test(tst-24): drive the .NET and Python clients against real in-process gRPC servers

TST-24 asked for per-client wire tests against a fake gateway. An audit first
corrected the finding's premise: Go, Rust, and Java already had them — bufconn,
a loopback tonic server, and InProcessServerBuilder respectively — each already
asserting the round trip, the server-observed bearer header, and the ReplayGap
sentinel. The two genuine gaps were .NET (every test substituted the transport
interface; the test project had no server package at all) and Python (stub
monkeypatching everywhere but one opt-in TLS test).

Both now serve mxaccess_gateway.v1.MxAccessGateway over a real transport —
Kestrel h2c and grpc.aio, each on an ephemeral loopback port — and drive the
ordinary public client API against it. Only the gateway's behaviour is canned;
the framing, serialization, metadata, and status codes are genuine. Four shapes
each: full round trip with every reply field asserted, the authorization header
as received by the server (including on the streaming RPC), the ReplayGap
sentinel surfaced as the client's typed signal, and a real PERMISSION_DENIED
mapping to the typed authorization error.

The .NET client was only ever compiled in CI, never tested, so the portable job
gains a dotnet test step.

Fixes a bug the new tests caught on their first run: Python's connect() built the
grpc.aio channel inside asyncio.to_thread, and a grpc.aio channel binds to the
event loop current on the constructing thread, so every non-stub connection
raised 'There is no current event loop in thread'. No mock-based test could see
it, and the test guarding the off-loop behaviour patched create_channel and so
asserted the bug. Split resolve_channel_security (blocking TOFU probe, off-loop)
from create_channel (on-loop); the guard tests now assert both halves.

clients/python/README.md

@@ -47,6 +47,19 @@ The tests import the generated gateway and worker stubs, run fake async gateway
 stubs, verify API key metadata, exercise stream cancellation, load shared value
 and command fixtures, and check deterministic CLI output.
 
+`tests/test_wire_fake_gateway.py` is the one suite that does **not** substitute a
+stub: it serves a canned `MxAccessGatewayServicer` from a real `grpc.aio` server
+on an ephemeral loopback port and drives the ordinary `GatewayClient` API against
+it. Only the gateway's behaviour is canned — the HTTP/2 framing, protobuf
+serialization, `authorization` metadata, and gRPC status codes are genuine, so it
+catches decode and metadata breaks a stub fake cannot see. No MXAccess, no worker,
+no TLS, so it runs in the default suite. See `docs/GatewayTesting.md`
+(Client Wire Tests) for the cross-client pattern.
+
+```powershell
+python -m pytest tests/test_wire_fake_gateway.py
+```
+
 ## Packaging
 
 Install the package in editable mode for local development:
@@ -398,6 +411,15 @@ point: the `require_certificate_validation=True` keyword on
 `--require-certificate-validation` CLI flag. See
 [Gateway Configuration](../../docs/GatewayConfiguration.md#automatic-self-signed-certificate).
 
+Channel construction is split in two: `resolve_channel_security(options)` performs
+the blocking part (the trust-on-first-use certificate probe) and
+`create_channel(options, security=...)` builds the channel. The async `connect`
+classmethods run the first off the event loop and the second on it, because a
+`grpc.aio` channel binds to the event loop current on the constructing thread —
+building it inside `asyncio.to_thread` raises
+`RuntimeError: There is no current event loop in thread 'asyncio_N'`. Callers that
+build their own channel should keep `create_channel` on the loop thread.
+
 ## CLI
 
 The CLI emits deterministic JSON for automation:

clients/python/src/zb_mom_ww_mxgateway/client.py

@@ -12,7 +12,7 @@ from .auth import merge_metadata
 from .errors import ensure_protocol_success, map_rpc_error
 from .generated import mxaccess_gateway_pb2 as pb
 from .generated import mxaccess_gateway_pb2_grpc as pb_grpc
-from .options import ClientOptions, create_channel
+from .options import ClientOptions, create_channel, resolve_channel_security
 
 
 class GatewayClient:
@@ -58,9 +58,13 @@ class GatewayClient:
         if stub is not None:
             return cls(options=resolved, stub=stub)
 
