H4: exception 01/02/03 (and undecodable data) are device refusals - per-tag
Bad with a once-per-tag Warning, no Degraded, and they advance
LastSuccessfulRead (the device answered), so an all-refused driver no longer
trips the 90 s watchdog. 04/05/06 still degrade (answered); 0A/0B degrade
without advancing. Health is one worst-wins write per batch instead of
last-writer-wins per tag. A comms failure on one unit fails the rest of
that unit's batch fast; an unreachable endpoint fails the whole batch fast.
H7: write transport failures map to BadCommunicationError, value overflow and
array length mismatch to BadOutOfRange, conversion failures to
BadTypeMismatch; writes report write-side health (a good write or a refusal
clears a write-side Degraded; a good read does not).
Lifecycle: every health write and the transport/table publish carry a
generation; a superseded init releases its transport and publishes nothing,
a read straddling a re-init cannot degrade the new transport, and the
shutdown's final write cannot clobber a newer init. The tag table is
build-and-swapped (re-init no longer merges into the old table). Teardown
joins the probe/re-probe loops (bounded); a probe answer advances
LastSuccessfulRead and an exception PDU from the device counts as Running.
Transports latch dispose atomically and never reconnect after it.
Refs #741
Claude-Session: https://claude.ai/code/session_01Ub9hGmYLQvn9XETzw7CMsT
P3: a torn-down socket is re-dialled inside the gate (one attempt per backoff
window) instead of failing fast with 'Transport not connected' until the 90 s
comms-loss watchdog re-initialised the driver.
L3: ConnectAsync disposes the previous/failed TcpClient, DNS runs under the
per-op timeout, and the reconnect never sleeps inside the gate (the old
10-attempt backoff held it for ~81 s).
H9: an MBAP response with no PDU, or an exception PDU with no code, is a
TruncatedFrame desync rather than an IndexOutOfRangeException.
Refs #741
Claude-Session: https://claude.ai/code/session_01Ub9hGmYLQvn9XETzw7CMsT