[Slach]

Why

Testflows jobs occasionally hang for 30-40 minutes in tests/testflows/infra/setup.sh on
apt-get install ffmpeg libsm6 libxext6. The ffmpeg dependency tree (~90 MB) sometimes downloads
from azure.archive.ubuntu.com at ~40 kB/s:

• run 40400570908, latest / functions: Fetched 85.4 MB in 38min 10s (37.3 kB/s)
• run 40351239613, fixed / functions: Fetched 91.5 MB in 29min 21s (52.0 kB/s)

The host ffmpeg is never used. Session video is recorded by the selenium/video:ffmpeg-8.1-*
sidecar container (se:recordVideo + video-image in infra/config.toml), which has its own ffmpeg.
The files only get copied out via the tmp/target mount.

Changes

• setup.sh: remove the second apt-get update and the apt-get install ffmpeg libsm6 libxext6.
• requirements.txt: opencv-python -> opencv-python-headless (same 4.9.0.80). The headless
  build has no X11 runtime deps, so libsm6/libxext6 are not needed. cv2 is only used for
  contour detection in steps/actions.py::check_screenshot (no GUI calls).

Local verification

• functions suite on Grafana latest: 13/13 scenarios OK. video.mp4 was still recorded in tmp/target.
• data_source_setup_connections (the suite that exercises check_screenshot / cv2): see the comment below.

🤖 Generated with Claude Code

[Slach]

Local verification on Grafana latest (macOS + OrbStack, SKIP_BUILDERS=1), with cv2 loaded from opencv-python-headless==4.9.0.80 (checked via cv2.__file__):

• data_source_setup_connections: this is the suite that calls actions.check_screenshot → cv2.cvtColor/inRange/findContours/minAreaRect/boxPoints. 19/19 scenarios OK, 8m27s, 9 checking screenshot steps passed. The only errored attempt is the usual first retry of wait for grafana to be started.
• functions: 13/13 scenarios OK. video.mp4 was recorded by the selenium/video sidecar.