typical Anthropic-style pricing.

Impact

For a session with InputInflation ≈ 1.8 and 305 requests / 128k prefix (a real production trace), the cache branch was estimated at ~41 units. The correct cost, computed with the same coverage/hit-rate observations but pricing inflation as cache_read, is roughly half that. This can make caching upstreams lose the cost race to no-cache alternatives that are actually more expensive in reality.

Fix

• Preserve originalInputTokens before the inflation multiply.
• suffix is now derived from the original count — only the newly appended user turn, which is never cacheable.
• Inflation flows into cacheable = prefix + inflationTokens. CoverageRatio then splits that pool as before.
• hits × cachedPortion × CacheReadPerToken and misses × cachedPortion × CacheWritePerToken now include the inflation share.
• no_cache branch is unchanged: on non-caching upstreams the inflation still bills at full input via features.InputTokens in NoCacheInputCost.
• breakEvenRequests mirrors the fix for consistency.

Tests

Two regression tests, both in internal/routing/cost_test.go:

1. TestEstimateWindowCostInflationBilledAsCacheRead — 1.8× inflation with hit_rate=1, coverage=1. CacheInputCost must stay ≈ 0 (all inflation goes through cache), CacheReadCost must match hits × 180 000 × 5e-7, CacheWriteCost must match 1 × 180 000 × 6.25e-6, cache path must win over no-cache.
2. TestEstimateWindowCostTrueSuffixStillFullPrice — 105k input with 100k reusable prefix and 1.5× inflation. Ensures the 5 000-token user turn is still billed at full InputPerToken (only inflation moves, not real suffix).

Full go test ./... and go vet ./... pass on Go 1.24.

Empirical verification

I validated the direction with a real A/B against production upstreams (Anthropic via a proxy). Same 144k-token payload sent to a no-cache endpoint and a cache-enabled endpoint. Anthropic usage reports 575k billed tokens for the cache endpoint (with 405k cache_read_input_tokens after warm-up) and 575k input_tokens for the no-cache endpoint. So the inflation tokens do bill — but on cache-supporting endpoints the majority is charged at cache_read, not full input, exactly what the new model reflects.