WriteToReadOnlyFile was listed in MalformedErrors, so ClassifyOutcome/
MapOutcome routed it to PermanentFail and the store-and-forward sink
dead-lettered every alarm event in the batch. But WriteToReadOnlyFile is
a connection-configuration fault (the write session was opened without
ReadOnly = false), not an event-payload fault — treating it as permanent
silently and permanently discards alarm events on a misconfigured or
regressed connection, which is data loss.
Move WriteToReadOnlyFile from MalformedErrors into ConnectionErrors. The
batch loop now aborts the batch, resets the connection (so the reconnect
path re-opens a writable ReadOnly = false session), and defers the
events as RetryPlease for the next drain tick.
Updated the ClassifyOutcome theory data and added a dedicated regression
test pinning WriteToReadOnlyFile -> RetryPlease.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
f982fa1f69