Status fields (_lastDrainUtc, _lastSuccessUtc, _lastError, _drainState,
_evictedCount) were written by the drain timer thread and read by
GetStatus() / health-check threads with no memory barrier, risking torn
DateTime? reads and stale DrainState observations.
- Added _statusLock object; all writes to status fields now happen inside
lock(_statusLock) blocks in DrainOnceAsync and DrainTimerCallback.
- GetStatus() snapshots all fields atomically under the same lock so the
Admin UI / /healthz endpoint always sees a consistent view.
- Regression test GetStatus_snapshot_is_consistent_under_concurrent_drain
drives status writes and reads from concurrent threads; asserts no throws.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
System.Threading.Thread is in the System.Threading namespace (not
System.Threading.Thread), so the existing ForbiddenNamespacePrefixes
entry "System.Threading.Thread" never matched — the namespace prefix
check compared against the type's containing namespace, which is
System.Threading. Move Thread into ForbiddenFullTypeNames (alongside
Environment / AppDomain / GC / Activator) where it is matched by exact
fully-qualified type name, which actually fires. Remove the dead
namespace-prefix entry and document why. The Rejects_Thread_new_at_compile
test now passes. (Core.Scripting-010.)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EnsureConnectedAsync now disposes and nulls any existing non-connected
client before creating a fresh one via _clientFactory.Create().
Previously the method reused a cached client via ConnectAsync, but a
client disposed by a HandleRecycle race or prior teardown would hit
FocasWireClient.ThrowIfDisposed on every subsequent call, leaving the
device permanently wedged with BadCommunicationError and no recovery
path until ReinitializeAsync.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add six previously-missing edge-case tests to WonderwareHistorianClientTests:
(2) WriteBatchAsync transport-drop catch path returns RetryPlease for all events;
(3) InvokeAsync second-attempt-also-fails propagates the exception;
(4) stalled sidecar fires OperationCanceledException within CallTimeout;
(5) HistoryAggregateType.Total throws NotSupportedException via ReadProcessedAsync;
(6) sidecar wrong-MessageKind reply throws InvalidDataException.
Extend FakeSidecarServer with DisconnectBeforeReply, ReplyWithWrongKind, and
StallAfterRequest test knobs to support these scenarios.
Add ContractsWireParityTests.cs (11 tests) to pin the MessagePack byte layout,
round-trip correctness, MessageKind enum values, and Framing constants — catching
silent [Key] index drift between the client and sidecar mirror copies without
requiring a cross-TFM (net10 vs net48) project reference.
Test count grew from 11 to 27; all 27 pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Guard all _health field accesses with Volatile.Read / Volatile.Write.
ReadAsync, WriteAsync, and ProbeLoopAsync run on different threads and
several updates are read-modify-write (new DriverHealth(_, _health.X, _)).
Without volatile semantics a concurrent update can be lost or a stale
LastSuccessfulRead timestamp propagated. DriverHealth is an immutable
record so Volatile is sufficient — no lock needed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
InitializeAsync catch block now mirrors ShutdownAsync teardown: cancels
and disposes probe CancellationTokenSources, calls DisposeRuntimes, and
clears _devices/_tagsByName before rethrowing. A caller that catches and
abandons (rather than retrying via ReinitializeAsync) no longer leaves
orphaned probe tasks or libplctag handles alive.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add regression tests for the Medium findings resolved in this series:
- AbCipDataType_maps_large_integer_types (theory: LInt→Int64, ULInt→UInt64,
UDInt→UInt32) and Read_UDInt_tag_returns_uint_value_not_negative_wrapped_int
cover Driver.AbCip-004.
- Structure_parent_tag_is_not_readable_after_member_fan_out,
InitializeAsync_throws_on_duplicate_tag_name, and
InitializeAsync_throws_when_member_name_collides_with_independent_tag
cover Driver.AbCip-005.
- Read_failure_evicts_runtime_so_next_read_creates_fresh_handle covers
Driver.AbCip-010.
AbCipDriverTests.AbCipDataType_maps_atomics_to_driver_types extended with
LInt/ULInt/UDInt assertions.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
DiscoverAsync now unconditionally emits SecurityClassification.ViewOnly
for every user-authored FOCAS tag. Previously the SecurityClass was
tag.Writable ? Operate : ViewOnly, but WireFocasClient.WriteAsync always
returns BadNotWritable — advertising Operate misleads OPC UA clients
and the DriverNodeManager ACL layer into granting write permission on
nodes that can never be written.
