# Plan — alarms over the mxaccessgw gateway

Coordinated epic across two repos:

- **`lmxopcua`** (this repo) — `c:\Users\dohertj2\Desktop\lmxopcua\`
- **`mxaccessgw`** — `c:\Users\dohertj2\Desktop\mxaccessgw\`

## Why

PR 7.2 (2026-04-30, commit `ae7106d`) retired the in-process v1 Galaxy stack
(`Driver.Galaxy.Host` / `.Proxy` / `.Shared` + `OtOpcUaGalaxyHost` Windows
service) and migrated Galaxy access to the in-process `GalaxyDriver` over
mxaccessgw's gRPC. In doing so, three v1 capabilities regressed:

1. **Native MxAccess alarm-event metadata** — v1's `GalaxyAlarmTracker`
   surfaced rich alarm transitions (operator comment, original raise time,
   ack time, alarm category, native severity). The current architecture
   reconstructs Part 9 transitions by subscribing to four sub-attribute
   value updates (`InAlarm`, `Acked`, `Priority`, `Description`) — fine for
   raise/clear but loses everything else.
2. **Native MxAccess Acknowledge semantics** — v1 called the MxAccess ack
   API directly from `GalaxyAlarmTracker`. Today, OPC UA acks are written
   into the `AckMsgWriteRef` sub-attribute — semantically valid but a
   round-trip through the value path that loses operator-comment fidelity.
3. **Alarm-historian write-back path** — `GalaxyHistorianWriter`
   implemented `IAlarmHistorianWriter` and forwarded scripted-alarm and
   Galaxy-native alarm transitions back to AVEVA Historian via
   `aahClientManaged`. PR 7.2 deleted it. `Phase7Composer.ResolveHistorianSink`
   now finds no writer and falls back to `NullAlarmHistorianSink`, so
   **scripted-alarm transitions queue locally and silently discard.**
   (Galaxy-native alarms still reach AVEVA Historian via the Galaxy template's
   own `HistorizeToAveva` toggle, independent of our sink — that path
   wasn't broken.)

`gateway.md` (mxaccessgw, line 8) explicitly commits the gateway to "full
MXAccess parity… preserve MXAccess behavior first… **native MXAccess event
families**." Today's gateway proto exposes only data-change families. Closing
the alarm regression and fulfilling that parity statement are the same task.

## Goals

- Restore all three regressed capabilities to feature parity with v1.
- Keep the v2 architectural split — gateway owns MxAccess transport;
  lmxopcua owns OPC UA Part 9 semantics, ACL/role enforcement, and
  multi-source aggregation (driver-native + scripted + sub-attribute).
- Preserve the value-driven sub-attribute path as a fallback for Galaxy
  templates that don't carry `$Alarm*` extensions.
- Land the work as a sequence of small, independently-reviewable PRs that
  alternate between repos in dependency order.

## Non-goals

- Reimplementing the Part 9 state machine inside mxaccessgw. The gateway
  stays UA-agnostic.
- Reworking the LDAP role-grant or OPC UA AlarmAck ACL surface — those
  already exist and route through `Server/Alarms/IAlarmAcknowledger`.
- Adding alarm support to non-Galaxy drivers (AbCip / FOCAS / OpcUaClient
  already have their own `IAlarmSource` implementations; Modbus / S7 /
  AbLegacy / TwinCAT don't have a native alarm bus and are out of scope).
- Altering Galaxy template conventions or `$Alarm*` extensions in the
  customer's Galaxy.

## Before → after

**Today (post-PR 7.2):**

```
MxAccess COM (gateway worker)
   │ data-change events only on the MxEvent stream
   ▼
GalaxyDriver (no IAlarmSource)
   │ IWritable / ISubscribable / ITagDiscovery only
   ▼
DriverNodeManager
   ├─ subscribes to four $Alarm* sub-attributes per condition
   ├─ AlarmConditionService rebuilds Part 9 transitions from value updates
   └─ DriverWritableAcknowledger writes AckMsgWriteRef on ack

Phase7Composer.ResolveHistorianSink → NullAlarmHistorianSink
   (scripted-alarm transitions queue → silently discarded)
```

