Review r0 I2 + M1/M2/M4/M7/M10.
- I2: one health writer (ApplyHealth) for reads AND writes, worst wins per
call: a comms fault => Degraded; otherwise any answer (Good value, device
refusal, Good or refused write) => Healthy; LastSuccessfulRead advances on
any answer. A refusal-only read used to leave a recovered driver Degraded,
and writes never touched health at all.
- M10: a no-reply write STATUS now degrades like the exception path did; a
Good/refused write records an answer on the device's backoff, a no-reply
one a fault (bit RMW included — it now reports its raw status).
- M2: a write inside the device's backoff window fails fast with
BadCommunicationError even when its handle is cached.
- M1: ShutdownAsync/Dispose write their final Unknown only while the epoch
they retired is still current (a newer initialise owns health otherwise).
- M4: an indirect bit write releases the previous index value's parent-word
handle (the per-word RMW gate is deliberately kept: removing a gate another
writer may hold would let two RMWs of one word race).
- M7: backoff decisions read an injectable Clock; the isolation tests drop
their stopwatch asserts and cross the window on a manual clock.
Refs #738
Claude-Session: https://claude.ai/code/session_01Ub9hGmYLQvn9XETzw7CMsT
Joseph Doherty
17ae2eea0b