fix+docs: make docker compose up -d actually work

[kdairatchi]

Summary

Running `docker compose up -d` against the published image crashed the container in a restart loop. Two real defects + one doc bug. All three fixed in this PR.

1. User/UID collision (runtime crash, exit 4)

Upstream `/opt/gem/run.sh` defaults `USER=gem USER_UID=1000 USER_GID=1000` and does `groupadd --gid 1000 gem`. But `cybersandbox/Dockerfile:143` already baked a `hunter` user at UID/GID 1000. Collision: `groupadd: GID '1000' already exists` → `set -e` kills the script → restart loop.
Fix: set `USER=hunter USER_UID=1000 USER_GID=1000` in the root compose env.

2. Premature privilege drop (runtime crash, exit 1)

`cybersandbox/Dockerfile:150` ends with `USER 1000`. Upstream's run.sh expects root at entry so it can useradd, `chown -R` /opt/jupyter, and write /etc/sudoers.d/$USER — it self-drops to $USER via `su -` before exec'ing supervisord. Running as UID 1000 means chown fails → `set -e` → crash. Running as root at entry does not weaken the app-layer posture: every supervised process (nginx, jupyter, code-server, mcp-hub, browser, VNC) still runs as hunter.
Fix: `user: "0:0"` on the sandbox service. Clean fix is to drop `USER 1000` from the Dockerfile + rebuild; tracked as a follow-up issue.

3. Two compose files, one README pointed at the wrong one

• `./docker-compose.yaml` — consumer setup (pulls the published image).
• `cybersandbox/docker-compose.yml` — maintainer setup (builds locally, mounts a personal Obsidian vault at `/mnt/c/Users/Dr34d/...`, binds `127.0.0.1:8082`).

README told new users `cd cybersandbox && docker compose up -d`, dropping them onto the maintainer file with host paths that don't exist on their machines.

Fix:

• Rename `cybersandbox/docker-compose.yml` → `cybersandbox/docker-compose.dev.yml` so the filename self-documents.
• Point root README at `./docker-compose.yaml` (run from repo root, no `cd`).
• Document the WSL `docker-credential-desktop.exe` failure mode in the README (this is what surfaced the whole problem).
• Update `cybersandbox/SETUP.md` commands to `-f docker-compose.dev.yml` and banner back to root quick-start.

Verification (done on this branch)

```
docker compose up -d # healthy in ~15s
docker compose ps # Up (healthy) 0.0.0.0:8080->8080
curl -s -o /dev/null -w "%{http_code}" :8080 # 200

python -c "from agent_sandbox import Sandbox; c=Sandbox(base_url='http://localhost:8080'); print(c.shell.exec_command(command='echo hi').data)"

session_id='...' output='hi' exit_code=0

```

All 11 supervisord services up (python-server, gem-server, browser, nginx, websocat, code-server, mcp-server-browser, jupyter, mcp-hub, openbox, tigervnc).

Follow-ups (not in this PR)

• Drop `USER 1000` from `cybersandbox/Dockerfile:150` and rebuild the image. That lets us remove the `user: "0:0"` override from compose. (Separate issue — needs a CI image build + new :latest push.)

Test plan

• `docker compose up -d` succeeds from repo root on WSL against published image
• HTTP 200 on :8080 after 15s
• Python SDK round-trip confirmed
• Both compose files still parse (`docker compose config --quiet`)
• Reviewer runs the same sequence on a fresh machine