~anthropic/claude-sonnet-latest → B (6:4)

Side B fixes real, concrete bugs (str/split with a string instead of regex, .getInputStream vs .getRequestBody, nil state/token causing NPEs, missing exception handling causing silent test failures) that were actively breaking OAuth E2E tests — these are genuine correctness fixes with lasting value. Side A adds a plausible feature (pool-scoped voting) with reasonable plumbing, but it's more speculative feature work with less certainty of correctness (e.g., new endpoint behavior, no new tests for the pool path) compared to B's targeted, verifiable bugfixes restoring broken test infrastructure.