Updated FocasCapabilityTests.DiscoverAsync_emits_pre_declared_tags to
assert ViewOnly for the writable-by-config tag so it matches the
corrected behaviour.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Mark _health volatile. The record-reference assignment is atomic, but
without an acquire/release memory barrier GetHealth() on another thread
can observe a stale snapshot indefinitely. volatile enforces the barrier
at read and write sites without a lock.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Throw InvalidOperationException at InitializeAsync when a tag's
DeviceHostAddress does not match any entry in the Devices list, naming
both the tag and the unresolved host. Previously the missing-device
check was guarded by a TryGetValue so a typo silently bypassed
capability-matrix validation and deferred the error to per-read
BadNodeIdUnknown — the opposite of the documented "fail at load" goal.
Also resolves findings 004, 005, and 006 in the same file:
- 004: DiscoverAsync now unconditionally emits ViewOnly for all user
tags; the Writable config field no longer influences security class
because the wire backend always returns BadNotWritable.
- 005: All _health reads use Volatile.Read and all writes use
Volatile.Write so concurrent readers observe a consistent reference
and read-modify-write sequences capture a stable snapshot.
- 006: EnsureConnectedAsync disposes and nulls any existing
non-connected client before creating a fresh one, preventing
ObjectDisposedException loops after a HandleRecycle race or teardown.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Runtimes and ParentRuntimes changed from Dictionary to ConcurrentDictionary.
EnsureTagRuntimeAsync and EnsureParentRuntimeAsync now use a per-key
GetCreationLock semaphore with a double-checked pattern: fast-path read
requires no lock; slow-path create+initialize+store is serialised per key
so a concurrent caller waits rather than creating a duplicate runtime that
would be leaked when DisposeRuntimes runs.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add `EvictRuntime` helper that removes + disposes a stale
`ConcurrentDictionary` entry. Call it from `ReadSingleAsync`,
`ReadGroupAsync`, and `WriteAsync` on non-zero libplctag status and
transport exceptions so the next call for the same tag re-creates a
fresh handle — mirroring the probe loop's recreate-on-failure pattern.
Value-conversion exceptions (NotSupportedException, FormatException,
InvalidCastException, OverflowException) are not transport faults and
do not evict the handle.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Apply _config.MaxValuesPerRead as a bucket cap in ReadAggregateAsync,
mirroring the existing cap in ReadRawAsync. Without this guard a processed
read over a wide time range with a small IntervalMs could accumulate an
unbounded HistorianAggregateSample list; if the serialised reply exceeded
the 16 MiB FrameWriter frame cap WriteAsync would throw and the client
correlation-id wait would hang. Truncation now logs a Warning with a hint
to widen IntervalMs or reduce the time range.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
DecodeValue for Bit with no bitIndex now reads the full 16-bit word via
GetInt16(0) and tests bit 0 instead of GetInt8(0), which only covered the
low byte and silently misread any bit in positions 8..15. The comment
explains the two decode paths (suffix-present vs suffix-absent).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add exponential backoff (250 ms → 500 ms → 1 s → 2 s → 4 s → 8 s cap) to
PipeServer.RunAsync after each connection-loop exception, replacing the spin
loop that previously pegged a CPU core and flooded the log on persistent errors
such as a duplicate pipe name or a failing PipeAcl.Create. After 20 consecutive
failures the method re-throws so the SCM / NSSM supervisor can restart the
sidecar cleanly. A clean connection (even a short-lived one) resets the counter.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
TryParse now rejects three classes of malformed PCCC address:
- Sub-element + bit-index together (e.g. T4:0.ACC/2) — never valid in PCCC
- File number on I/O/S system files (e.g. I3:0, S2:1) — single-letter only
- Sub-element on non-T/C/R files (e.g. B3:0.DN, N7:0.FOO) — only Timer,
Counter, and Control files carry structured elements
New helper predicates IsNoFileNumberLetter / IsSubElementFileLetter
keep the parser's intent clear. Regression tests added in AbLegacyAddressTests.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add ScriptSandboxTests cases for all forbidden-namespace deny-list
vectors that lacked test coverage: System.Threading.Thread,
System.Threading.Tasks.Task.Run (newly denied per Core.Scripting-003),
System.Runtime.InteropServices.Marshal, and Microsoft.Win32.Registry.
