complete_quest proxies to the API so the panel tap is one write

WHY: Feature 18 makes the integration a client of the NestQuest API
service — every database write must go through the API.  The
complete_quest service still wrote the local database and fired the
completed event itself, which would double-fire once the SSE
subscription re-fires the API's transition (and stops being true the
moment the local database is dropped).  The service now proxies to the
entry's ONE shared API client (complete_instance: the panel complete
route) and keeps only the old immediate-update behaviour: a successful
call refreshes the coordinator so the sensors reflect the completion
at once; the bus event comes solely from the SSE stream.  Typed API
failures (and an unconfigured client, and a non-panel actor) surface
as HomeAssistantError with token-free messages instead of being
swallowed; the payload contract (instance_id, actor=panel,
actor_child_id) is unchanged for the panel cards.

Tests: new tests/test_complete_quest_proxy.py asserts the call reaches
the shared client with the right arguments, that no local write or
local event happens on the API path, that a non-2xx/transport failure
raises a service error chained to the typed error (never containing
the token), and that the unconfigured stub fails clearly.  Existing
tests that drove complete_quest through the local database move to the
API path via the harness stand-in client (which performs the same core
completion and publishes the same SSE frames the API route does) and
assert their bus events after the frames are delivered through the
real SSE subscription; the shim-level tests drive the events shim
directly, the way its remaining callers invoke it.

Task 47c7c90e (complete_quest proxies to the API, Feature 18).

custom_components/nestquest/services.py

@@ -6,7 +6,10 @@ call time through the injected ``find_runtime`` callable.  Permission
 checks are Feature 09 later tasks — this module only validates payload
 shape and forwards to the Feature 03/06/08 business layers (the
 presence writes included, through :mod:`.presence_management` — the
-ONE implementation the API admin plane shares).
+ONE implementation the API admin plane shares).  Since Feature 18 the
+integration is a client of the NestQuest API service, so
+``complete_quest`` proxies to the API's panel complete route through
+the entry's shared API client instead of writing any database.
 ``export_history_csv`` is registered with a schema but raises until
 Feature 13.  ``regenerate`` is supplied by the caller so
 the Feature 07 handler stays the one in ``__init__``.
@@ -26,6 +29,7 @@ from . import children as children_layer
 from . import completion as completion_layer
 from . import presence_management as presence_management_layer
 from . import quest_definitions as quest_definitions_layer
+from .api_client import NestQuestApiError
 from .const import (
     DOMAIN,
     DOMAIN_SERVICES,
@@ -42,7 +46,7 @@ from .const import (
     SERVICE_UNCOMPLETE_QUEST,
     SERVICE_UPDATE_QUEST_DEFINITION,
 )
-from .events import fire_quest_completed, fire_quest_uncompleted
+from .events import fire_quest_uncompleted
 from .permissions import permission_gate
 from .recurrence import ScheduleRule
 from .service_policy import SERVICE_POLICY
@@ -301,22 +305,50 @@ def _complete_quest(
     async def _handler(call: Any) -> None:
         runtime = _require_runtime(hass, find_runtime)
         instance_id = call.data["instance_id"]
-        # ``appended`` is decided under the same lock as the write, so
-        # a concurrent duplicate call (a double tap) can not produce a
-        # second transition event: the loser appends nothing.
-        result = await completion_layer.complete_instance(
-            runtime.database,
-            instance_id,
-            **_actor_kwargs(call),
-        )
-        if result.appended:
-            await fire_quest_completed(
-                hass,
-                runtime.database,
-                instance_id,
-                was_on_time=result.was_on_time,
+        # The panel contract (decision 9): a tap completes as the
+        # tapped child profile, the only actor shape the API's panel
+        # route accepts.  The payload schema is unchanged — the panel
+        # cards call it with instance_id, actor=panel, actor_child_id.
+        actor = _actor_kwargs(call)
+        if actor["actor_source"] != ACTOR_PANEL:
+            raise HomeAssistantError(
+                "nestquest.complete_quest completes through the API's "
+                "panel route as the tapped child profile; actor must "
+                "be 'panel' with actor_child_id"
             )
-            await _refresh_entities(runtime)
+        # The entry's ONE shared API client (built at setup): the
+        # coordinator's snapshot poll, the SSE subscription, and this
+        # proxy all ride the same instance.  An entry with no panel
+        # token runs the unconfigured stub, which has no completion
+        # surface — fail fast with a clear, token-free message.
+        complete_instance = getattr(
+            getattr(runtime.coordinator, "api_client", None),
+            "complete_instance",
+            None,
+        )
+        if complete_instance is None:
+            raise HomeAssistantError(
+                "NestQuest API is not configured; cannot complete a "
+                "quest"
+            )
+        try:
+            await complete_instance(instance_id, actor["actor_child_id"])
+        except NestQuestApiError as err:
+            # The typed API failure surfaces as a service error the
+            # caller sees — never swallowed, never retried here.  The
+            # typed message names method, path, and status only; the
+            # panel token never appears in it.
+            raise HomeAssistantError(
+                f"NestQuest quest completion failed: {err}"
+            ) from err
+        # NO local completion write and NO local event: the API service
+        # emits the transition on its SSE stream and the integration's
+        # subscription re-fires it on the bus — firing it here too
+        # would double-fire.  The immediate refresh keeps the old
+        # behaviour where the sensors reflect the completion the
+        # moment the service call returns instead of waiting for the
+        # poll interval.
+        await _refresh_entities(runtime)
 
     return _with_errors(_handler)