fix: log a missing MQTT client at debug when the tree is transport-free by design - #20

Closes #11. The debug/warning split you confirmed, with the predicate exactly as you
specified it.

The predicate

Device._transport_free() — the root holds no client and was given no config to build one
from:

root = self.root()
return root.mqttc is None and root._mqtt_cfg is None

_mqtt_cfg is the term that separates the two cases. A root told nothing is passive by
request; a root told how to build a client and holding none is broken. That is the "you
forgot to start the root" warning the issue is careful to preserve, and it stays exactly as
loud as it was.

Property and Node resolve it through node -> device -> root, mirroring the walk already
in Property.start_mqtt_client. An incomplete chain falls through as not transport-free,
so a half-built tree stays loud rather than going quiet — silencing a detached property would
hide a real bug.

Before / after

Same 31-device / 120-property transport-free tree, same workload (build, set_state(READY),
publish every property, subscribe every settable):

Identical event count — nothing is dropped, only reclassified, so DEBUG still shows every
one. The blunt workaround (logging.getLogger("homie").setLevel(logging.ERROR)) is no longer
needed, which matters because it also hid the genuine warnings.

The twelve sites

All now route through one module-level helper, so the severity rule is stated once:

def _log_missing_client(message: str, *, by_design: bool, level: int = logging.WARNING) -> None:
    logger.log(logging.DEBUG if by_design else level, message)

propertyGetMqttClient, propertyStartMqttClient, propertyPublishValue,
propertyClearValue, propertySetSubscribe, nodeGetMqttClient, deviceGetMqttClient,
deviceStartMqttClient, deviceStop, deviceDeleteAllFromMqtt, deviceClearTopic,
devicePublish.

One judgement call worth your eye

devicePublishNoMqttClient was already info on main, not warning. Transport-free drops
it to debug with the rest, but the expected-a-client case keeps info rather than being
promoted — reading "WARNING otherwise" literally would have made one site louder than it is
today, which is a behaviour change the issue did not ask for. That is what the level=
parameter expresses. Say the word if you'd rather have it uniform.

Bring-your-own-transport is not transport-free

#14 landed after the issue was filed, so worth stating explicitly: an injected client sets
mqttc, so the predicate is False and a missing client on that path still warns. That is
the right answer — the caller handed a client over, so its absence is an anomaly, not a
request. There is a test pinning it.

Rebased onto #19. The predicate reads mqttc and _mqtt_cfg, which #19 retyped without
changing their meaning, so it is unaffected; _owned_client is deliberately not part of it,
since an SDK-built client that is missing is exactly the case that should stay loud.

Tests

Six in TestTransportFreeLogSeverity: no WARNING anywhere in a transport-free tree while
DEBUG records are present; a config-bearing root still warns; an injected client is not
transport-free; the predicate resolves identically from Device, Node and Property; a
detached property stays loud; and devicePublish keeps info when a client was expected.

pytest: 533 passed, on 3.14 and on 3.10. ruff check . and ruff format --check .
clean.

Also confirmed while here

Thanks for the 3.10-3.13 matrix in #16 — that closes the gap where the declared floor wasn't
exercised by CI. This branch is green across it.