-        # create_channel may perform a blocking TLS certificate probe (TOFU
-        # default); run it off the event loop so connect never freezes the loop.
-        channel = await asyncio.to_thread(create_channel, resolved)
+        # Resolving security may perform a blocking TLS certificate probe (TOFU
+        # default); run that off the event loop so connect never freezes it. The
+        # channel itself must be built on the loop thread — a grpc.aio channel
+        # binds to the loop current on the constructing thread, and a worker
+        # thread has none.
+        security = await asyncio.to_thread(resolve_channel_security, resolved)
+        channel = create_channel(resolved, security=security)
         return cls(
             options=resolved,
             stub=pb_grpc.MxAccessGatewayStub(channel),

clients/python/src/zb_mom_ww_mxgateway/galaxy.py

@@ -21,7 +21,12 @@ from .auth import merge_metadata
 from .errors import MxGatewayError, map_rpc_error
 from .generated import galaxy_repository_pb2 as galaxy_pb
 from .generated import galaxy_repository_pb2_grpc as galaxy_pb_grpc
-from .options import BrowseChildrenOptions, ClientOptions, create_channel
+from .options import (
+    BrowseChildrenOptions,
+    ClientOptions,
+    create_channel,
+    resolve_channel_security,
+)
 
 _DISCOVER_HIERARCHY_PAGE_SIZE = 5000
 _BROWSE_CHILDREN_PAGE_SIZE = 500
@@ -70,9 +75,13 @@ class GalaxyRepositoryClient:
         if stub is not None:
             return cls(options=resolved, stub=stub)
 
-        # create_channel may perform a blocking TLS certificate probe (TOFU
-        # default); run it off the event loop so connect never freezes the loop.
-        channel = await asyncio.to_thread(create_channel, resolved)
+        # Resolving security may perform a blocking TLS certificate probe (TOFU
+        # default); run that off the event loop so connect never freezes it. The
+        # channel itself must be built on the loop thread — a grpc.aio channel
+        # binds to the loop current on the constructing thread, and a worker
+        # thread has none.
+        security = await asyncio.to_thread(resolve_channel_security, resolved)
+        channel = create_channel(resolved, security=security)
         return cls(
             options=resolved,
             stub=galaxy_pb_grpc.GalaxyRepositoryStub(channel),

clients/python/src/zb_mom_ww_mxgateway/options.py

@@ -105,7 +105,72 @@ def _split_authority(endpoint: str) -> tuple[str, int]:
     return (host or "localhost", int(port))
 
