[earakely-scale]

FD-5184. The conformance probe's Fixture Desk was written for X11's Tk. On macOS, cua-driver failed four full-probe checks because of how Aqua's Tk reports keys and the wheel, not because of the driver. The fixture now records the same effects under X11, Aqua and Windows. The probe's checks are unchanged.

What changes

In computer/conformance/fixture.py:

• Keysyms: Aqua's Tk names , and space by their character. The fixture records them under X11's names, comma and space.
• Select-all: Tk binds Ctrl+/ to select-all on X11 only; Aqua's is Cmd+A. The fixture binds Ctrl+/ itself.
• The wheel:
  • Tk 8.6 on X11 reports buttons 4 and 5, as before.
  • Aqua, Windows and Tk 9 report <MouseWheel>. It is counted by its delta: 120 a notch where Tk scales it, else 1 a unit.
  • Tk 9 reports a scroll posted in pixels, which is how a macOS driver posts it, as <TouchpadScroll>. It is counted by the lines it moves, 10 px a line, because macOS coalesces quick scrolls into one event.
• Windowing system: layout.json records which one the fixture ran under.

How it was tested

• macOS, cua-driver. Cua's agentenv-cua plugin at the head of feat(agentenv-cua): AgentEnv plugin for macOS VMs (lume provider, cua-driver macOS driver) trycua/cua#6405, on a bench image built with its recipe from Cua's base, ghcr.io/trycua/macos:26-20261021-fca2193 (macOS 26.5.2, Tk 9 in the driver's Python). It ran on an EC2 mac2-m2.metal, with the full probe through a loopback relay.
  • Released protocol (0.1.449): 44 passed and 4 failed. The failures were typed key events, the select-all chord, a repeated key, and wheel directions.
  • This branch: 47 passed, 1 failed, 1 not applicable. Every wheel and trackpad check passes, in five runs on three boots.
  • The one steady failure left is the driver's. A raw log of every Tk event showed each typed character's key-down with the right key, but its key-up with keycode 0, the A key. So an app sees "a" released after each character.
    • Keys cua-driver presses (Backspace, Return, Tab, Control) release correctly, so the check keeps failing it.
  • The drag check also failed on 3 of 7 first probes after a boot, its start 3 to 9 px off. It never failed on a later probe.
    • The raw log shows the pointer entering the pad at the drag's start, (50, 50). The driver's mouse-down then arrived 0.5 s later at (56, 55), already along the path. Tk reports each event's own coordinates, so this is the driver's timing, not the platform.
    • One of those runs also missed a first click on the button.
• Linux, on an EC2 VM (c8i.4xlarge, Ubuntu 24.04, Tk 8.6):
  • ruff check passed;
  • the protocol's conformance unit tests: 55 passed, including the new tests of the wheel's and the trackpad's step counts, at the deltas macOS gave;
  • integration: the reference driver's image passing the probe at full, and cua-driver's gate 3 on the desktop computer: 2 passed.

🤖 Generated with Claude Code