Mouse accuracy
Claim. The target's own kernel sees the absolute mouse within 1 unit of 32767 across a grid (about 0.06 px at 1080p). In X the cursor lands within 1 px of the requested pixel.
Method. live_mouse_test.py sends raw mouse events through the Comet, then drives a graphical session (minimal Xorg, openbox, xterm, xev, xdotool). Each result is checked a second way on the target over ssh: the kernel input events, and X's own pointer position. The cursor sprite found on the KVM's own picture was also compared with the request. Target: Ubuntu 26.04 at 1920x1080.
Sample size. 17 checks on raw events and 26 on the graphical session. The grid size is not recorded in the source note. Examples reported: 5 asked, 4 landed; 1914 asked, 1914 landed.
Result. Kernel: within 1/32767. X: within 1 px. Clicks arrive as buttons 1, 2, 3 at the exact pixel; the wheel as 4 and 5; a drag delivers motion with the button held and releases at the end point. Click on OCR word boxes double-clicked two words about 66 px apart and X reported exactly each word selected.
Limits.
- One Comet, one target, one resolution.
- The BIOS tested ignores the mouse, so none of this applies there. Use keys in a BIOS.
- A Linux text console with
gpmrunning takes the mouse for itself. - Absolute mouse only appears once the target enumerates it.
- The per-check pass counts are as reported in the notes.
Raw data. Test script scripts/kvm/live_mouse_test.py and summary in docs/kvm/README.md (section "Measured on psycho"), both in this repository. Rerun needs a Comet and the target set up with scripts/kvm/target-mouse-tools.sh.