fix(york_region): honour the classified markers in _call_client

Addresses the Codex P2 on PR #5. Correct — and the same generalisation gap as
the previous three findings: I converted the raise in york_region/client.py
without checking the handler that catches it.

york_region carries no @upstream_guard; its _call_client helper is the
catch-all (that is why Phase 20.2 skipped its 28 tools). So NotFound raised
for an unknown dataset id fell past NoPortalError and the HTTP arms into the
generic `except Exception` and a routine missing record was reported as an
upstream outage. Marker arms now come first.

Audited the rest behaviourally rather than assuming york_region was the only
one: every module whose own client raises a marker now returns the right code
(york_region/manitoba/saskatchewan/statcan NotFound, ircc/saskatchewan/
manitoba InvalidInput). A first pass over ALL tools reported 118 mishandlings
— that was my probe passing junk arguments like station="x", which trips each
tool's own validation before the mock is reached. Restricted to real paths,
mishandlings are zero.

test_markers_survive_the_module_handler covers the four modules that raise
markers, so a handler that drops one fails loudly. Verified non-vacuous:
reverting york_region/tools.py fails exactly that case.

3170 passed, 97.15% coverage.

Author: mariomeyer

src/mcp_canada/modules/york_region/tools.py

@@ -32,6 +32,7 @@
 )
 from mcp_canada.modules.york_region.constants import PORTAL_URLS
 from mcp_canada.shared.envelope import make_error, make_response
+from mcp_canada.shared.errors import InvalidInput, NotFound, UpstreamData
 
 API_NAME = "arcgis-hub"
 
@@ -49,8 +50,14 @@ async def _call_client(
 ) -> dict[str, Any]:
     """Run a client coroutine and wrap result with make_response / make_error.
 
-    Centralises NoPortalError -> NOT_FOUND, HTTP 404 -> NOT_FOUND,
-    other HTTPStatusError -> UPSTREAM_ERROR, generic Exception -> UPSTREAM_ERROR.
+    Centralises the classified markers, NoPortalError -> NOT_FOUND,
+    HTTP 404 -> NOT_FOUND, other HTTPStatusError -> UPSTREAM_ERROR,
+    generic Exception -> UPSTREAM_ERROR.
+
+    The marker arms must come first. These tools carry no ``@upstream_guard``
+    (this helper is their catch-all), so without them an unknown dataset id
+    raised as ``NotFound`` fell into the generic arm below and a routine
+    missing record was reported as an upstream outage.
     """
     try:
         data, cached = await coro
@@ -61,6 +68,12 @@ async def _call_client(
             cached=cached,
             lang=lang,
         )
+    except InvalidInput as e:
+        return make_error("INVALID_INPUT", str(e), lang=lang)
+    except NotFound as e:
+        return make_error("NOT_FOUND", str(e), lang=lang)
+    except UpstreamData as e:
+        return make_error("UPSTREAM_ERROR", f"upstream returned unusable data: {e}", lang=lang)
     except NoPortalError as e:
         return make_error("NOT_FOUND", str(e), lang=lang)
     except httpx.HTTPStatusError as e:

tests/test_error_classification_defaults.py

@@ -169,3 +169,46 @@ def test_invalid_input_is_a_valueerror_for_backwards_compatibility():
     from mcp_canada.shared.errors import InvalidInput
 
     assert issubclass(InvalidInput, ValueError)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("module", "client_attr", "tool_name", "kwargs", "marker", "expected"),
+    [
+        ("york_region", "fetch_get_dataset_details", "york_region_get_dataset_details",
+         {"dataset_id": "bogus"}, "NotFound", "NOT_FOUND"),
+        ("manitoba", "fetch_livestock_prices", "manitoba_get_livestock_prices",
+         {}, "InvalidInput", "INVALID_INPUT"),
+        ("saskatchewan", "fetch_fire_bans", "saskatchewan_get_fire_bans",
+         {"ban_scope": "urban"}, "InvalidInput", "INVALID_INPUT"),
+        ("ircc", "fetch_permanent_residents", "ircc_get_permanent_residents",
+         {}, "InvalidInput", "INVALID_INPUT"),
+    ],
+)
+async def test_markers_survive_the_module_handler(
+    module, client_attr, tool_name, kwargs, marker, expected
+):
+    """Classifying at the raise site is useless if the handler drops it.
+
+    york_region carries no @upstream_guard — its `_call_client` helper is the
+    catch-all — so a NotFound raised for an unknown dataset id fell into the
+    generic arm and a routine missing record was reported as an upstream
+    outage. Caught by Codex on PR #5. This covers each module whose own client
+    raises a marker.
+    """
+    import importlib
+    from unittest.mock import AsyncMock, patch
+
+    import mcp_canada.shared.errors as errors
+
+    tools = importlib.import_module(f"mcp_canada.modules.{module}.tools")
+    client = importlib.import_module(f"mcp_canada.modules.{module}.client")
+    target = tools if hasattr(tools, client_attr) else client
+    exc = getattr(errors, marker)("probe")
+
+    with patch.object(target, client_attr, new=AsyncMock(side_effect=exc)):
+        result = await getattr(tools, tool_name)(**kwargs)
+
+    assert result["error"]["code"] == expected, (
+        f"{module}.{tool_name} dropped a {marker} raised by its client"
+    )