docs: Arc mouse click-loss investigation (open)

[bearyjd]

Summary

Adds docs/research/arc-mouse-click-loss-2026-10-03.md, a working record of an open investigation into dropped left clicks (worst on the Microsoft Arc Touch BT mouse). Docs only; no image, kargs, or test changes.

Key findings so far (details and captures in the doc):

• The mouse is Bluetooth LE HID-over-GATT.
• Of 10 physical clicks, only 5 reached evdev and 3 were seen over the air, so about half are lost before the kernel (mouse button/firmware or the BLE link).
• A release/re-press 6-7 ms apart inside one press shows button chatter at the source.
• Failed clicks also never appear in kwin's own event list, though a fresh libinput context sees them.
• Reconnecting the mouse changed nothing, which weakens the stale-per-device-state / debounce hypothesis.
• Ruled out: the 20260920 to 20260927 image diff (only containerd + selinux-policy), SELinux denials, input-remapper, BT controller autosuspend, compositor stall, terminal mouse reporting.

A machine-local libinput quirk experiment (ModelBouncingKeys=1 for the Arc mouse in /etc/libinput/local-overrides.quirks) is applied but not in the repo. The doc records the stanza and the revert steps.

Test plan

• Pair the mouse to another device and count clicks (decides worn mouse vs this laptop's BLE handling)
• Log out/in with the quirk applied and record how long until clicks fail again
• If the quirk helps, decide whether to bake it into the image (hardware-tuning change per CLAUDE.md; update CODEMAPS and tests/smoke.sh in that change)
• Update or close out the doc with the final root cause

docs/** is in paths-ignore for build.yml, so no image rebuild is triggered.