fix(testing): wire publishers to router-registered subscribers

[lesnik512]

Closes #86.

The bug

TestTimersBroker.create_publisher_fake_subscriber scanned broker._subscribers to find the real subscriber a publisher should be wired to. That list holds only endpoints registered directly on the broker — upstream's Registrator.subscribers is what includes routers':

@property
def subscribers(self) -> list["SubscriberUsecase[MsgType]"]:
    return [*self._subscribers, *(sub for r in self.routers for sub in r.subscribers)]

So anything declared on a TimersRouter was never matched, is_real came back False, and the fallback ran broker.subscriber(publisher.config.topic) — the unprefixed topic. Behind a prefix the fake landed on a topic nothing publishes to.

is_real=True is what makes upstream mirror the real handler's calls into the publisher's recorder, so the user-visible failure was:

async with TestTimersBroker(broker):
    await publisher.publish("ping", activate_in=timedelta(0))
    publisher.mock.assert_called_once_with("ping")

AssertionError: Expected 'mock' to be called once. Called 0 times.

on a publish that did happen. Anyone testing a prefixed router could not assert on their publisher.

The fix

Scan broker.subscribers, and register the fallback under full_topic rather than the raw topic. Both halves are needed: the first finds the router's subscriber, the second puts the fake on the right topic when there genuinely isn't one.

TimersBroker.subscribers now narrows the base's list[SubscriberUsecase[TimerMessage]] to list[TimersSubscriber], which is what _subscribers: list[TimersSubscriber] already claimed one line above. Narrowing at the broker rather than casting at each call site also retired an existing typing.cast in FakeTimersProducer.publish, which ty then flagged as redundant.

Before / after

Same script, three declarations:

declaration	subscribers after entering the test broker	publisher.mock
direct on broker	['reminders'] → unchanged	records → records
router, no prefix	['reminders', 'reminders'] → ['reminders']	records → records
router, prefix app:	['reminders', 'app:reminders'] → ['app:reminders']	silent → records

The no-prefix row was wrong too — a duplicate subscriber on a topic that already had one — it just happened to keep working, because the fake carries no user handler and the topic still matched. Only the prefixed row failed visibly.

The stray channel this surfaced through in #85 is gone as well:

channels: {"app:reminders:Handle": "app:reminders", "app:reminders:Publisher": "app:reminders"}

Previously a third reminders:PublisherResponseSubscriber sat alongside those.

Tests

Written first, red before the change:

FAILED tests/test_unit.py::test_publisher_is_wired_to_a_subscriber_declared_on_a_prefixed_router
  AssertionError: Expected 'mock' to be called once. Called 0 times.
FAILED tests/test_unit.py::test_router_publisher_does_not_add_a_second_subscriber[no-prefix]
  AssertionError: assert ['reminders', 'reminders'] == ['reminders']
FAILED tests/test_unit.py::test_router_publisher_does_not_add_a_second_subscriber[prefixed]
  AssertionError: assert ['reminders', 'app:reminders'] == ['app:reminders']

The existing coverage of this method was test_create_publisher_fake_subscriber_is_instance_method, which checks the signature only. Every behavioural path through it went through a broker with no router, so the whole router branch was unexercised while the gate read 100%.

Not a 0.7.6 regression

Pre-existing. Registrator.subscribers has had the router-aware definition throughout 0.7; nothing in #85 touched this path.

Verification

just lint clean (eof-fixer, ruff format, ruff check, ty). Full suite against Redis 8: 170 passed, 100% coverage.