Adds a proof-of-concept desired-state convergence mechanism for
harmony_agent: the central platform writes a DesiredStateConfig
to NATS KV at desired-state.<agent-id>; the agent watches the key,
executes the command via sh -c, and writes the result to
actual-state.<agent-id>.
- New module: harmony_agent/src/desired_state.rs (~150 lines)
- New types: DeploymentConfig::DesiredState, DesiredStateConfig,
ActualState, ExecutionStatus
- ADR: docs/adr/021-agent-desired-state-convergence.md
This is an initial happy path for review. The architecture
needs further evaluation against the mini-kubelet vision before
hardening (security, sandboxing, signing, resource limits).
Jean-Gabriel Gill-Couture
9cd1713788