 
-def create_channel(options: ClientOptions) -> grpc.aio.Channel:
+@dataclass(frozen=True)
+class ChannelSecurity:
+    """Transport security resolved for one channel.
+
+    `credentials` is `None` for a plaintext channel. `target_name_override` is
+    the SNI/authority override the TOFU path needs, kept separate from the
+    caller's explicit `server_name_override` so the caller always wins.
+    """
+
+    credentials: grpc.ChannelCredentials | None = None
+    target_name_override: str | None = None
+
+
+def resolve_channel_security(options: ClientOptions) -> ChannelSecurity:
+    """Resolve transport security for `options`, running any blocking probe.
+
+    This is the only blocking part of channel construction: the TOFU path opens
+    a real TCP+TLS socket to fetch the server's certificate. It is split out of
+    `create_channel` because a `grpc.aio` channel binds to the event loop
+    *current on the constructing thread*, so the channel itself must be built on
+    the loop thread — building it inside `asyncio.to_thread` raises
+    ``RuntimeError: There is no current event loop in thread 'asyncio_N'``. The
+    async `connect` classmethods therefore run this function off the loop and
+    then call `create_channel` on it.
+    """
+
+    if options.plaintext:
+        return ChannelSecurity()
+
+    if options.ca_file:
+        root_certificates = Path(options.ca_file).read_bytes()
+        return ChannelSecurity(
+            credentials=grpc.ssl_channel_credentials(root_certificates=root_certificates)
+        )
+
+    if options.require_certificate_validation:
+        return ChannelSecurity(credentials=grpc.ssl_channel_credentials())
+
+    # Lenient default: grpc-python has no per-channel skip-verify, so fetch the
+    # server's certificate (unverified) and pin it for this channel (TOFU).
+    # The probe opens a real blocking TCP+TLS socket, so it MUST be bounded —
+    # a black-holed / firewall-drop host would otherwise hang on the OS default
+    # connect timeout (minutes). Bound it by call_timeout (or a short fixed
+    # fallback) so the dial fails fast as a transport error.
+    host, port = _split_authority(options.endpoint)
+    probe_timeout = options.call_timeout if options.call_timeout else _TOFU_PROBE_TIMEOUT_SECONDS
+    try:
+        presented = ssl.get_server_certificate((host, port), timeout=probe_timeout)
+    except OSError as error:
+        raise MxGatewayTransportError(
+            f"failed to fetch TLS certificate from {options.endpoint}: {error}"
+        ) from error
+    # The gateway self-signed cert always carries a "localhost" SAN, so default
+    # the SNI/target-name override to it when none was supplied, tolerating
+    # dial-by-IP or hostname mismatch.
+    return ChannelSecurity(
+        credentials=grpc.ssl_channel_credentials(root_certificates=presented.encode("ascii")),
+        target_name_override="localhost",
+    )
+
+
+def create_channel(
+    options: ClientOptions,
+    *,
+    security: ChannelSecurity | None = None,
+) -> grpc.aio.Channel:
     """Create a plaintext or TLS `grpc.aio` channel from client options.
 
     The TLS default is lenient: grpc-python has no per-channel skip-verify, so
@@ -113,48 +178,29 @@ def create_channel(options: ClientOptions) -> grpc.aio.Channel:
     as the channel's only trust root (trust-on-first-use). Set
     `require_certificate_validation=True` to force system-trust verification, or
     pass `ca_file` to verify against a specific CA — both bypass the TOFU path.
+
+    Pass *security* to reuse a `ChannelSecurity` already resolved off the event
+    loop by `resolve_channel_security`; omit it and this call resolves (and may
+    block) inline. Must run on the thread owning the event loop the channel will
+    be used from.
     """
 
+    security = security if security is not None else resolve_channel_security(options)
+
     channel_options: list[tuple[str, str | int]] = [
         ("grpc.max_receive_message_length", options.max_grpc_message_bytes),
         ("grpc.max_send_message_length", options.max_grpc_message_bytes),
     ]
     if options.server_name_override:
         channel_options.append(("grpc.ssl_target_name_override", options.server_name_override))
+    elif security.target_name_override:
+        channel_options.append(("grpc.ssl_target_name_override", security.target_name_override))
 
-    if options.plaintext:
+    if security.credentials is None:
         return grpc.aio.insecure_channel(options.endpoint, options=channel_options)
 