**After this epic:**

```
MxAccess COM (gateway worker)
   │ data-change   ──┐
   │ alarm-transition │
   │ write-complete   ├─► single MxEvent stream (new family added)
   ▼                  ▼
GalaxyDriver : ITagDiscovery, IReadable, IWritable, ISubscribable, IRediscoverable,
               IHostConnectivityProbe, IAlarmSource             ← restored
   ├─ EventPump dispatches OnAlarmTransition family → IAlarmSource.OnAlarmEvent
   ├─ AcknowledgeAsync → gateway RPC AcknowledgeAlarm
   └─ QueryActiveAlarmsAsync → gateway RPC QueryActiveAlarms (ConditionRefresh)

DriverNodeManager
   ├─ rich alarm events from IAlarmSource.OnAlarmEvent → AlarmConditionService
   ├─ value-driven sub-attribute path STILL WORKS for templates without $Alarm
   ├─ DriverWritableAcknowledger preserved as fallback for the value path
   └─ ScriptedAlarmEngine output continues to feed AlarmConditionService

Phase7Composer.ResolveHistorianSink → GatewayAlarmHistorianWriter
   ├─ scripted-alarm transitions → SqliteStoreAndForwardSink
   └─ drain worker → gateway RPC WriteHistorianEvent → AVEVA Historian
```

## Architecture decisions

**D1 — Where the Part 9 state machine runs.** Stays in lmxopcua's
`AlarmConditionService`. Gateway is UA-agnostic. ScriptedAlarmEngine produces
Part 9 transitions with no MxAccess origin; the aggregator must live where all
sources converge.

**D2 — Where authz on Acknowledge runs.** Stays in lmxopcua. The OPC UA
`AlarmConditionState.OnAcknowledge` delegate already checks the session's
roles for `AlarmAck` against the LDAP/role-grant ACL. The gateway should
never be reachable in a way that bypasses that check.

**D3 — How rich alarm events reach OPC UA clients.** New `MxEventFamily`
on the existing `StreamEvents` RPC (no second stream). Adds latency
parity with data-change events, reuses the bounded-channel + worker-side
delivery semantics already documented in `gateway.md`.

**D4 — Sub-attribute fallback path stays.** Some Galaxy templates won't
have `$Alarm*` extensions yet; the existing value-driven path remains the
only way to surface alarms for those templates. Both paths feed
`AlarmConditionService`. Driver-native events take precedence when both
are present (more authoritative, lower latency).

**D5 — Where the historian writer lives.** As a new RPC on the gateway
(`WriteHistorianEvent`). The Wonderware sidecar's existing
`WriteAlarmEvents` IPC slot stays unwired and is deleted as part of this
epic — the gateway is the canonical place for "write to AVEVA Historian"
since the gateway already owns AVEVA-COM access. This also means the
sidecar (long term) only does *reads* and could potentially retire entirely
if the historian-client REST migration (`docs/plans/...`) lands.

## Track A — mxaccessgw changes

All five PRs land in `c:\Users\dohertj2\Desktop\mxaccessgw\`.

### PR A.1 — proto: add alarm-transition event family + ack/query RPCs

**Files** (`src\MxGateway.Contracts\Protos\mxaccess_gateway.proto`):

1. Extend `MxEventFamily` (line 403):
   ```
   MX_EVENT_FAMILY_ON_ALARM_TRANSITION = 5;
   ```

2. Extend `MxEvent.body` oneof (line 395) with:
   ```
   OnAlarmTransitionEvent on_alarm_transition = 24;
   ```

3. New message `OnAlarmTransitionEvent` after the existing event-family
   bodies (line 425+). Carry the full MxAccess alarm payload — alarm name,
   source object reference, alarm-type-name (e.g. "AnalogLimitAlarm.HiHi"),
   transition kind enum (`Raise` / `Acknowledge` / `Clear`), severity (raw
   numeric — keep MxAccess scale; mapping to OPC UA 0-1000 happens
   server-side in lmxopcua), `original_raise_timestamp`,
   `transition_timestamp`, optional `operator_user`, optional
   `operator_comment`, alarm `category` string, alarm `description`. Mirror
   the field set documented in v1's `GalaxyAlarmTracker`.