The 001/002 type-granular and node-form vectors were already covered by
the -001/-002 resolution commits. All 79 tests pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
In TimedScriptEvaluator.RunAsync, the catch (TimeoutException) block
now checks ct.IsCancellationRequested before throwing
ScriptTimeoutException, so a caller cancellation that races a timeout
deterministically surfaces as OperationCanceledException regardless of
which WaitAsync observes first. Regression test
Caller_cancellation_wins_even_when_timeout_fires_first added.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EnqueueAsync used synchronous SQLite I/O (conn.Open / ExecuteNonQuery /
COUNT(*)) on the caller's thread, blocking the alarm-emitting thread under
write contention with the drain worker. The cancellationToken parameter was
silently ignored.
- EnqueueAsync converted to genuine async: OpenAsync / ExecuteNonQueryAsync /
ExecuteScalarAsync used throughout; ct threaded to every await.
- ApplyPragmasAsync added alongside the existing ApplyPragmas helper so
the WAL + busy_timeout PRAGMAs are applied on the async open path too.
- EnforceCapacityAsync added to handle capacity eviction on the async path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
DependencyExtractor.VisitInvocationExpression now additionally checks
that the member-access receiver is the identifier "ctx" before treating
a GetTag / SetVirtualTag call as a ScriptContext dependency. This
prevents spurious dependencies when a script defines a local helper type
with a matching method name and calls it as other.GetTag("X"). Test
Ignores_member_access_GetTag_on_non_ctx_receiver added.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The ConcurrentDictionary + TryAdd/dispose-loser pattern for Runtimes
and ParentRuntimes was already applied as part of the Driver.AbCip-008
fix. Recording resolution with evidence rather than applying a
duplicate change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add System.Threading.Tasks to ForbiddenNamespacePrefixes so scripts
cannot use Task.Run / Parallel to spawn background work that outlives
the per-evaluation timeout. Document the unbounded-memory accepted
trade-off and the Task denial rationale in docs/VirtualTags.md (new
"Known resource limits" subsection) and cross-reference from
docs/ScriptedAlarms.md.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Extract the string-vs-numeric value selection from raw and at-time read
loops into a SelectValue helper method. aahClientManaged's HistoryQueryResult
has no data-type field in the bound SDK version, so the heuristic (prefer
StringValue when non-empty and Value==0) is unavoidable; the helper now
documents the limitation explicitly in its XML doc so the known edge case
(numeric tag at exactly zero with a formatted StringValue) is self-evident.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`PlcTagHandle` and `DeviceState.TagHandles` were dead scaffolding: the
`ReleaseHandle` no-op never called `plc_tag_destroy` and the dict was
never populated. Removed the file, the dead dict, and its
`DisposeHandles` loop. Updated the `AbCipDriver` class doc to document
that native lifetime is owned by libplctag.NET `Tag.Dispose()` (invoked
from `DisposeHandles`) with the library's own finalizer covering any
GC-collected instances. Two test methods that only exercised the dead
`PlcTagHandle` class removed from `AbCipDriverTests`.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`current & widthMask` was already applied in `WriteBitInWordAsync` by
the -001 High finding fix, making the 16-bit sign-extension hazard fully
neutralised. No further code change required; mark Resolved.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Normalise req.Events to Array.Empty<AlarmHistorianEventDto>() immediately
after MessagePack deserialization in HandleWriteAlarmEventsAsync. MessagePack
deserializes an absent or explicit-nil array field as null, not Array.Empty,
so a peer that sends a null Events array would trigger a NullReferenceException
on either .Length dereference (no-writer branch or catch block), leaving the
client correlation-id wait hanging with no reply frame.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Document explicitly that WriteBatchAsync never returns PermanentFail because
the WriteAlarmEventsReply wire contract carries only a bool-per-event (no
unrecoverable/transient distinction). Add a <remarks> XML block explaining
the structural limitation, why poison events retry rather than dead-letter,
and that a coordinated per-event status enum extension to the .NET 4.8
sidecar is a tracked follow-up. Add inline NOTE comments in both the
success and catch paths for discoverability.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Introduce DeserializeSampleValue() helper that enforces a 64 KiB per-sample
ValueBytes size cap before calling MessagePackSerializer.Deserialize<object>,
and documents that the default StandardResolver (primitive-only, no typeless
