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

B fixes real, concrete bugs (str/split with a string instead of regex, reading from the wrong stream, -1 vs 0 content-length, unhandled nil token/state causing crashes) that were actually breaking E2E auth tests, plus adds exception handling so the mock server doesn't die silently. A is a larger feature/refactor (theme cookie plumbing, room-scoped URL rewriting) which is plausibly useful but is unreviewed new surface area with added complexity across many files, whereas B's changes are small, targeted, and verifiably correct fixes to broken test infrastructure.