[argszero]

Summary

Fixes the P0 cache-billing bug (rant 2026-08-23T08:20:38): DeepSeek's native top-level prompt_cache_hit_tokens spelling was silently dropped by serde, so cache hits were billed at the full cache-miss price (~30x overcharge), and input tokens were not disjointed (double-billing risk).

Changes

1. StreamUsage (src/sse.rs): added #[serde(default)] prompt_cache_hit_tokens: u32 — DeepSeek's native top-level spelling (prompt_tokens = prompt_cache_hit_tokens + prompt_cache_miss_tokens).
2. extract_cache_read_tokens: now supports all three spellings with priority DeepSeek → OpenAI → Anthropic (prompt_cache_hit_tokens → prompt_tokens_details.cached_tokens → cache_read_input_tokens).
3. All 6 record_usage call sites: input is now disjoint — prompt_tokens.saturating_sub(cached) (or .max(0.0) for the f64 paths) — so cached tokens are never billed twice (miss price + hit price).
4. Downstream forwarding (build_anthropic_usage_json): already disjoints (input_tokens = prompt_tokens - cached - cache_creation) and is now effective since cached is actually extracted for DeepSeek.

Tests

Added 3 spelling tests, each asserting the (input, cached, output) triple:

• deepseek_cache_hit_tokens_spelling — {prompt_tokens:100, prompt_cache_hit_tokens:90, completion_tokens:50} → (10, 90, 50); also verifies downstream input_tokens:10 + cache_read_input_tokens:90
• openai_cached_tokens_spelling_disjoint — OpenAI prompt_tokens_details.cached_tokens → (10, 90, 50)
• anthropic_cache_read_spelling_disjoint — Anthropic cache_read_input_tokens → (10, 90, 50)

Related

Rant 2026-08-23T08:20:38 — 缓存计费 bug（P0，多收 ~30 倍）

Verification

• cargo test — 145/145 pass (142 baseline + 3 new)
• cargo fmt --check clean
• cargo clippy --all-targets -- -D warnings clean

Explicitly NOT changed (per rant)

• Price table untouched (values match official CNY pricing)
• No new DB fields/migrations (reuses existing cached_tokens column)
• No reasoning-token / output-billing changes