Closes the gap surfaced by Phase 7 live smoke (#240): DriverInstance rows in
the central config DB had no path to materialise as live IDriver instances in
DriverHost, so virtual-tag scripts read BadNodeIdUnknown for every tag.
## DriverFactoryRegistry (Core.Hosting)
Process-singleton type-name → factory map. Each driver project's static
Register call pre-loads its factory at Program.cs startup; the bootstrapper
looks up by DriverInstance.DriverType + invokes with (DriverInstanceId,
DriverConfig JSON). Case-insensitive; duplicate-type registration throws.
## GalaxyProxyDriverFactoryExtensions.Register (Driver.Galaxy.Proxy)
Static helper — no Microsoft.Extensions.DependencyInjection dep, keeps the
driver project free of DI machinery. Parses DriverConfig JSON for PipeName +
SharedSecret + ConnectTimeoutMs. DriverInstanceId from the row wins over JSON
per the schema's UX_DriverInstance_Generation_LogicalId.
## DriverInstanceBootstrapper (Server)
After NodeBootstrap loads the published generation: queries DriverInstance
rows scoped to that generation, looks up the factory per row, constructs +
DriverHost.RegisterAsync (which calls InitializeAsync). Per plan decision
#12 (driver isolation), failure of one driver doesn't prevent others —
logs ERR + continues + returns the count actually registered. Unknown
DriverType (factory not registered) logs WRN + skips so a missing-assembly
deployment doesn't take down the whole server.
## Wired into OpcUaServerService.ExecuteAsync
After NodeBootstrap.LoadCurrentGenerationAsync, before
PopulateEquipmentContentAsync + Phase7Composer.PrepareAsync. The Phase 7
chain now sees a populated DriverHost so CachedTagUpstreamSource has an
upstream feed.
## Live evidence on the dev box
Re-ran the Phase 7 smoke from task #240. Pre-#248 vs post-#248:
Equipment namespace snapshots loaded for 0/0 driver(s) ← before
Equipment namespace snapshots loaded for 1/1 driver(s) ← after
Galaxy.Host pipe ACL denied our SID (env-config issue documented in
docs/ServiceHosting.md, NOT a code issue) — the bootstrapper logged it as
"failed to initialize, driver state will reflect Faulted" and continued past
the failure exactly per plan #12. The rest of the pipeline (Equipment walker
+ Phase 7 composer) ran to completion.
## Tests — 5 new DriverFactoryRegistryTests
Register + TryGet round-trip, case-insensitive lookup, duplicate-type throws,
null-arg guards, RegisteredTypes snapshot. Pure functions; no DI/DB needed.
The bootstrapper's DB-query path is exercised by the live smoke (#240) which
operators run before each release.
Closes the historian leg of Phase 7. Scripted alarm transitions now batch-flow
through the existing Galaxy.Host pipe + queue durably in a local SQLite store-
and-forward when Galaxy is the registered driver, instead of being dropped into
NullAlarmHistorianSink.
## GalaxyHistorianWriter (Driver.Galaxy.Proxy.Ipc)
IAlarmHistorianWriter implementation. Translates AlarmHistorianEvent →
HistorianAlarmEventDto (Stream D contract), batches via the existing
GalaxyIpcClient.CallAsync round-trip on MessageKind.HistorianAlarmEventRequest /
Response, maps per-event HistorianAlarmEventOutcomeDto bytes back to
HistorianWriteOutcome (Ack/RetryPlease/PermanentFail) so the SQLite drain
worker knows what to ack vs dead-letter vs retry. Empty-batch fast path.
Pipe-level transport faults (broken pipe, host crash) bubble up as
GalaxyIpcException which the SQLite sink's drain worker translates to
whole-batch RetryPlease per its catch contract.
## GalaxyProxyDriver implements IAlarmHistorianWriter
Marker interface lets Phase7Composer discover it via type check at compose
time. WriteBatchAsync delegates to a thin GalaxyHistorianWriter wrapping the
driver's existing _client. Throws InvalidOperationException if InitializeAsync
hasn't connected yet — the SQLite drain worker treats that as a transient
batch failure and retries.
## Phase7Composer.ResolveHistorianSink
Replaces the injected sink dep when any registered driver implements
IAlarmHistorianWriter. Constructs SqliteStoreAndForwardSink at
%ProgramData%/OtOpcUa/alarm-historian-queue.db (falls back to %TEMP% when
ProgramData unavailable, e.g. dev), starts the 2s drain timer, owns the sink
disposable for clean teardown. When no driver provides the writer, keeps the
NullAlarmHistorianSink wired by Program.cs (#246).
DisposeAsync now also disposes the owned SQLite sink in the right order:
bridge → engines → owned sink → injected fallback.
## Tests — 7 new GalaxyHistorianWriterMappingTests
ToDto round-trips every field; preserves null Comment; per-byte outcome enum
mapping (Ack / RetryPlease / PermanentFail) via [Theory]; unknown byte throws;
ctor null-guard. The IPC round-trip itself is covered by the live Host suite
(task #240) which constructs a real pipe.
Server.Phase7 tests: 34/34 still pass; Galaxy.Proxy tests: 25/25 (+7 = 32 total).
## Phase 7 production wiring chain — COMPLETE
- ✅#243 composition kernel
- ✅#245 scripted-alarm IReadable adapter
- ✅#244 driver bridge
- ✅#246 Program.cs wire-in
- ✅#247 this — Galaxy.Host historian writer + SQLite sink activation
What unblocks now: task #240 live OPC UA E2E smoke. With a Galaxy driver
registered, scripted alarm transitions flow end-to-end through the engine →
SQLite queue → drain worker → Galaxy.Host IPC → Aveva Historian alarm schema.
Without Galaxy, NullSink keeps the engines functional and the queue dormant.
