fix(gateway): recognise DeepSeek's cache-hit token spelling for non-streamed responses

[argszero]

Summary

Non-streamed responses from an OpenAI-protocol upstream that reports cache hits with DeepSeek's native spelling (prompt_cache_hit_tokens) are currently billed as if nothing was cached: cached_tokens is recorded as 0 and the cached input is charged at the full cache-miss rate.

The streaming paths already recognise this spelling — UsageCapture::finish (cached 三拼写兼容, rant 2026-08-23T14:05:02) and sse.rs::extract_cache_read_tokens both try DeepSeek-native first, then OpenAI, and all six sse.rs::record_usage call sites were fixed for it. CHANGELOG.md scopes both of those fixes to the passthrough/streaming path (v0.7.10 / v0.7.11), so the non-streaming parser is a missed sibling rather than a deliberate exemption. This PR closes that gap.

Related Issue

No tracking issue exists for this; none is fabricated. Verified closingIssuesReferences is empty on purpose.

Changes

• src/gateway.rs::parse_usage, openai_chat arm: the cached lookup now tries usage.prompt_cache_hit_tokens (DeepSeek native) first and falls back to usage.prompt_tokens_details.cached_tokens (OpenAI) — the same priority the two streaming implementations use. The disjoint return (prompt_tokens − cached).max(0.0) is unchanged, as is every other protocol arm.
• src/gateway.rs: the doc comment above parse_usage (from 2026-08-20T10:17:27) listed only two cache spellings for openai_chat and was stale after this change; it now names both.
• No configuration or data-structure change, so no example-file sync is needed.

Measured impact

Fixture: prompt_tokens=1000, prompt_cache_hit_tokens=900, completion_tokens=50, prices 10 / 20 USD/M for miss / output (with a cache-hit price of 2 USD/M).

path	cached_tokens	cost
non-stream, DeepSeek spelling — before	0	0.0792
non-stream, DeepSeek spelling — after	900	0.02736
non-stream, OpenAI spelling	900	0.02736
stream, DeepSeek spelling	900	0.02736

So the same request is charged 2.9x more before the fix. With no cache-hit price configured at all (the shape the earlier probe used), the ratio is 5.5x, because then the entire input is charged at the miss price. Under DeepSeek's real peak rates (3.0 / 0.10) a cache hit is effectively billed at the miss price, i.e. its own unit price is wrong by 30x.

This is an overcharge, and it also corrupts usage_records.cached_tokens — the column the transaction page renders as the cache-hit breakdown. It is reachable on the plain DeepSeek path: deepseek-paygo exposes openai_chat at https://api.deepseek.com, and DeepSeek's native non-streaming usage uses the top-level spelling. The code path depends on the upstream request shape, not the client's stream flag, so an anthropic- or responses-protocol client cross-translated to an openai_chat upstream hits parse_usage as well — and there the user sees cache_read_input_tokens: 900 in the response while being billed for a miss.

Tests

• cargo test 全部通过 — 175 passed / 0 failed (baseline 172 + 3 new)
• cargo fmt --check 通过
• 新增/更新了单元测试（如适用）

Added tests:

• parse_usage_openai_deepseek_native_cache_spelling — the DeepSeek-native body must yield (100, 900, 50), with the OpenAI spelling as a positive control and a third case proving the DeepSeek spelling wins when both are present (same priority as streaming).
• usage_parsers_agree_on_cache_spelling — feeds the same three usages to both parse_usage and UsageCapture::finish and asserts the two parsers return identical triples, so the stream/non-stream split cannot silently diverge again. The third case also pins that a body with no cache field still means "no cache" (cached = 0) — that is absence-of-a-field, not the "upstream reported nothing" case C2037 addressed.
• e2e_nonstream_deepseek_cache_spelling_billing — end-to-end through the real forward path against a fake upstream emitting the DeepSeek spelling, asserting usage_records.cached_tokens = 900, tokens = 1050, the settled cost, and the consumer/owner balance split.

A/B verification (reverting only the cached lookup, keeping the tests) turns exactly the new assertions red while all four pre-existing parse_usage tests stay green, and the failure prints the defect: left: (1000.0, 0.0, 50.0) / right: (100.0, 900.0, 50.0).

Checklist

• 分支命名符合约定（feat/ /fix/ /docs/ ...）
• Commit message 使用 Conventional Commits 格式
• 单一职责，改动最小化

Deliberately out of scope: refactoring the three protocol arms into a shared helper (they legitimately differ — anthropic's input_tokens is already disjoint while openai/responses are not), handling prompt_cache_miss_tokens, and any new gate (a missing recognition in one function is caught by a test, not a scanner).