4. New RPC on `MxAccessGateway` service (line 11):
   ```
   rpc AcknowledgeAlarm(AcknowledgeAlarmRequest) returns (AcknowledgeAlarmReply);
   rpc QueryActiveAlarms(QueryActiveAlarmsRequest) returns (stream ActiveAlarmSnapshot);
   ```

   `AcknowledgeAlarmRequest` carries `session_id`, `alarm_full_reference`,
   `comment`, `user_principal`. Reply carries `MxStatusProxy`.

   `QueryActiveAlarmsRequest` carries `session_id`, optional
   `alarm_filter_prefix` (for ConditionRefresh on a sub-tree).
   `ActiveAlarmSnapshot` carries the same fields as
   `OnAlarmTransitionEvent` plus `current_state` enum (`Active` /
   `ActiveAcked` / `Inactive`).

**Tests** (`MxGateway.Tests` — proto/codegen sanity):

- Round-trip Serialize→Deserialize for the new messages with all-fields
  populated and empty-optional-fields cases.
- `MxEvent.body` oneof selection guard — supplying multiple bodies
  rejected.

**Out of scope:** worker-side wiring (PR A.2), gateway-side dispatch (PR A.3).
PR A.1 is a pure contract-surface change; nothing functional yet.

### PR A.2 — worker: subscribe to MxAccess alarm event source

**Files** (`src\MxGateway.Worker\` — net48/x86):

The MxAccess Toolkit exposes alarm subscription separately from data
subscription. Per AVEVA's MXAccess C++ Toolkit reference (canonical doc
referenced from `gateway.md`), alarm events arrive through the
`IAlarmEventSink` interface registered against the MxAccess `Alarms`
collection of an open session, OR via the MxAccess "alarm provider"
subscription pattern (depends on Toolkit version on the worker host —
verify against the version actually deployed in the worker bin during
PR A.2).

1. Worker subscribes to MxAccess alarms once per session, with a single
   sink that fans out into the same bounded channel the data-change pump
   uses (`MxGateway.Worker\Eventing\EventChannel.cs` or whatever the worker
   currently calls its sink — verify name during the PR).
2. Sink translates each MxAccess alarm event into a `WorkerEvent` proto
   (defined in `mxaccess_worker.proto`) carrying the new
   `OnAlarmTransitionEvent` body. Reuses the existing `worker_sequence`
   counter so ordering is preserved across families.
3. Worker honours the same backpressure rules as data-change events —
   newest-dropped on full channel, single dropped-counter metric per
   family.

**Tests** (`MxGateway.Worker.Tests`):

- Fake `IAlarmEventSink` source emits canned transitions; assert the
  worker forwards each as the right `WorkerEvent` shape.
- Cancellation test — closing the session unsubscribes from MxAccess
  alarms cleanly (no leaked sinks if the worker is recycled mid-session).

**Out of scope:** any gateway-side dispatch, any RPC handler — PR A.2
is worker-internal.

### PR A.3 — gateway: dispatch OnAlarmTransition + implement AcknowledgeAlarm

**Files** (`src\MxGateway.Server\`):

1. The session-level event multiplexer (`Sessions\SessionEventStream.cs`
   or equivalent — verify name during PR) recognizes the new
   `WorkerEvent` body and forwards as an `MxEvent` with family
   `MX_EVENT_FAMILY_ON_ALARM_TRANSITION` to the gRPC
   `StreamEvents` consumer.
2. New RPC handler `AcknowledgeAlarm` builds an MxAccess `WorkerCommand`
   carrying an `AlarmAcknowledgeCommand` (new in `mxaccess_worker.proto`
   under PR A.1). Forwarded to the worker; reply mapped to
   `AcknowledgeAlarmReply` with the MxAccess `MxStatus` proxy populated.
3. AuthN — same API-key + scope check as existing RPCs. Add a new scope
   `invoke:alarm-ack` (mirrors `invoke:write` granularity); existing keys
   without it return `PERMISSION_DENIED`.