-    if options.ca_file:
-        root_certificates = Path(options.ca_file).read_bytes()
-        credentials = grpc.ssl_channel_credentials(root_certificates=root_certificates)
-    elif options.require_certificate_validation:
-        credentials = grpc.ssl_channel_credentials()
-    else:
-        # Lenient default: grpc-python has no per-channel skip-verify, so fetch the
-        # server's certificate (unverified) and pin it for this channel (TOFU).
-        # The probe opens a real blocking TCP+TLS socket, so it MUST be bounded —
-        # a black-holed / firewall-drop host would otherwise hang on the OS default
-        # connect timeout (minutes). Bound it by call_timeout (or a short fixed
-        # fallback) so the dial fails fast as a transport error. The async
-        # `connect` classmethods run this off the event loop (asyncio.to_thread).
-        host, port = _split_authority(options.endpoint)
-        probe_timeout = options.call_timeout if options.call_timeout else _TOFU_PROBE_TIMEOUT_SECONDS
-        try:
-            presented = ssl.get_server_certificate((host, port), timeout=probe_timeout)
-        except OSError as error:
-            raise MxGatewayTransportError(
-                f"failed to fetch TLS certificate from {options.endpoint}: {error}"
-            ) from error
-        credentials = grpc.ssl_channel_credentials(root_certificates=presented.encode("ascii"))
-        # The gateway self-signed cert always carries a "localhost" SAN, so default
-        # the SNI/target-name override to it when none was supplied, tolerating
-        # dial-by-IP or hostname mismatch.
-        if not options.server_name_override:
-            channel_options.append(("grpc.ssl_target_name_override", "localhost"))
-
     return grpc.aio.secure_channel(
         options.endpoint,
-        credentials,
+        security.credentials,
         options=channel_options,
     )

clients/python/tests/test_client_session.py

@@ -12,6 +12,7 @@ from zb_mom_ww_mxgateway import client as client_module
 from zb_mom_ww_mxgateway import galaxy as galaxy_module
 from zb_mom_ww_mxgateway.galaxy import GalaxyRepositoryClient
 from zb_mom_ww_mxgateway.generated import mxaccess_gateway_pb2 as pb
+from zb_mom_ww_mxgateway.options import ChannelSecurity
 
 
 @pytest.mark.asyncio
@@ -21,11 +22,12 @@ async def test_gateway_connect_forwards_require_certificate_validation(
     """The connect convenience kwarg must reach ClientOptions (Client.Python-027)."""
     captured: dict[str, Any] = {}
 
-    def fake_create_channel(options: ClientOptions) -> object:
+    def fake_resolve(options: ClientOptions) -> ChannelSecurity:
         captured["options"] = options
-        return object()
+        return ChannelSecurity()
 
-    monkeypatch.setattr(client_module, "create_channel", fake_create_channel)
+    monkeypatch.setattr(client_module, "resolve_channel_security", fake_resolve)
+    monkeypatch.setattr(client_module, "create_channel", _stub_create_channel)
     monkeypatch.setattr(client_module.pb_grpc, "MxAccessGatewayStub", lambda channel: object())
 
     await GatewayClient.connect(
@@ -43,11 +45,12 @@ async def test_galaxy_connect_forwards_require_certificate_validation(
     """GalaxyRepositoryClient.connect must thread the flag too (Client.Python-027)."""
     captured: dict[str, Any] = {}
 
-    def fake_create_channel(options: ClientOptions) -> object:
+    def fake_resolve(options: ClientOptions) -> ChannelSecurity:
         captured["options"] = options
-        return object()
+        return ChannelSecurity()
 
-    monkeypatch.setattr(galaxy_module, "create_channel", fake_create_channel)
+    monkeypatch.setattr(galaxy_module, "resolve_channel_security", fake_resolve)
+    monkeypatch.setattr(galaxy_module, "create_channel", _stub_create_channel)
     monkeypatch.setattr(
         galaxy_module.galaxy_pb_grpc, "GalaxyRepositoryStub", lambda channel: object()
     )
@@ -61,52 +64,67 @@ async def test_galaxy_connect_forwards_require_certificate_validation(
 
 
 @pytest.mark.asyncio
-async def test_gateway_connect_runs_create_channel_off_the_event_loop(
+async def test_gateway_connect_splits_probe_off_loop_and_channel_on_loop(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None:
-    """connect must run the blocking channel factory off the loop (Client.Python-028)."""
-    ran_in_thread: dict[str, bool] = {}
+    """The blocking probe runs off the loop; the channel is built on it.
 
