Live run against ab_server (slc500/micrologix/plc5 x libplctag/native): with
coalescing on, the blackholed device's 20 tags are one block, i.e. one comms
fault per poll, so its backoff window opens on the second poll rather than the
first. The gate now asserts the actual contract: at most two timeouts across
three polls, the live device Good on every poll, and the third poll fast.
Refs #738
Claude-Session: https://claude.ai/code/session_01Ub9hGmYLQvn9XETzw7CMsT
Review 07 P7 (High): reads walked every reference of every device serially
with no backoff, so a dead device cost Timeout per tag per poll (50 tags x 2 s
= 100 s) and starved every healthy device long enough to trip the comms-loss
watchdog against a reachable one.
- ReadAsync buckets references by device and reads up to 8 devices
concurrently; each device's tags are still read in order.
- Each device has a ConnectionBackoff (1 s doubling to 30 s). Two consecutive
comms faults mark it down: the rest of that pass and every later pass fail
fast with BadCommunicationError and touch no wire until the window expires or
any answer (data path or probe) resets it. One timing-out tag does not
blacken its answering siblings.
- Cold connects are single-flight per device (ConnectGate), replacing the
per-key creation locks; a cached handle never touches the gate.
- A status/exception proving the device ANSWERED with a deterministic refusal
(not found, not allowed, out of bounds, type/size mismatch) is per-tag Bad,
keeps its handle, warns once per tag and advances LastSuccessfulRead instead
of evicting and degrading. Writes classify the same way.
- Health is written once per read: Degraded if any device faulted, Healthy if
anything read Good, LastSuccessfulRead advanced when any device answered.
Adds an env-gated ab_server case with a blackholed sibling device.
Refs #738
Claude-Session: https://claude.ai/code/session_01Ub9hGmYLQvn9XETzw7CMsT
Joseph Doherty