**Tests** (`MxGateway.Tests`, `MxGateway.IntegrationTests`):

- Unit: dispatch test — fake worker emits an `AlarmTransition` event;
  assert the gateway forwards it on the live `StreamEvents` channel of
  every subscribed session.
- Integration: end-to-end against the real worker (requires the parity
  rig setup — see `docs\v2\Galaxy.ParityRig.md` in lmxopcua for the
  MxAccess-installed dev box prerequisites). Trigger a real Galaxy
  alarm, assert the gateway emits `OnAlarmTransition`. Acknowledge via
  the new RPC, assert the alarm transitions to `ActiveAcked` and an
  `Acknowledge` transition event is emitted back.
- AuthN: existing key without `invoke:alarm-ack` scope rejected.

### PR A.4 — gateway: ConditionRefresh snapshot via QueryActiveAlarms

**Files** (`src\MxGateway.Server\`, `src\MxGateway.Worker\`):

1. Worker exposes a `QueryActiveAlarmsCommand` that walks the session's
   active-alarm collection and streams snapshots back through the
   existing command-reply channel. The MxAccess Toolkit's
   `Alarms.GetActive()` (verify exact API name during PR) is the
   underlying call.
2. Gateway RPC `QueryActiveAlarms` opens a server-streaming reply,
   batches snapshots through.
3. AuthN — new scope `invoke:alarm-query` (separate from ack so a
   read-only client can refresh without ack rights).

**Tests:**

- Worker-test: synthetic active set of 0 / 1 / 100 alarms; assert
  pagination respects worker channel capacity.
- Integration: against the parity rig, assert a ConditionRefresh after
  reconnect returns every alarm currently `Active` or `ActiveAcked` in
  the Galaxy.

### PR A.5 — gateway: WriteHistorianEvent RPC for sink write-back

**Files** (`src\MxGateway.Server\`, `src\MxGateway.Worker\`,
`src\MxGateway.Contracts\Protos\mxaccess_gateway.proto`).

1. New RPC `WriteHistorianEvent(WriteHistorianEventRequest) →
   WriteHistorianEventReply`. Request carries an
   `AlarmHistorianRecord` mirroring the existing
   `Core.AlarmHistorian.AlarmHistorianEvent` payload (alarm id,
   equipment path, alarm name, alarm-type-name, severity, event kind,
   message, user, comment, timestamp).
2. Worker maps the record onto `aahClientManaged`'s alarm-event
   write API (the same path v1's `GalaxyHistorianWriter` used). Worker
   batches up to N records per write to amortize the COM round-trip.
3. AuthN — new scope `invoke:historian-write`. Cross-cutting with
   `invoke:write` — keys for OPC UA servers that publish historian
   data must hold both.

**Tests:**

- Worker test: fake `aahClientManaged` writer; assert batching
  semantics + retry-on-Bad-status-code behaviour matches v1's
  `GalaxyHistorianWriter` (per-row outcome reporting).
- Integration: write a record, query it back via existing Historian
  read APIs, assert round-trip fidelity.

**Sequencing within Track A:** A.1 → A.2 → A.3 → A.4 → A.5. A.1 is
mechanical; A.2 + A.3 are the load-bearing changes that unlock lmxopcua
side. A.4 + A.5 can ship after lmxopcua starts consuming A.3 output.

## Track B — lmxopcua changes

All five PRs land in `c:\Users\dohertj2\Desktop\lmxopcua\`. Each B-PR
depends on a specific A-PR — see the sequencing matrix below.

### PR B.1 — EventPump: dispatch OnAlarmTransition family

**Depends on:** A.1 (proto), A.3 (gateway dispatching the new family).

**Files:**

- `src\ZB.MOM.WW.OtOpcUa.Driver.Galaxy\Runtime\EventPump.cs:160` —
  current `Dispatch(MxEvent ev)` returns early for any non-`OnDataChange`
  family. Add a branch:
  ```csharp
  switch (ev.Family) {
    case MxEventFamily.OnDataChange: DispatchDataChange(ev); break;
    case MxEventFamily.OnAlarmTransition: DispatchAlarmTransition(ev); break;
    default: return;
  }
  ```
- New `DispatchAlarmTransition` translates the proto event into an
  `AlarmEventArgs` (existing type from `Core.Abstractions`) and raises an
  internal event the driver subscribes to.
- New `MxAccessSeverityMapper` in `Driver.Galaxy\Runtime\` — maps the
  MxAccess raw severity into the `AlarmSeverity` enum + the OPC UA
  numeric severity (250 / 500 / 700 / 900 ladder per v1's
  `AlarmTracking.md`).

