[Detail Bug] Cat plugin: decorative cat stops wandering when app window loses focus (alt-tab/minimize)

Start in apps/knx-gui/src/knx_gui/plugins/cat/follower.py at CatFollower.render(), then reproduce the behavior by alt-tabbing or minimizing the application while the cat is wandering. Verify that an invalid unfocused-window mouse position no longer cancels wandering and that normal mouse movement still triggers the sprint back to the corner.

Detail Bug Report

https://app.detail.dev/org_62aa40f5-2c23-4914-a665-3bb2068af20e/bugs/bug_5bb67883-0bde-4cba-abdb-60af32f757bf

Introduced in 187848f4163973ac8cc27ec8c31301b58af25e72 by @kewde on May 17, 2026

Summary

• Context: CatFollower (apps/knx-gui/src/knx_gui/plugins/cat/follower.py) is the decorative-cat plugin. docs/plugin.md lists cat as the only plugin with no panels and no service — "Cosmetic desktop cat follower." A sprite sits at the bottom-left corner, wanders to a random spot along the bottom every 3–8 s of idle time, and sprints back to the corner when the mouse moves. The idle-wander feature was introduced by commit 187848f ("refactor(cat): replace click-following with idle wander behavior"): "Cat now wanders slowly when user is away and sprints back to corner when mouse movement is detected."
• Bug: mouse_moved is computed from io.mouse_pos before the validity guard on the next line that rejects Dear ImGui's "no mouse" sentinel (-FLT_MAX, -FLT_MAX), so when the mouse position is the sentinel, mouse_moved is True on every frame and the spook-to-corner branch fires the frame after every wander attempt — aborting the wander. The actual trigger is the application window losing focus (alt-tab to another application, or the window being minimized / obscured): Dear ImGui's GLFW backend sets io.MousePos = ImVec2(-FLT_MAX, -FLT_MAX) each frame when the window is unfocused.
• Actual vs. expected: When the window is unfocused, the cat remains pinned within ~_WANDER_DIST of the corner (wander attempts are immediately cancelled). Expected: the cat continues the normal idle-wander behavior.
• Impact: Low / cosmetic. While the app window is unfocused, the decorative cat does not roam along the bottom as intended.

Code with Bug

apps/knx-gui/src/knx_gui/plugins/cat/follower.py, CatFollower.render():

        io = imgui.get_io()
        mx, my = io.mouse_pos.x, io.mouse_pos.y
        mouse_moved = (
            abs(mx - self._last_mouse_x) > _MOUSE_MOVE_THRESHOLD
            or abs(my - self._last_mouse_y) > _MOUSE_MOVE_THRESHOLD
        )  # <-- BUG 🔴 computed before mouse-pos validity check; -FLT_MAX sentinel makes this True every frame
        if abs(mx) < 100_000 and abs(my) < 100_000:
            self._last_mouse_x, self._last_mouse_y = mx, my

Spook branch that then cancels wandering:

        if mouse_moved and (self._wandering or self._returning) and not self._sprinting:
            self._sprinting = True
            self._wandering = False
            self._returning = False
            self._target_x = corner_x
            self._target_y = ground_y

Explanation

• When the Dear ImGui GLFW backend reports no mouse (window unfocused), it sets io.mouse_pos to (-FLT_MAX, -FLT_MAX).
• The code correctly avoids saving this sentinel into _last_mouse_x/_last_mouse_y, so those values remain frozen at the last real cursor position.
• However, mouse_moved is computed before the validity guard. With (mx, my) == (-FLT_MAX, -FLT_MAX) and _last_mouse_* frozen, abs(mx - last) > _MOUSE_MOVE_THRESHOLD is True every frame.
• As a result, on any frame after wandering starts (self._wandering=True), the spook branch triggers immediately and retargets the cat back to the corner, effectively preventing any visible wander while unfocused.

Codebase Inconsistency

In commit 187848f, the old logic early-returned on invalid mouse values (if abs(mx) > 100_000 or abs(my) > 100_000: return), preventing the sentinel from affecting state. The refactor removed the early return and introduced the new ordering where mouse_moved is computed from unvalidated values.

Recommended Fix

Compute mouse_moved only when the mouse position is valid (reuse the existing validity check):

        _valid = abs(mx) < 100_000 and abs(my) < 100_000
        mouse_moved = _valid and (
            abs(mx - self._last_mouse_x) > _MOUSE_MOVE_THRESHOLD
            or abs(my - self._last_mouse_y) > _MOUSE_MOVE_THRESHOLD
        )
        if _valid:
            self._last_mouse_x, self._last_mouse_y = mx, my

History

This bug was introduced in commit 187848f4. That refactor ("replace click-following with idle wander behavior") replaced the prior click-to-follow logic — which early-returned from render() on if abs(mx) > 100_000 or abs(my) > 100_000: return, so the sentinel never reached any state — with the new idle-wander/sprint-back design, and in doing so split the validity check into a non-returning guard that freezes _last_mouse_* while computing mouse_moved before that guard runs.