-    def fake_create_channel(options: ClientOptions) -> object:
-        # If this runs on the event loop thread, get_running_loop() succeeds.
-        try:
-            asyncio.get_running_loop()
-            ran_in_thread["off_loop"] = False
-        except RuntimeError:
-            ran_in_thread["off_loop"] = True
-        return object()
-
-    monkeypatch.setattr(client_module, "create_channel", fake_create_channel)
+    Client.Python-028 required the blocking TOFU probe off the event loop. The
+    channel itself must nonetheless be constructed *on* the loop thread: a
+    ``grpc.aio`` channel binds to the loop current on the constructing thread,
+    and a ``to_thread`` worker has none, so building it off-loop raises
+    ``RuntimeError: There is no current event loop``. Assert both halves.
+    """
+    where = _record_connect_threads(monkeypatch, client_module)
     monkeypatch.setattr(client_module.pb_grpc, "MxAccessGatewayStub", lambda channel: object())
 
     await GatewayClient.connect(endpoint="gateway.example:5001")
 
-    assert ran_in_thread["off_loop"] is True
+    assert where == {"resolve_off_loop": True, "create_on_loop": True}
 
 
 @pytest.mark.asyncio
-async def test_galaxy_connect_runs_create_channel_off_the_event_loop(
+async def test_galaxy_connect_splits_probe_off_loop_and_channel_on_loop(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None:
-    """GalaxyRepositoryClient.connect must also run the probe off the loop (Client.Python-028)."""
-    ran_in_thread: dict[str, bool] = {}
-
-    def fake_create_channel(options: ClientOptions) -> object:
-        try:
-            asyncio.get_running_loop()
-            ran_in_thread["off_loop"] = False
-        except RuntimeError:
-            ran_in_thread["off_loop"] = True
-        return object()
-
-    monkeypatch.setattr(galaxy_module, "create_channel", fake_create_channel)
+    """GalaxyRepositoryClient.connect splits the probe and the channel the same way."""
+    where = _record_connect_threads(monkeypatch, galaxy_module)
     monkeypatch.setattr(
         galaxy_module.galaxy_pb_grpc, "GalaxyRepositoryStub", lambda channel: object()
     )
 
     await GalaxyRepositoryClient.connect(endpoint="gateway.example:5001")
 
-    assert ran_in_thread["off_loop"] is True
+    assert where == {"resolve_off_loop": True, "create_on_loop": True}
+
+
+def _stub_create_channel(options: ClientOptions, *, security: ChannelSecurity) -> object:
+    return object()
+
+
+def _on_event_loop_thread() -> bool:
+    try:
+        asyncio.get_running_loop()
+    except RuntimeError:
+        return False
+    return True
+
+
+def _record_connect_threads(monkeypatch: pytest.MonkeyPatch, module: Any) -> dict[str, bool]:
+    """Patch *module*'s channel helpers to record which thread each ran on."""
+    where: dict[str, bool] = {}
+
+    def fake_resolve(options: ClientOptions) -> ChannelSecurity:
+        where["resolve_off_loop"] = not _on_event_loop_thread()
+        return ChannelSecurity()
+
+    def fake_create_channel(options: ClientOptions, *, security: ChannelSecurity) -> object:
+        where["create_on_loop"] = _on_event_loop_thread()
+        return object()
+
+    monkeypatch.setattr(module, "resolve_channel_security", fake_resolve)
+    monkeypatch.setattr(module, "create_channel", fake_create_channel)
+    return where
 