**Tests** (`tests\ZB.MOM.WW.OtOpcUa.Driver.Galaxy.Tests\Runtime\`):

- `EventPumpAlarmTests` — feed three synthetic MxEvents (raise / ack /
  clear); assert each fires `OnAlarmEvent` on the driver with correct
  payload.
- Severity-mapping table tests — every documented MxAccess severity
  level → expected (`AlarmSeverity`, OPC UA numeric) tuple.

### PR B.2 — GalaxyDriver re-implements IAlarmSource

**Depends on:** A.3 (`AcknowledgeAlarm` RPC available), B.1 (event
dispatch).

**Files:**

- `src\ZB.MOM.WW.OtOpcUa.Driver.Galaxy\GalaxyDriver.cs:28` — extend the
  class declaration:
  ```csharp
  public sealed class GalaxyDriver
      : IDriver, ITagDiscovery, IReadable, IWritable, ISubscribable,
        IRediscoverable, IHostConnectivityProbe, IAlarmSource, IDisposable
  ```
- Implement the four `IAlarmSource` members:
  - `SubscribeAlarmsAsync` — no-op returning a sentinel handle. The
    driver is already subscribed for data; alarm events arrive on the
    same event stream once the gateway emits the new family. (Same
    pattern AbCip uses today — see `Driver.AbCip\AbCipDriver.cs:208`.)
  - `UnsubscribeAlarmsAsync` — no-op.
  - `OnAlarmEvent` — wired to the EventPump branch added in B.1.
  - `AcknowledgeAsync` — calls the new gateway RPC via the
    `IGalaxyAlarmAcknowledger` abstraction (new file, mirrors the
    `IGalaxyDataWriter` pattern), with `GatewayGalaxyAlarmAcknowledger`
    as the production implementation in `Runtime\`. Resilience wrapping
    via `AlarmSurfaceInvoker` per existing pattern.
- `DriverInstanceFactory` for Galaxy registers
  `IGalaxyAlarmAcknowledger` alongside the existing data writer.

**Tests:**

- Subscribe-noop returns a non-null handle; unsubscribe accepts it.
- Acknowledge — fake `IGalaxyAlarmAcknowledger` records the call; assert
  the request shape and resilience-pipeline routing.
- End-to-end test in `Driver.Galaxy.Tests` — fake gateway emits a
  raise-then-ack event sequence; assert the driver fires `OnAlarmEvent`
  twice with matching alarm-id correlation.

### PR B.3 — DriverNodeManager: route to driver-native when present

**Depends on:** B.2.

**Files:**

- `src\ZB.MOM.WW.OtOpcUa.Server\OpcUa\DriverNodeManager.cs` — when
  registering an `AlarmConditionState` for a Galaxy variable, check
  whether the driver is `IAlarmSource`. If yes, prefer the
  `OnAlarmEvent`-driven path; the value-driven sub-attribute path
  becomes the secondary path that handles transitions the driver-native
  stream missed (network blip, gateway restart, gw missing the
  `$Alarm*` extension on this template).
- `Server\Alarms\AlarmConditionService` — already accepts events from
  multiple sources; only addition is a `DriverEventOrigin` enum on
  internal transitions so the dedup logic prefers the richer
  driver-native record over a stale sub-attribute synthesis.
- `IAlarmAcknowledger` resolution in `DriverNodeManager` —
  prefer the driver's `IAlarmSource.AcknowledgeAsync` over
  `DriverWritableAcknowledger` when both are available. Keep
  `DriverWritableAcknowledger` as the fallback for templates without
  `$Alarm*` extensions.

**Tests:**

- Two-source-fan-in test: same alarm condition receives both a
  driver-native ack event and a sub-attribute value update for the same
  transition; assert no duplicate Part 9 transition fires.
- Acknowledger routing — driver implements `IAlarmSource` →
  ack-via-RPC; driver implements only `IWritable` → ack-via-write
  (existing path).

