v0.7.2

v0.7.2 (2026-07-19)

This release is published under the MIT License.

Bug Fixes

• Release the credential when stream teardown fails (b9d5cb5)

finalize() marked itself done, then awaited the upstream stream's aexit, and only afterwards ran the finalizers. The finalizers are what release the key locked by key_concurrency: false, so if that await raised — a broken upstream connection, or cancellation from the request timeout — the release was skipped. finalized was already True, so no later call could recover it, and nothing expires a key lock: the key left rotation permanently. With a two-key pool that is an outage.

Run the finalizers from a finally block, and swallow errors within each so one bad finalizer cannot strand the rest.

Add e2e coverage for credential release under key_concurrency: false, driving a real proxy process against a real upstream — client hangs up mid-stream, upstream fails mid-stream, request timeout cancels a stalled stream, and the happy path so a fix cannot work by never locking. These already passed before the change, which is worth recording: the paths I could drive end to end were releasing correctly, and the defect above is one I could only reproduce by forcing teardown to fail. The two unit tests do fail without the fix.

Also document the behaviour that actually looks like a stuck queue. Key rate limits are not a fast-fail path like the IP and user quotas: a request waits for a key until the queue expiry and then 504s. Sustained demand above keys x key_rate_limit therefore leaves a permanently full queue, and a max_size far above rate x expiry only buys a longer wait before the same failure. Measured: 1 key at 1/2s with a 6s expiry serves 3 of 12 concurrent requests and expires the other 9 at 6.1s.

Test harness gains key_concurrency and request_timeout_seconds knobs and a stream-hang upstream endpoint.

---

Detailed Changes: v0.7.1...v0.7.2