 
 @pytest.mark.asyncio

clients/python/tests/test_wire_fake_gateway.py

@@ -0,0 +1,288 @@
+"""Wire-level tests: the Python client against a real localhost gRPC server.
+
+Every other test in this suite substitutes a fake *stub* object for
+``pb_grpc.MxAccessGatewayStub``, so nothing between the client wrapper and the
+generated stub is exercised: no HTTP/2 framing, no protobuf serialization, no
+call metadata, no gRPC status translation. That leaves a class of contract break
+— a field the gateway populates but the client never decodes, metadata the
+client believes it sends but does not, a status code it maps differently once it
+arrives as a real ``grpc.RpcError`` — invisible to the default suite.
+
+These tests close that gap by serving the real ``mxaccess_gateway.v1.MxAccessGateway``
+service from an in-process ``grpc.aio`` server bound to ``127.0.0.1:0`` and
+driving the ordinary public client API against it. The bytes on the wire are the
+real ones; only the gateway's *behavior* is canned. No MXAccess, no worker, no
+network beyond loopback, so this runs everywhere the normal suite runs.
+
+See ``docs/GatewayTesting.md`` (Client Wire Tests) for the shared pattern and its
+counterpart in the .NET client.
+"""
+
+from __future__ import annotations
+
+import socket
+from collections.abc import AsyncIterator, Awaitable, Callable
+
+import grpc
+import pytest
+import pytest_asyncio
+
+from zb_mom_ww_mxgateway import ClientOptions, GatewayClient
+from zb_mom_ww_mxgateway.errors import MxGatewayAuthorizationError
+from zb_mom_ww_mxgateway.events import ReplayGap
+from zb_mom_ww_mxgateway.generated import mxaccess_gateway_pb2 as pb
+from zb_mom_ww_mxgateway.generated import mxaccess_gateway_pb2_grpc as pb_grpc
+
+API_KEY = "mxgw_wiretest_secret"
+SESSION_ID = "wire-session-1"
+SERVER_HANDLE = 4242
+ITEM_HANDLE = 77
+
+
+def _free_port() -> int:
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as sock:
+        sock.bind(("127.0.0.1", 0))
+        return int(sock.getsockname()[1])
+
+
+def _ok() -> pb.ProtocolStatus:
+    return pb.ProtocolStatus(code=pb.PROTOCOL_STATUS_CODE_OK)
+
+
+class FakeGateway(pb_grpc.MxAccessGatewayServicer):
+    """Canned gateway serving the four session RPCs over a real transport.
+
+    Replies are shaped like the gateway's own: an OK ``ProtocolStatus``, the
+    echoed session id, and the typed payload the client wrapper reads (for
+    example ``RegisterReply.server_handle``). Set ``deny`` to make ``Invoke``
+    abort with ``PERMISSION_DENIED`` so the client's gRPC-status mapping is
+    exercised against a genuine ``grpc.RpcError`` rather than a hand-built one.
+    """
+
+    def __init__(self, *, deny: bool = False, replay_gap: pb.ReplayGap | None = None) -> None:
+        self.deny = deny
+        self.replay_gap = replay_gap
+        self.endpoint = ""
+        self.metadata_by_method: dict[str, str] = {}
+        self.open_request: pb.OpenSessionRequest | None = None
+        self.invoke_request: pb.MxCommandRequest | None = None
+        self.stream_request: pb.StreamEventsRequest | None = None
+        self.close_request: pb.CloseSessionRequest | None = None
+
+    def _record(self, method: str, context: grpc.aio.ServicerContext) -> None:
+        for key, value in context.invocation_metadata() or ():
+            if key == "authorization":
+                self.metadata_by_method[method] = value
+
+    async def OpenSession(  # noqa: N802 - generated gRPC method name
+        self, request: pb.OpenSessionRequest, context: grpc.aio.ServicerContext
+    ) -> pb.OpenSessionReply:
+        """Answer ``OpenSession`` with a fully populated reply."""
+        self._record("OpenSession", context)
+        self.open_request = request
+        return pb.OpenSessionReply(
+            session_id=SESSION_ID,
+            backend_name="fake-backend",
+            worker_process_id=1234,
+            worker_protocol_version=1,
+            capabilities=["events", "invoke"],