### PR B.4 — IAlarmHistorianWriter via gateway

**Depends on:** A.5 (`WriteHistorianEvent` RPC available).

**Files:**

- New `src\ZB.MOM.WW.OtOpcUa.Driver.Galaxy\Runtime\GatewayAlarmHistorianWriter.cs`
  implementing `IAlarmHistorianWriter`. Calls the gateway RPC from
  Track A.5 with the same batch + per-row outcome semantics v1's
  `GalaxyHistorianWriter` exposed.
- `GalaxyDriverFactory` registers it as a singleton tied to the
  `DriverInstance`.
- `Server\Phase7\Phase7Composer.ResolveHistorianSink` — already scans
  registered drivers for an `IAlarmHistorianWriter`. Once GalaxyDriver
  exposes one, `SqliteStoreAndForwardSink` boots with a real writer
  attached and the `NullAlarmHistorianSink` fallback no longer applies
  on Galaxy installs.
- Delete `WriteAlarmEventsRequest` / `WriteAlarmEventsReply` /
  `IAlarmEventWriter` from the Wonderware sidecar
  (`src\ZB.MOM.WW.OtOpcUa.Driver.Historian.Wonderware\Ipc\Contracts.cs`,
  `Ipc\HistorianFrameHandler.cs`, `Ipc\Framing.cs`). The historian
  sidecar becomes read-only — matches the audit done earlier.

**Tests:**

- `GatewayAlarmHistorianWriter` against a fake gRPC server — single
  record, batch, per-row failure modes (Ack / RetryPlease /
  PermanentFail).
- `Phase7Composer` end-to-end — register a Galaxy driver, assert
  `ResolveHistorianSink` picks `SqliteStoreAndForwardSink` with the
  new writer attached.

### PR B.5 — docs + memory housekeeping

**Depends on:** B.1 / B.2 / B.3 / B.4 all green on the parity rig.

**Files:**

- `docs\drivers\Galaxy.md` — current text says the driver implements
  five capability interfaces; update to seven (`IAlarmSource`,
  `IAlarmHistorianWriter`-via-companion).
- `docs\AlarmTracking.md` — promote a fresh top-level doc that
  describes the v2-final architecture (driver-native primary path +
  sub-attribute fallback + scripted-alarm aggregation). Cross-link from
  `docs\README.md`. The v1 archive stays as historical record.
- `docs\v1\AlarmTracking.md` — extend the existing historical banner
  with "Restored to functional parity in this epic — see
  `docs\AlarmTracking.md` for current state."
- Memory entries (`C:\Users\dohertj2\.claude\projects\…\memory\`):
  - Update `project_galaxy_via_mxgateway.md` — add the alarm path
    restoration.
  - Update `project_server_history_alarm_subsystems.md` — note that
    `Phase7Composer.ResolveHistorianSink` now finds a writer on
    Galaxy installs.
- `docs\plans\alarms-over-gateway.md` (this file) — banner the doc
  `✅ Completed YYYY-MM-DD — historical record.` matching the existing
  v2-mxgw plan retirement convention.

## Sequencing matrix

```
Track A (mxaccessgw)              Track B (lmxopcua)
─────────────────────────         ─────────────────────────
A.1 proto                         (waits)
   │
   ├──────────────────────────►   B.1 EventPump branch
A.2 worker subscription                  │ uses proto types only
   │                                     │ unit-testable without live gw
   │
A.3 gateway dispatch + ack RPC ──►B.2 GalaxyDriver : IAlarmSource
   │                                     │
   │                              ──►B.3 DriverNodeManager routing
   │
A.4 ConditionRefresh                     │ (B.3 closes the loop with A.4
   │                                       once ConditionRefresh wired)
   │
A.5 WriteHistorianEvent ─────────►B.4 GatewayAlarmHistorianWriter
                                        │ + sidecar write-path deletion
                                  ──►B.5 docs + memory
```

A.1 + B.1 can land in parallel (B.1's tests use proto types without
needing a running gateway). B.1 stays inert until A.3 ships the gateway
dispatch — which is fine; the dispatch branch is a no-op until events
arrive.

## Test gates

Per PR: unit tests pass + build green + analyzer clean (Roslyn
OTOPCUA0001 still wraps every alarm-capability call through
`AlarmSurfaceInvoker`).

