GeezerVison

A rejected update_settings() call can leave its valid prefix queued even though no SETTINGS frame was emitted. A later acknowledgement then applies an unsent value, contrary to the SETTINGS synchronization described in RFC 9113 section 6.5.3: in a paired client/server reproduction, one side changes its concurrent-stream limit to 3 while the peer still knows 100.

I validate the complete batch against the existing rules before queueing any values. This preserves the existing error type, code and message, connection-state error precedence, valid update ordering, and unknown setting codes. The change is limited to InvalidSettingsValueError validation failures.

Twelve new cases fail before the fix; all 16 regression/control cases pass afterward. The full suite passes 1,678 tests with 100% statement and branch coverage on Python 3.10, 3.12 and 3.13. Ruff, strict-bytes mypy, manifest/package checks, strict distribution checks and the documentation build pass. A fresh installed wheel also passes 12 paired-peer checks outside the checkout. No sockets or external service are needed for these tests.