or dynamic-type resolution) is in use. Both ToSnapshots and AlignAtTimeSnapshots
route through the new helper. Add inline XML comments to the two NuGetAuditSuppress
entries in the csproj recording the advisory title, why each does not apply to
this module's primitive-only deserialization, and when to revisit.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Structure tags with declared Members no longer register the bare parent
name in `_tagsByName` — reading it would return Good/null, which is
misleading. Clients read individual member paths. Both the member
fan-out and the scalar-tag paths now perform a duplicate-key check that
throws `InvalidOperationException` naming both colliding entries (fail-
fast, consistent with the AbCipHostAddress validation pattern).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`ToDriverDataType` mapped LInt/ULInt to Int32 (truncation) and UDInt
to Int32 (negative wrap for values > Int32.MaxValue). DriverDataType
already carries Int64/UInt64/UInt32, so map each Logix 64-bit and
unsigned-32-bit type to the correct member. `DecodeValueAt` in
`LibplctagTagRuntime` updated to return uint/ulong for UDInt/ULInt
so the runtime value type agrees with the declared OPC UA type.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replace the synchronous non-cancellable _stream.ReadByte() for the kind byte
in FrameReader.ReadFrameAsync with an async ReadExactAsync(new byte[1], ct)
call so the full frame read honours the EffectiveCallTimeout-linked token
and cannot wedge the call gate when the sidecar stalls mid-frame.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
WriteCommand.ParseValue wraps FormatException/OverflowException as
CliFx CommandException so a bad --value yields a clean one-line CLI error
naming the value and target type instead of a raw .NET stack trace.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Driver.AbCip.Cli-001: WriteCommand.ParseValue wraps FormatException/
OverflowException as CommandException so bad --value input yields a clean
CLI error instead of a raw stack trace.
Driver.AbCip.Cli-002: probe/read/subscribe commands reject Structure types
up front (RejectStructure helper), matching the write guard.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Added missing test coverage identified in the -005 finding:
- FormatTable_with_empty_input_returns_header_only: verifies the -004 fix
(empty batch read returns header+separator rather than throwing).
- FormatStatus_with_sub_code_bits_resolves_to_named_class: Theory exercising
the -002 high-word mask path (e.g. 0x80050001 → "BadCommunicationError").
- FormatStatus_unknown_sub_code_falls_back_to_severity_class: Theory for the
-002 severity-class fallback (unknown sub-codes still emit Good/Uncertain/Bad).
- New DriverCommandBaseTests class: four tests covering verbose/non-verbose
Serilog level selection, ConfigureLogging idempotency, and FlushLogging.
Also corrected the stale FormatStatus_unknown_codes_fall_back_to_hex_only
expectation (0xDEADBEEF now resolves to "Bad" via the severity-class fallback
introduced by -002, not bare hex) and fixed the FormatTable empty-input crash
(guard rows.Length == 0 before calling Enumerable.Max).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ConfigureLogging is now idempotent via a _loggingConfigured guard field so
repeated calls from subclasses do not abandon and leak the previous logger.
The previous Log.Logger is disposed before overwriting to release its
console-sink resources cleanly.
A new protected static FlushLogging() helper calls Log.CloseAndFlush() so
commands can guarantee buffered output is flushed in their finally blocks
before the process exits — important for the long-running subscribe verb.
XML doc updated to reflect call-once semantics and document FlushLogging().
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
FormatStatus now matches named codes against code & 0xFFFF0000 (high-word
mask) rather than exact equality, so status codes carrying sub-code or flag
bits in the low 16 bits (e.g. 0x80050001) still resolve to their named class.
For codes not in the named shortlist a severity-class fallback using the top
2 bits always emits Good / Uncertain / Bad rather than bare hex.
Updated the stale FormatStatus_unknown_codes_fall_back_to_hex_only test (its
expectation became invalid once the severity-class fallback was added) and
added new Theory cases exercising both the high-word matching and the
severity-class fallback paths.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Core.VirtualTags-002: cold-start guard publishes BadWaitingForInitialData
instead of silently returning a stale value.
Core.VirtualTags-003: Load detects duplicate Path values and keys the
upstream-subscription loop off the registered tag set.
Core.VirtualTags-005: VirtualTagSource fires the initial-data callback per
path before registering the change observer, fixing an ordering race.
Core.VirtualTags-008: DependencyGraph caches topological rank, lowering
per-change-event cost from O(V+E) to O(closure).