End-of-epic gate: re-run the parity rig (`docs\v2\Galaxy.ParityRig.md`)
with these scenarios added:

1. **Native alarm raise** — Galaxy `$Alarm*` raise with operator-time
   metadata appears as an OPC UA Part 9 transition with full payload
   (no longer reconstructed from sub-attribute writes).
2. **Native ack** — OPC UA client acks; assert the gateway records the
   ack against MxAccess directly (not via sub-attribute write); operator
   comment present in the resulting `Acknowledged` transition.
3. **ConditionRefresh after reconnect** — disconnect the GalaxyDriver,
   raise three alarms in Galaxy, reconnect; assert all three appear in
   the next ConditionRefresh.
4. **Historian write-back** — fire a scripted alarm; assert it arrives in
   AVEVA Historian via the gateway path (use the existing Historian
   sidecar's read API to query it back).
5. **Sub-attribute fallback still works** — disable `IAlarmSource` on
   the GalaxyDriver via test seam, fire a sub-attribute value change;
   assert Part 9 transition still raised.

Soak target: 24h × 1k tags (light) — same parity-rig harness but
extended to also subscribe to alarms. Pass criterion: zero dropped
alarm transitions, zero state-machine inversions, zero unhandled
exceptions in the AlarmSurfaceInvoker pipeline.

## Risks and mitigations

| Risk | Mitigation |
|---|---|
| MxAccess Toolkit alarm subscription API differs across installed AVEVA versions | PR A.2 verifies against the worker-host's installed Toolkit version; documents the exact API used. Pin the worker DLL set per major MxAccess version if needed. |
| Worker-side alarm subscription leaks between sessions if cleanup is wrong | PR A.2 includes a session-recycle test that asserts no `IAlarmEventSink` instances remain registered after Close. |
| Gateway adds a new auth scope (`invoke:alarm-ack`); existing keys lack it | PR A.3 + A.5 ship with a one-time bootstrap migration: keys with `invoke:write` get the new scope auto-granted on the dev rig and parity rig. Production keys are reissued via `apikey rotate-key` (existing CLI). |
| Two simultaneous alarm sources (driver-native + sub-attribute) double-fire transitions | PR B.3 dedup is the load-bearing design. End-to-end test #1 covers it explicitly. |
| Historian write-back batch fails mid-batch — partial success | The existing `SqliteStoreAndForwardSink.HistorianWriteOutcome` per-row enum + dead-letter retention already handles this; PR A.5 just exposes the same outcome shape over gRPC. |
| Sidecar write-path deletion in B.4 leaves orphan IPC frames in old client builds | The frame-kind enum is forward-compatible (`MessageKind.WriteAlarmEventsRequest = 0x20`). Old clients sending the request to a new sidecar receive `Unsupported message kind`; new clients never send it. Acceptable — same-version deploy is the existing rollout convention. |

## Roll-out

Track A lands first onto `mxaccessgw/main`, deployed to the parity rig.
Track B lands onto `lmxopcua/master` once A.3 is live on the rig — earlier
Track B PRs can target a feature branch (`feat/alarms-over-gateway`) and
merge to master after the rig is fully green.

## Back-out

Each PR is individually revertable. The cleanest back-out point is at
the gateway-side enum extension: removing `MX_EVENT_FAMILY_ON_ALARM_TRANSITION`
from the proto means EventPump silently drops alarm events again and
GalaxyDriver's `OnAlarmEvent` never fires — but the sub-attribute fallback
path still produces functional alarms, so the OPC UA surface degrades to
v2-current behaviour without breaking. PR B.4 is the only one with a
non-trivial back-out (re-add the deleted sidecar IPC slot if revert
needed); land B.4 last and only after end-of-epic gate is green.

