mirror of
https://github.com/PerpetualSoftware/pad.git
synced 2026-09-23 02:53:31 +00:00
xarmian
be47896e13
Codex round 16. The positive-id check lived inside the PREFIXED branch of decodePayload, so a bare payload carrying id 0 or a negative was accepted: delivered with no SSE cursor for the client to advance to, and — once an epoch has been adopted — read as the sequence going backwards and used to discard every replay buffer. The check is now a function both branches call, which is the by-construction form rather than a second copy that can drift. Same route as an unreadable payload: this workspace's coverage ends and the next resume says so. The failing case that made this worth finding is the one my own round-11 test did not cover — it exercised the prefixed form only, because that is the form the guard was written next to. Symmetry asserted in the rejections table now, for both forms. Claude-Session: https://claude.ai/code/session_01JVDBKbgn3Xt7ndW1YoYd8X