lifecycle: topological resolver, wait_for_health, graceful shutdown

[pierrick-fonquerne]

Coordinate the lifecycle of every resource declared in the manifest: resolve their start order, gate readiness on a health policy, and tear everything down cleanly on signal.

Scope

• Dependency graph builder from the parsed manifest.
• Topological sort with cycle detection.
• wait_for_health startup policy reusing the runtime healthcheck primitive.
• Coordinated shutdown on SIGINT and SIGTERM with a configurable graceful timeout before force-kill.
• Tracing spans per resource lifecycle.

Acceptance criteria

• A cycle in the dependency graph produces a clear error before any runtime call is made.
• Resources start in topological order and only after their dependencies are healthy.
• Ctrl+C from the CLI stops every resource in reverse topological order within the configured timeout.
• A failure during start triggers a clean teardown of the resources already running.