# Live-test incident — 2026-08-30

This report records the failed extended Chameleon live test so the prototype
is not mistaken for release-ready software.

## What happened

OmaFrames `0.5.0-prototype` loaded into Hyprland 0.56.2 and attached both Vines
and Chameleon to one existing ChatGPT window and two Foot windows opened after
load. Runtime size, speed, jump interval, palette, and reduced-motion values
were changed through Hyprland's transient Lua configuration. A full-screen
still captured the procedural chameleon perched on a tiled window edge at 2× scale.

The failure began when an attempted full-screen video recording launched
`gpu-screen-recorder`. The screen became mostly pink, input stopped responding,
and the display later went black. Remote attempts to unload the plugin or
restart Hyprland through IPC could not complete. A targeted kill of Hyprland's
service cgroup also failed to recover the GPU, and the machine was rebooted.

## Evidence retained after reboot

The previous boot journal establishes this order:

1. At 09:39:50 the `amdgpu` driver reported a VCE ring timeout and explicitly
   named process `gpu-screen-reco` and its command-submission thread.
2. The driver began a GPU reset, reported VRAM loss, and the ChatGPT GPU process
   terminated with `SIGTRAP`; its core command line identifies it as
   `--type=gpu-process` on `/dev/dri/renderD129`.
3. At 09:40:21 the AMD graphics ring self-test failed with timeout `-110`.
4. Kernel workers and a ChatGPT graphics thread remained blocked for more than
   122 seconds in AMDGPU cleanup paths.
5. Hyprland's watchdog later raised `SIGABRT` while OmaFrames was loaded. Its
   crash report is truncated after the plugin list, consistent with the GPU
   already being wedged.
6. Subsequent AMD GPU recovery attempts failed with `-22`, including firmware
   reload, VBLANK, and page-flip timeouts.

There was no memory or swap exhaustion. The failed recorder left only an
88-byte MP4 header, which was removed. Source changes and both valid still
captures survived.

## Conclusion and limits

The evidence proves that the system-wide failure began in the AMD VCE video
encode path invoked by `gpu-screen-recorder`; the first kernel error names that
process. It does **not** prove that OmaFrames is free of compositor bugs: the
plugin was loaded in-process, and the intended flight, close-during-flight,
fullscreen, and unload tests did not complete. Conversely, there is no evidence
that Chameleon initiated the first GPU timeout.

Future live tests on this machine should avoid `gpu-screen-recorder`, retain a
remote recovery channel independent of the graphical session, and validate the
airborne render pass in a nested or disposable compositor before another host
session load.