## Out of scope (explicit)

- **Other alarm sources beyond Galaxy.** AbCip / FOCAS / OpcUaClient
  drivers already implement `IAlarmSource`; they're untouched.
- **Modbus / S7 / AbLegacy / TwinCAT alarms.** None of those protocols
  has a native alarm bus. Alarms on those drivers, if needed, ship via
  the scripted-alarm path.
- **Multi-Galaxy ack routing.** Today's gateway model is one Galaxy per
  session; if a deployment splits across galaxies, each gets its own
  GalaxyDriver and they don't cross-talk. No change.
- **OPC UA Part 9 advanced features** beyond the current scope —
  shelving, subscribed-to-events-only, branch-state for re-trigger
  semantics. Future epic if a customer asks.
- **Insight / cloud Historian write-back path.** Track A.5 targets the
  on-prem AVEVA Historian via aahClientManaged. The cloud variant
  would mirror the same gateway RPC over the REST API discussed in
  `docs/histsdk` — separate epic.

## File inventory (touched)

**mxaccessgw:**

- `src\MxGateway.Contracts\Protos\mxaccess_gateway.proto` (A.1, A.5)
- `src\MxGateway.Contracts\Protos\mxaccess_worker.proto` (A.2, A.4, A.5)
- `src\MxGateway.Worker\…\Eventing\` (A.2, A.3, A.4)
- `src\MxGateway.Worker\…\Commands\` (A.3, A.4, A.5)
- `src\MxGateway.Server\Sessions\SessionEventStream.cs` (A.3)
- `src\MxGateway.Server\Rpc\` (A.3, A.4, A.5)
- `src\MxGateway.Server\Auth\Scopes.cs` (A.3, A.4, A.5)
- `MxGateway.Tests`, `MxGateway.Worker.Tests`, `MxGateway.IntegrationTests`

**lmxopcua:**

- `src\ZB.MOM.WW.OtOpcUa.Driver.Galaxy\Runtime\EventPump.cs` (B.1)
- `src\ZB.MOM.WW.OtOpcUa.Driver.Galaxy\Runtime\MxAccessSeverityMapper.cs` *(new — B.1)*
- `src\ZB.MOM.WW.OtOpcUa.Driver.Galaxy\Runtime\IGalaxyAlarmAcknowledger.cs` *(new — B.2)*
- `src\ZB.MOM.WW.OtOpcUa.Driver.Galaxy\Runtime\GatewayGalaxyAlarmAcknowledger.cs` *(new — B.2)*
- `src\ZB.MOM.WW.OtOpcUa.Driver.Galaxy\Runtime\GatewayAlarmHistorianWriter.cs` *(new — B.4)*
- `src\ZB.MOM.WW.OtOpcUa.Driver.Galaxy\GalaxyDriver.cs` (B.2)
- `src\ZB.MOM.WW.OtOpcUa.Driver.Galaxy\GalaxyDriverFactory.cs` (B.2, B.4)
- `src\ZB.MOM.WW.OtOpcUa.Server\OpcUa\DriverNodeManager.cs` (B.3)
- `src\ZB.MOM.WW.OtOpcUa.Server\Alarms\AlarmConditionService.cs` (B.3)
- `src\ZB.MOM.WW.OtOpcUa.Server\Phase7\Phase7Composer.cs` (B.4)
- `src\ZB.MOM.WW.OtOpcUa.Driver.Historian.Wonderware\Ipc\Contracts.cs` (B.4 — deletions)
- `src\ZB.MOM.WW.OtOpcUa.Driver.Historian.Wonderware\Ipc\HistorianFrameHandler.cs` (B.4 — deletions)
- `src\ZB.MOM.WW.OtOpcUa.Driver.Historian.Wonderware\Ipc\Framing.cs` (B.4 — deletions)
- `tests\ZB.MOM.WW.OtOpcUa.Driver.Galaxy.Tests\Runtime\` (B.1, B.2)
- `tests\ZB.MOM.WW.OtOpcUa.Server.Tests\Alarms\` (B.3)
- `tests\ZB.MOM.WW.OtOpcUa.Driver.Historian.Wonderware.Tests\` (B.4 — drop deleted-contract tests)
- `docs\drivers\Galaxy.md` (B.5)
- `docs\AlarmTracking.md` *(new — B.5)*
- `docs\v1\AlarmTracking.md` (B.5 — banner update)
- `docs\plans\alarms-over-gateway.md` (B.5 — completion banner)

Total: ~12 source files added/modified in mxaccessgw; ~17 in lmxopcua;
~10 test files. Should land in 4-6 weeks of focused work given the
parity-rig dependency for end-to-end validation.