+            gateway_protocol_version=3,
+            protocol_status=_ok(),
+        )
+
+    async def Invoke(  # noqa: N802 - generated gRPC method name
+        self, request: pb.MxCommandRequest, context: grpc.aio.ServicerContext
+    ) -> pb.MxCommandReply:
+        """Answer ``Invoke`` with a Register reply, or deny when configured."""
+        self._record("Invoke", context)
+        self.invoke_request = request
+        if self.deny:
+            await context.abort(grpc.StatusCode.PERMISSION_DENIED, "invoke scope required")
+        return pb.MxCommandReply(
+            session_id=request.session_id,
+            correlation_id=request.client_correlation_id,
+            kind=request.command.kind,
+            protocol_status=_ok(),
+            hresult=0,
+            register=pb.RegisterReply(server_handle=SERVER_HANDLE),
+        )
+
+    async def StreamEvents(  # noqa: N802 - generated gRPC method name
+        self, request: pb.StreamEventsRequest, context: grpc.aio.ServicerContext
+    ) -> AsyncIterator[pb.MxEvent]:
+        """Stream an optional replay-gap sentinel followed by one data change."""
+        self._record("StreamEvents", context)
+        self.stream_request = request
+        if self.replay_gap is not None:
+            # The sentinel shape the gateway emits: family unspecified, body
+            # unset, only replay_gap populated.
+            yield pb.MxEvent(session_id=request.session_id, replay_gap=self.replay_gap)
+        yield pb.MxEvent(
+            session_id=request.session_id,
+            family=pb.MX_EVENT_FAMILY_ON_DATA_CHANGE,
+            server_handle=SERVER_HANDLE,
+            item_handle=ITEM_HANDLE,
+            value=pb.MxValue(int32_value=17),
+            quality=192,
+            worker_sequence=9,
+            on_data_change=pb.OnDataChangeEvent(),
+        )
+
+    async def CloseSession(  # noqa: N802 - generated gRPC method name
+        self, request: pb.CloseSessionRequest, context: grpc.aio.ServicerContext
+    ) -> pb.CloseSessionReply:
+        """Answer ``CloseSession`` with a closed final state."""
+        self._record("CloseSession", context)
+        self.close_request = request
+        return pb.CloseSessionReply(
+            session_id=request.session_id,
+            final_state=pb.SESSION_STATE_CLOSED,
+            protocol_status=_ok(),
+        )
+
+
+ServeGateway = Callable[..., Awaitable[FakeGateway]]
+
+
+@pytest_asyncio.fixture
+async def serve_gateway() -> AsyncIterator[ServeGateway]:
+    """Yield a factory that serves a :class:`FakeGateway` on loopback.
+
+    Each call starts its own server on a free port and records it for teardown,
+    so a test can serve a differently-configured gateway without a fixture per
+    variant.
+    """
+    servers: list[grpc.aio.Server] = []
+
+    async def _start(**kwargs: object) -> FakeGateway:
+        fake = FakeGateway(**kwargs)  # type: ignore[arg-type]
+        server = grpc.aio.server()
+        pb_grpc.add_MxAccessGatewayServicer_to_server(fake, server)
+        port = _free_port()
+        server.add_insecure_port(f"127.0.0.1:{port}")
+        await server.start()
+        servers.append(server)
+        fake.endpoint = f"127.0.0.1:{port}"
+        return fake
+
+    try:
+        yield _start
+    finally:
+        for server in servers:
+            await server.stop(grace=None)
+
+
+async def _connect(fake: FakeGateway) -> GatewayClient:
+    return await GatewayClient.connect(
+        ClientOptions(
+            endpoint=fake.endpoint,
+            api_key=API_KEY,
+            plaintext=True,
+            call_timeout=10.0,
+        )
+    )
+
+
+@pytest.mark.asyncio
+async def test_session_round_trip_decodes_real_wire_bytes(serve_gateway: ServeGateway) -> None:
+    """Open, invoke, stream, and close against a real server over loopback."""
+    wire_gateway = await serve_gateway()
+    client = await _connect(wire_gateway)
+    try:
+        session = await client.open_session(client_session_name="wire-test")
