[argszero]

Summary

#433 shipped cache-write (cache-creation) billing with one test on each half of the feature:

• both_usage_parsers_read_anthropic_cache_creation_tokens feeds the parsers only;
• settle_bills_cache_write_tokens_at_the_configured_price feeds settle_usage directly.

The propagation between those halves was crossed by no test. This PR adds the missing end-to-end coverage: one test per path, each posting a real request through the router to a fake Anthropic upstream and reading the ledger back.

path	propagation point	new test
non-streaming	parse_usage four-tuple → settle_usage (4th arg)	e2e_anthropic_cache_creation_reaches_the_ledger
same-protocol streaming	UsageCapture::finish four-tuple → flow-tail settle_usage	e2e_anthropic_stream_cache_creation_reaches_the_ledger

Each asserts:

• transactions.cache_write_tokens carries the upstream usage.cache_creation_input_tokens;
• usage_records.cache_write_tokens carries the same count;
• tokens still means input + cached + output (the write segment stays in its own column, as feat(billing): bill cache-write (cache-creation) tokens #433 specified);
• the charge equals the configured price × write tokens.

Why the price must be configured in the test

Every shipped model sets cache_write_input_per_m = 0, so at the default price the cache-write term is zero and a test cannot tell "the token reached the ledger" from "the token was dropped". Both tests therefore set a non-zero price on their own models row and assert the money alongside the count.

Measured teeth (not just green)

Each seam was mutated on its own, and each mutation reddens exactly one test:

• non-stream settle_usage 4th argument → 0.0: e2e_anthropic_cache_creation_reaches_the_ledger fails at the cache_write_tokens assertion (left: 0.0, right: 400.0), the streaming test stays green.
• streaming settle_usage last argument → 0.0: e2e_anthropic_stream_cache_creation_reaches_the_ledger fails at its cache_write_tokens assertion (left: 0.0, right: 400.0), the non-stream test stays green.

Before this PR both mutations left the whole suite green.

Related Issue

Follow-up to #433, which is already merged — this closes the coverage gap that feature left, it does not change behaviour. No production code is touched: the diff is test-only (src/gateway.rs, +226 lines inside mod tests).

Tests

• Existing tests pass — cargo test: 590 passed, 0 failed (588 before this PR)
• New tests added — the two end-to-end tests above
• cargo fmt --check clean
• cargo clippy --all-targets -- -D warnings clean