Core.VirtualTags-012: added 9 engine tests; CoerceResult null-return now
maps to BadInternalError as the code comment intended.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Core.Abstractions-001: PollGroupEngine compares array values with structural
equality so a driver returning a fresh T[] each poll no longer fires spuriously.
Core.Abstractions-002: PollOnceAsync guards reader result cardinality and
throws a descriptive InvalidOperationException on mismatch instead of a
swallowed ArgumentOutOfRangeException that stalled the subscription.
Core.Abstractions-003: the poll loop Task is tracked; Unsubscribe/DisposeAsync
await loop completion before disposing the CTS.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Mark Core.ScriptedAlarms-002, -004, -005, -007, -012 as Resolved with
one-line descriptions. Update open-findings count from 11 to 6.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add engine-level tests covering the six gaps identified in the finding:
(1) timed-shelve auto-expiry driven via injectable clock + RunShelvingCheckForTest
hook so timer tests are deterministic;
(2) ConfirmAsync, TimedShelveAsync/UnshelveAsync round-trip, EnableAsync engine
methods exercised end-to-end;
(3) OnEvent subscriber-throws isolation — engine state advances and stays
operational after a subscriber throws;
(4) IAlarmStateStore.SaveAsync failure leaves in-memory state unchanged (locks in
the persist-before-update invariant from finding-007);
(5) second LoadAsync does not leak the old timer (regression for finding-002);
(6) AreInputsReady cold-start guard correctly blocks on Bad/missing inputs and
allows Uncertain-quality inputs through.
Expose RunShelvingCheckForTest() internal method on ScriptedAlarmEngine to
support deterministic timer tests.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
SubscribeAsync now wraps each driver handle in a private HostBoundHandle
that carries the resolved host name. UnsubscribeAsync unwraps it and
routes through the recorded host's resilience pipeline, correctly
charging the subscription's originating host's circuit breaker/bulkhead
instead of always using the default host. Falls back to the default
host for handles not created by this invoker. Two regression tests
added; update findings.md Open count from 10 to 6.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
BuildAddressSpaceAsync now checks _disposed (throws ObjectDisposedException)
and tears down the previous alarm forwarder + clears the sink registry
before re-walking, so a Galaxy-redeploy rebuild does not leak the old
forwarder and double-deliver alarm transitions. Three regression tests
added: double-build does not double-fire, sink count is correct after
rebuild, and post-dispose call throws.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Change ClusterEntry from sealed record to sealed class so TryUpdate
uses reference equality for the CAS comparison. Prune now uses a
read-compute-TryUpdate retry loop that restarts when a concurrent
Install updates the entry between the read and the write, preventing
a race that could silently drop the just-installed newest generation.
Two regression tests added to PermissionTrieCacheTests.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add FolderSegment member to NodeAclScopeKind; update WalkSystemPlatform
to report NodeAclScopeKind.FolderSegment (not Equipment) for each
visited Galaxy folder level, so MatchedGrant.Scope in
AuthorizationDecision.Provenance correctly distinguishes Galaxy folder
grants from UNS Equipment grants in the audit trail and Admin UI
diagnostics. Three regression tests added to PermissionTrieTests.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Reorder persist/update in ApplyAsync, ReevaluateAsync, and ShelvingCheckAsync:
SaveAsync is now called before the in-memory _alarms entry is advanced. A store
failure therefore leaves both the persisted and in-memory views at the prior state
rather than diverging, maintaining the invariant that startup recovery reflects
actual persisted state.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add _disposed re-checks inside ReevaluateAsync and ShelvingCheckAsync after
acquiring _evalGate so callbacks in flight when Dispose() runs bail out cleanly
instead of mutating _alarms or writing to a disposed store. Drop the
_alarms.Clear() from Dispose() — clearing outside the gate races concurrent
reads and is unnecessary since the object is being discarded.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Split the LoadAsync seed-read + subscribe loop: ReadTag seed fills _valueCache
first, then persisted-state restore runs, then _loaded = true, then SubscribeTag
is called. Any synchronous initial push from the upstream now arrives after
_alarms is fully initialised and _loaded = true, so ReevaluateAsync will queue
correctly behind the gate rather than racing the half-built state.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Dispose any existing _shelvingTimer before reassigning it inside LoadAsync so
that a second LoadAsync call does not leak the old timer and leave two timers
running concurrently against the same engine state.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
