test(mtconnect): live-Agent docker fixture + integration suite (Tasks 19+20)

The first and only thing in the MTConnect workstream that exercises the driver
against a real Agent; everything else runs on canned XML.

Fixture (Docker/): two services, both stock images with this folder bind-mounted.
- `agent`   mtconnect/agent:2.7.0.12 on :5000. NOTE the repo name: the source
            project is mtconnect/cppagent but the PUBLISHED image is
            mtconnect/agent; mtconnect/cppagent does not exist on Docker Hub.
- `adapter` a stdlib SHDR feeder. Required, not optional: the agent image ships
            only the binary + schemas/styles, so a lone Agent answers /probe and
            then reports every observation UNAVAILABLE forever, which can prove
            nothing about reads or streaming.

Devices.xml seeds one DataItem per inference branch (SAMPLE+units, TIME_SERIES,
PART_COUNT/LINE_NUMBER, controlled vocab, DATA_SET, CONDITION), gives every named
item `name != id` (the inverse of the hand-authored unit fixtures), and leaves one
EVENT permanently unfed so UNAVAILABLE -> BadNoCommunication has a live subject.

Suite (12 tests) asserts STRUCTURALLY against the Agent's own /probe response, never
by literal id, so it survives an edit to Devices.xml and can be pointed at a real
machine tool via MTCONNECT_AGENT_ENDPOINT. It skips cleanly (12/12) when no Agent
answers, and each test carries a hard [Fact(Timeout)] so a half-up fixture cannot
wedge a build.

Three findings from bringing the fixture up, each now pinned in a comment:
- agent.cfg must be pure ASCII; one non-ASCII byte in a COMMENT makes the config
  parser reject the whole file with a bare "Failed / Stopped at line: N" and exit.
- `sampleCount` is an attribute of a TIME_SERIES observation, not of a DataItem
  declaration. A real Agent meeting one on a declaration DROPS THE ENTIRE DATA ITEM
  from the device model, so the inference only ever sees null and a live TIME_SERIES
  tag is always variable-length. Asserted, not ignored.
- The Agent's own <Agent> self-model publishes update-rate SAMPLEs that tick whether
  or not any adapter is attached. Excluding them is load-bearing: with them included
  the "live stream delivered a changed value" test passed with the fixture's data
  source deliberately stopped.

MTConnectError-under-HTTP-200 is NOT covered live and cannot be: cppagent 2.7 answers
an unknown device 404 and an out-of-range sequence 400, each with a well-formed error
body. That shape stays canned-only, and the suite says so.

infra/README.md

@@ -37,7 +37,7 @@ ssh dohertj2@10.100.0.35 'docker ps --filter label=project=lmxopcua --format "{{
   | Stack dir | Purpose |
   |---|---|
   | `/opt/otopcua-mssql` | central SQL (always-on) |
-  | `/opt/otopcua-modbus` · `/opt/otopcua-abcip` · `/opt/otopcua-s7` · `/opt/otopcua-opcuaclient` · `/opt/otopcua-mqtt` | driver fixtures |
+  | `/opt/otopcua-modbus` · `/opt/otopcua-abcip` · `/opt/otopcua-s7` · `/opt/otopcua-opcuaclient` · `/opt/otopcua-mqtt` · `/opt/otopcua-mtconnect` | driver fixtures |
   | `~/otopcua-ablegacy` · `~/otopcua-focas` | driver fixtures (user-owned) |
   | `~/otopcua-harness` | Host.IntegrationTests real-mode deps (SQL + GLAuth) — see §4 |
 
@@ -71,10 +71,11 @@ when it isn't running. Bring one up, `dotnet test`, tear down. Repo compose live
 | FOCAS | `otopcua-focas-sim` (vendored focas-mock) | `10.100.0.35:8193` | `docker compose up -d` (stack `~/otopcua-focas`) |
 | MQTT | `eclipse-mosquitto:2.0.22` (+ a publisher sidecar) | `10.100.0.35:8883` TLS+auth · `:1883` plaintext+auth | `MQTT_FIXTURE_USERNAME=otopcua MQTT_FIXTURE_PASSWORD=<pw> docker compose up -d` (stack `/opt/otopcua-mqtt`) |
 | MQTT — Sparkplug B | `otopcua-sparkplug-sim` (project-owned C# edge-node simulator) | same broker; group **`OtOpcUaSim`**, edge nodes **`EdgeA`** / **`EdgeB`**, `EdgeA` device **`Filler1`**, ~2 s cadence | `docker compose --profile sparkplug up -d` (same stack). Answers rebirth NCMDs; restart it to force a fresh birth |
+| MTConnect | `mtconnect/agent:2.7.0.12` + `python:3.13-alpine` adapter | `http://10.100.0.35:5000` | `docker compose up -d --wait` |
 
 **Endpoint overrides** point a suite at a real PLC instead of the sim:
 `MODBUS_SIM_ENDPOINT` · `AB_SERVER_ENDPOINT` · `S7_SIM_ENDPOINT` · `OPCUA_SIM_ENDPOINT` ·
-`MQTT_FIXTURE_ENDPOINT` / `MQTT_FIXTURE_PLAIN_ENDPOINT`.
+`MQTT_FIXTURE_ENDPOINT` / `MQTT_FIXTURE_PLAIN_ENDPOINT` · `MTCONNECT_AGENT_ENDPOINT`.
 
 > **MQTT needs credentials + material, unlike the other fixtures.** The broker has **no anonymous
 > fallback on either listener** (deliberate — an anonymous broker would let a bad auth config pass),
@@ -84,6 +85,13 @@ when it isn't running. Bring one up, `dotnet test`, tear down. Repo compose live
 > copy of the CA: `scp dohertj2@10.100.0.35:/opt/otopcua-mqtt/secrets/ca.crt /tmp/mqtt-fixture-ca.crt`.
 > It is also the only fixture whose services carry the `project: lmxopcua` label (see CLAUDE.md).
 
+> **MTConnect is a TWO-service stack.** The `mtconnect/agent` image ships no simulator, so a lone
+> Agent answers `/probe` and reports every observation `UNAVAILABLE` forever. The `adapter` service
+> (a stdlib SHDR feeder, `Docker/adapter.py`) is the data source; the Agent dials out to it. Its
+> suite also probes over **HTTP**, not TCP — an Agent's port accepts connections before its device
+> model is parsed. Full detail + the seeded device model:
+> [`tests/Drivers/ZB.MOM.WW.OtOpcUa.Driver.MTConnect.IntegrationTests/Docker/README.md`](../tests/Drivers/ZB.MOM.WW.OtOpcUa.Driver.MTConnect.IntegrationTests/Docker/README.md).
+
 > **Fixture-cycling gotcha:** profile-gated services can share a port; a plain `docker compose down`
 > may leave a stale container bound. Force-remove before switching profiles:
 > `docker rm -f $(docker ps -aq --filter name=otopcua-<driver>)` then `up --force-recreate`.