[argszero]

Summary

Fix incomplete P0 cache-billing fix from PR #128 (rant 2026-08-23T08:20:38): PR #128 only patched src/sse.rs (cross-protocol conversion path), but the same-protocol passthrough path — UsageCapture::finish() in src/gateway.rs — was missed. Per host's rant 2026-08-23T14:05:02, cached tokens were double-counted into the total (observed: per-call uncached 241,776 instead of 112, total ≈2× on dev).

Root cause (gateway.rs openai branch of finish())

• input was set to the full prompt_tokens (includes cache-hit portion) — not disjoint
• cached only recognized OpenAI's prompt_tokens_details.cached_tokens spelling — DeepSeek's native top-level prompt_cache_hit_tokens was dropped (which made the miss even worse on DeepSeek responses)

Changes

1. UsageCapture::finish() openai branch: cached extraction now three-spelling compatible with the same priority as extract_cache_read_tokens in sse.rs — DeepSeek native prompt_cache_hit_tokens → OpenAI prompt_tokens_details.cached_tokens
2. Disjoint input at return (covers openai + anthropic branches uniformly): (input - cached).max(0.0) — matches the already-fixed logic in sse.rs; max guards underflow
3. Tests (extended usage_capture_parses_openai_and_anthropic):
   • DeepSeek spelling: {"prompt_tokens":100,"prompt_cache_hit_tokens":90,"completion_tokens":50} → input=10, cached=90
   • OpenAI spelling: prompt_tokens_details.cached_tokens=90 → input=10, cached=90
   • Anthropic: cache_read_input_tokens=70 → input=10, cached=70
   • Original no-cached cases still pass (10-0=10, 80-0=80)

Expected effect (correct behavior, not a bug)

• Points drop sharply: input billed at ~100 uncached + ~241k cached rate (0.05/1M) instead of 241k full rate
• Transaction record: 输入(非缓存)≈100, 输入(缓存)≈24万, 总 Token≈24万 — consistent with real EMRG usage

Related

• Rant: 2026-08-23T14:05:02.990820+08:00 (P0 缓存计费修复不完整：UsageCapture.finish() 同协议透传路径漏改)
• Follow-up to PR fix(gateway): DeepSeek cache-hit tokens billing + disjoint input (rant 2026-08-23T08:20:38) #128 (326d00f)

Tests

• cargo test 145/145 pass (incl. new cached cases)
• cargo fmt --check clean
• cargo clippy --all-targets clean