feat(gui): make the Behavior Input node placeable and configurable

[gbradham]

Follow-up to #131. That PR landed the closed loop and proved it with an integration test — but the Behavior Input node had no presence in the app at all: no palette entry, no port configuration, no properties. The loop worked when a graph was built in Python and could not be built in the Builder, which is where experiments are actually authored.

The sharp one: ports

setup_node_ports falls back to ([">in"], [">out"]) for any type it doesn't list. BehaviorInputNode has zero inputs and four outputs — Active, Behavior, On Enter, On Exit. So the fallback drew the wrong shape, and On Enter (output index 2) was unreachable: a connection drawn on the canvas would have wired Active instead. The exact connection #131's integration test makes programmatically could not be drawn by hand.

Maimu gets an entry in the same table, which only relabels its ports from in/out to exec/exec — its shape was already right by luck.

The behavior is picked, not typed

A mistyped label is silently inert. The node compares the emitted string against the configured one, so "freezng" simply never matches: the stimulus never fires and nothing anywhere says why. That is the worst failure mode available to a closed-loop stimulus, so the properties panel offers a dropdown of the loaded model's own vocabulary.

It stays an editable combo on purpose:

• a graph can be authored before any model is loaded, which is normal
• a label saved from a model that isn't currently loaded is preserved, not dropped

How the vocabulary gets there. Over LiveSignalBus, which is already the boundary between vision and the flow. The camera panel publishes the model's classes when inference goes live — the same place it already hands them to the preview overlay. The alternative, having the properties panel reach into the camera panel for a worker thread's attribute, would couple the flow editor to the vision UI's internals.

Also

• A Behavior section in the node library, so the node can be placed at all
• BehaviorInput in the interface category, so it styles like the other trigger nodes
• The confirmation window as a spin box floored at 1 (one frame is no confirmation — the configuration the node exists to prevent), with the latency stated in the panel rather than only in the docstring: at 30 fps, 5 frames is about 167 ms on top of inference

Tests — 15 new

tests/unit/gui/test_behavior_input_editor.py: the palette offers the button; the ports are the node's real four with On Enter at index 2 and Active as data not exec; Maimu no longer takes the generic fallback; the dropdown is populated from the model; a saved behavior shows; a label saved before the model loaded survives; choosing and editing both persist; the panel opens with no bus at all (vision not running); the bus hands back a copy rather than its own list; and loading a model publishes its classes to the bus.

Verification

PYTHONPATH=src QT_QPA_PLATFORM=offscreen pytest tests/ — 3642 passed, 3 skipped, ruff and black clean.