Summary

Four cases where Chronicle reported success while quietly doing nothing.

Changed

[OnceOnly] on a reactor handler method now takes effect — the handler is skipped for replayed events while the rest of the reactor still replays. Only class-level placement was honored before.
The kernel now warns when it skips a concurrency check because the append declared a concurrency scope with no expected sequence number.

A 16.12 or newer kernel can now read constraint definitions written by an older kernel, so an in-place upgrade no longer breaks constraint registration and client startup.
Recovering a failed partition no longer clears it without re-running the handler, silently losing the missed side effect.
The observer management API's try-recover-failed-partition and replay routes were swapped, so recovery replayed and replay retried.
A job step that handled no events no longer overwrites the sequence number recorded by an earlier step that did.