fix(opcua): replay writes deferred during an off-thread plan; cache-ref and guard fixes (Lane B fix round 1)

C1 (critical): B6 computes the plan off the pinned thread while B1 made writes
non-creating, so a value/status/alarm write for a node the in-flight plan adds
was skipped, and change-driven drivers never republish it (BadWaitingForInitialData
indefinitely; on a full rebuild every re-subscribe value landed on discarded
nodes). OpcUaPublishActor now applies writes immediately AND remembers the
latest per (realm, id) while a plan is in flight or queued (value+status
compose; alarm state/quality coalesce to the latest, which the node's delta
gate makes inert on an existing condition), replays it after each apply before
answering, and drops it once the pipeline is idle. Bounded (250k keys by
default; beyond it new keys are not held, logged once).

I2: the data-plane-authority parse runs in its own try inside the plan task; a
throw is logged and never degrades the address-space apply.

I3: the SDK's DeleteMonitoredItems releases an item's component-cache
reference on a Good delete; a tombstoned item's reference was already
released, so the delete took one from the re-created node's entry. Released
items are remembered and the orphan delete takes one back first. Event items
never held a reference (EventManager + SubscribeToEvents), so the three
event-item releases are gone.

I4: direct tests for the HistoryUpdate override's per-request updater check
and the HistoryReadEvents override's source guard (both reached by restoring
the bit after a reset - the B7 race window); fixed the false remark.

Minors: the MaterialiseAlarmCondition kind-swap tombstones the dropped
instance; a bytes-less rebuild arriving mid-plan queues behind it (replies in
order); a PlanReady/PlanFailed from a previous actor incarnation is ignored and
its requester answered degraded; the tie-cluster end bound of an uncapped read
fits against the backend ceiling (defence in depth - guarded by a sweep test);
the lock-time benchmark stays in the default run, headroom noted.

Refs #732
Claude-Session: https://claude.ai/code/session_01Ub9hGmYLQvn9XETzw7CMsT

tests/Server/ZB.MOM.WW.OtOpcUa.OpcUaServer.Tests/NodeManagerLockTimeBenchmarkTests.cs

@@ -20,6 +20,11 @@ namespace ZB.MOM.WW.OtOpcUa.OpcUaServer.Tests;
 ///     </list>
 ///     The wall-clock budgets are generous for CI hardware; the before/after numbers are in the commit
 ///     message. P3 is pinned by the bookkeeping's own visit counter, which does not depend on the machine.
+///     Kept in the default run on purpose (not a Benchmark category the CI filter would have to learn to exclude):
+///     measured after the fix at 5–6 ms (P1) and 58–64 ms (P2) on a loaded dev machine against the 1 s budget —
+///     ≥ 15× headroom, so it does not flake on slow CI. The flip side: the pre-fix code (75–150 ms / 470–680 ms) also
+///     fits 1 s at these sizes, so the wall-clock legs guard against gross (order-of-magnitude) regressions only; the
+///     machine-independent P3 visit counter is the strict pin.
 /// </summary>
 [Trait("Category", "Benchmark")]
 public sealed class NodeManagerLockTimeBenchmarkTests(ITestOutputHelper output) : IDisposable