Review finding. The comments described MAX_DELIVERY_ATTEMPTS as a per-batch
budget mirroring the Python core's. It is not: Python puts the attempt count in
the claim filename so it follows one batch, while this counter lives in the
closure and counts consecutive failed flushes, so events captured during an
outage join the same queue and are dropped with it.
Per-batch accounting would mean an attempt count on every event. The queue is
already bounded, so the simpler rule stands; the comments now describe it rather
than the Python one.
Claude-Session: https://claude.ai/code/session_01C7tEmH86HAr7GoAAKCEHZb
Saket Aryan
7900a5d8d1