Add provisioner target control and graceful agent deprovisioning

- Add IProvisionerStateProvider interface for decoupling orchestrator
  HTTP layer from provisioner implementations
- HordeProvisioner implements the interface directly, exposing target,
  active, estimated core counts and per-agent provisioning status
- Add orchestrator dashboard provisioner panel with editable target
  core count, active/estimated/agents/draining metric tiles
- Add per-agent status badges in agents table (active/draining/deallocated)
  with dimming for non-provisioned or deallocated workers
- Add GET/POST /orch/provisioner/status and /orch/provisioner/target
  endpoints; include provisioner stats in WebSocket push
- Add session lifecycle HTTP endpoints on compute nodes:
  GET /compute/session/status, POST /compute/session/drain,
  POST /compute/session/sunset (with shutdown callback)
- Implement graceful deprovisioning in HordeProvisioner:
  - Best-fit agent selection: largest agent that fits the remaining
    excess, breaking ties by workload (least busy first)
  - Never deprovision below target core count
  - Drain via HTTP: signal drain, poll status, send sunset on completion
  - Configurable grace period (--horde-drain-grace-period, default 300s)
  - Anti-oscillation: scale-up requires full agent-sized gap below target;
    estimated cores adjusted at drain-mark time, not agent-exit time
- Fix JSON-to-CbObject number parsing: integral JSON numbers are now
  stored as integers instead of Float64 so AsInt32/AsInt64 work
- Add effective URL to curl HTTP client error logging
- Add WorkerAnnotator callback to OrchestratorService::GetWorkerList
  for per-worker provisioner status injection
- Remove standalone compute/orchestrator.html and compute/compute.html;
  redirect /dashboard/compute/ to SPA

0 files changed, 0 insertions, 0 deletions