container: give agent runs docker, gh auth and a git identity

[ClaydeCode]

Why

A headless run started by the scheduler or the Pebble worker could not commit, could not push, and could not talk to Docker. Verified in the running container on the VM today:

• gh auth status → not logged into any GitHub hosts. GH_TOKEN is exported only inside run_cycle() of the Freeshard loop, and that loop is off on the VM (CLAYDE_FS_ENABLED=false), so nothing ever set it. The credential helper is !gh auth git-credential, so pushes fail too.
• git commit in the container → Author identity unknown. No user.name/user.email is configured anywhere, although CLAYDE_GIT_NAME and CLAYDE_GIT_EMAIL exist in config.
• No docker binary in the image.

CLAUDE.md already documented the first two as if they were true (lines 5 and 148). This makes them true.

The motivating case is running the app-repository update-apps pass as a weekly scheduled task: update/update.py apply gates every app on docker compose pull --dry-run, then commits, pushes and opens a PR.

What changed

• config.bootstrap_process_env() exports GH_TOKEN and sets the global git identity from settings. Called from both entry points (orchestrator._run_with_pebble, freeshard.entry.run_loop) instead of happening as a side effect of one loop; the in-cycle export and its test move here.
• Dockerfile installs docker-ce-cli + docker-compose-plugin from Docker's Debian repo, codename taken from /etc/os-release so a base-image bump does not silently break it.
• docker-compose.yml mounts /var/run/docker.sock and adds the host's docker group via group_add: ["${DOCKER_GID:-111}"], needed because the container runs as uid 1000.

The socket is the full one. An agent in this container can control every container on the host. That is deliberate and was weighed against running the update pass outside the container; a read-only socket proxy would also work if we later want it.

Verified before opening this PR

Built the Docker-install layer on python:3.13-slim and ran the resulting client against the host daemon as uid 1000:

Check	Result
docker --version / docker compose version in the built image	29.8.1 / v5.5.1
docker version as -u 1000:1000 --group-add 111	reaches the daemon (29.2.1)
same, without --group-add	permission denied ... unix:///var/run/docker.sock
docker compose pull --dry-run -q on a rendered app template (audiobookshelf:2.36.0)	exit 0
same with a bogus tag	exit 1, manifest unknown

The last two are the gate update.py depends on, and they discriminate.

uv run pytest: 418 passed.

Deploy note

~/clayde/docker-compose.yml on the VM is maintained by hand and needs the same two edits, plus DOCKER_GID=111 in ~/clayde/.env (this host's getent group docker). The container must be recreated, not just restarted, for the mount to apply.

Recommended reading order

1. src/clayde/config.py — the new function
2. src/clayde/orchestrator.py, src/clayde/freeshard/entry.py — call sites
3. src/clayde/freeshard/loop.py — the removed duplicate
4. Dockerfile, docker-compose.yml
5. tests/test_config.py, tests/freeshard/test_loop.py