[cayossarian]

Refs #223.

Every MQTT message on a parent/child panel paid for two things that grow with the tree:

• Routing. ControllerRoutes.dispatch tested each recorded subscription in turn with topic_matches_sub, which builds a fresh matcher per call. That's four routes per device, and the panel's own topics, recorded first, are matched last.
• Readiness. SpanMqttClient._on_message asks is_ready() before and after each message, and every call walked the declared tree twice.

Changes

• fix(transport): the recorded patterns live in one paho MQTTMatcher. Dispatch looks up the matching patterns in the typed route table, and the most recently recorded one still wins, decided by a recording number. _release indexes the new pattern once instead of once per held topic. paho leaves paho/mqtt/matcher.py unannotated, so mypy gets untyped_calls_exclude = ["paho.mqtt.matcher"], with a comment on the one shape that keeps its Any contained.
• fix(adapter): is_ready() keeps its answer and forgets it on the Controller's own events: a device's first $state, a change of state, a removal, a $description, and the panel's info/model value. The next call works it out again. The tree walk is unchanged.
• client.py is unchanged.

Tree changes stay correct:

• Commissioning or decommissioning arrives as a $description or a removal, which updates the routes and clears readiness.
• A rename in the SPAN app is only an info/name value. It doesn't touch routes or readiness, and it still reaches the snapshot.
• A rebuilt adapter starts with no remembered answer.

Measured

Per message, both is_ready() calls plus handle_message, on an M-series Mac. The reference tree was grown by cloning circuits.

Devices	Before	After
14	79–115 µs	~2 µs
30	87–246 µs	~2 µs
46	97–378 µs	~2 µs

Not yet measured on a CM4; the issue reporter has offered to test.

Tests

• Transport: the most recent route beats a more specific one, unsubscribing leaves a sibling or longer pattern routing, and a dropped route that is recorded again picks up what was held.
• Readiness: after every message, the answer is compared with a fresh adapter fed the same messages. That covers 5 arrival orders and 5 tree-change scenarios. A reading or a rename causes no tree walk.

Falsified:

• A first-match dispatch fails only the new overlap test; the existing ordering tests can't tell recency from specificity.
• Leaving removed patterns in the matcher fails 4 transport tests.
• Disabling each of the five readiness events in turn fails at least one test each time.

Full suite: 1872 passed, 7 skipped (live capture absent), 9 xfailed. All pre-commit hooks pass.