Fix heavy-build rootfs provisioning

[H3xept]

Fixes #21.

What

• add a 64 GiB heavy-build curated image with Docker Engine, Buildx, and Compose
• make placement image-aware across local and remote nodes and return 503 no_capacity with reason image_unavailable before creating a sandbox record
• wait for docker info during guest boot and include the daemon log when startup fails
• make node provisioning boot the image through the real API/Firecracker path and verify Compose, a nested container, DNS, and HTTPS egress
• document how the curated image is staged

Why

The catalog advertised heavy-build, but node provisioning did not build its rootfs. The scheduler assumed curated images were present, so requests reached Firecracker, failed while staging a missing artifact, and left a failed sandbox record behind a generic HTTP 500.

The control plane now queries each worker's local runtime through the authenticated internal node API. Inventory checks run concurrently and each remote call has a two-second total timeout. A timeout or error makes only that worker ineligible, while nodes that report the requested artifact remain eligible for base, browser, and heavy-build placement.

Failure mode

Provisioning now stops if the Docker-capable image cannot boot or if Docker, Compose, nested containers, DNS, or bridge egress do not work. At runtime, a missing image produces a stable capacity response instead of a partial sandbox lifecycle.

Checks

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace (59 sandboxd unit tests and 22 integration tests passed)
• cargo audit (passed; one repository-allowed anyhow warning reported)
• cd sdk/typescript && npm ci && npm run build && npm pack --dry-run
• cd sdk/python && python -m build
• shellcheck deploy/provision-node.sh deploy/images/heavy-build/sandbox-init
• built the image recipe locally and verified Docker, Compose, containerd/runc, legacy iptables, jq, and git
• verified the Docker-enabled Firecracker path with an equivalent Docker 27 DinD custom image: cgroup v2, overlay2, Compose, a nested Alpine container, DNS, and HTTPS egress
• the provisioner now enforces the same runtime checks against the curated heavy-build image before a node is accepted; this exact target-image smoke runs during Linux/KVM provisioning