+        assert session.session_id == SESSION_ID
+        assert session.open_reply.backend_name == "fake-backend"
+        assert list(session.open_reply.capabilities) == ["events", "invoke"]
+
+        server_handle = await session.register("wire-test-client")
+        assert server_handle == SERVER_HANDLE
+        assert wire_gateway.invoke_request is not None
+        assert wire_gateway.invoke_request.command.kind == pb.MX_COMMAND_KIND_REGISTER
+        assert wire_gateway.invoke_request.command.register.client_name == "wire-test-client"
+
+        events = [event async for event in session.stream_events()]
+        assert len(events) == 1
+        event = events[0]
+        assert not isinstance(event, ReplayGap)
+        assert event.family == pb.MX_EVENT_FAMILY_ON_DATA_CHANGE
+        assert event.server_handle == SERVER_HANDLE
+        assert event.item_handle == ITEM_HANDLE
+        assert event.value.int32_value == 17
+        assert event.quality == 192
+        assert event.worker_sequence == 9
+        assert event.HasField("on_data_change")
+
+        close_reply = await session.close()
+        assert close_reply.final_state == pb.SESSION_STATE_CLOSED
+        assert wire_gateway.close_request is not None
+        assert wire_gateway.close_request.session_id == SESSION_ID
+    finally:
+        await client.close()
+
+
+@pytest.mark.asyncio
+async def test_api_key_reaches_the_server_on_every_rpc(serve_gateway: ServeGateway) -> None:
+    """The bearer header is on the wire for unary and streaming calls alike.
+
+    Stub-substituting tests can only assert what the client *passes*; this
+    asserts what the server *receives*, which is the property that matters.
+    """
+    wire_gateway = await serve_gateway()
+    client = await _connect(wire_gateway)
+    try:
+        session = await client.open_session(client_session_name="wire-test")
+        await session.register("wire-test-client")
+        async for _ in session.stream_events():
+            break
+        await session.close()
+    finally:
+        await client.close()
+
+    expected = f"Bearer {API_KEY}"
+    assert wire_gateway.metadata_by_method == {
+        "OpenSession": expected,
+        "Invoke": expected,
+        "StreamEvents": expected,
+        "CloseSession": expected,
+    }
+
+
+@pytest.mark.asyncio
+async def test_replay_gap_sentinel_survives_the_wire(serve_gateway: ServeGateway) -> None:
+    """A resumed stream surfaces the gateway's sentinel as a typed ``ReplayGap``."""
+    replay_gap_gateway = await serve_gateway(
+        replay_gap=pb.ReplayGap(requested_after_sequence=3, oldest_available_sequence=8)
+    )
+    client = await _connect(replay_gap_gateway)
+    try:
+        session = await client.open_session(client_session_name="wire-test")
+        items = [item async for item in session.stream_events(after_worker_sequence=3)]
+    finally:
+        await client.close()
+
+    assert len(items) == 2
+    gap = items[0]
+    assert isinstance(gap, ReplayGap)
+    assert gap.requested_after_sequence == 3
+    assert gap.oldest_available_sequence == 8
+    assert gap.resume_after_worker_sequence == 7
+    assert not isinstance(items[1], ReplayGap)
+    assert items[1].family == pb.MX_EVENT_FAMILY_ON_DATA_CHANGE
+
+    assert replay_gap_gateway.stream_request is not None
+    assert replay_gap_gateway.stream_request.after_worker_sequence == 3
+
+
+@pytest.mark.asyncio
+async def test_permission_denied_maps_to_authorization_error(
+    serve_gateway: ServeGateway,
+) -> None:
+    """A real ``PERMISSION_DENIED`` status becomes the typed client error."""
+    denying_gateway = await serve_gateway(deny=True)
+    client = await _connect(denying_gateway)
+    try:
+        session = await client.open_session(client_session_name="wire-test")
+        with pytest.raises(MxGatewayAuthorizationError):
+            await session.register("wire-test-client")
+    finally:
+        await client.close()