Add auto-provisioning and TcpServer integration to TrackerManager

[jonnyspicer]

Summary

Wire the TcpServer into TrackerManager, add auto-provisioning of tracker instances for unknown nodes, and handle disconnect/reconnect lifecycle.

Context

Depends on #5 (DetectionSource protocol) and #6 (TcpServer + TcpReceiver).

See design spec: docs/superpowers/specs/2026-03-11-tcp-push-detection-forwarding-design.md in the retina monorepo.

Changes

Updated: TrackerManager (tracker_host/manager.py)

• Start TcpServer on init if tcp_server.enabled in config
• For mode: "tcp" trackers in config: create TcpReceiver, register with TcpServer, create TrackerInstance with that receiver
• For mode: "http" trackers: existing behavior (create DetectionFetcher-backed instance)
• Auto-provisioning callback: register a callback with TcpServer for unknown node_ids. When a frame arrives from an unknown node with a valid token:
  • Auto-assign a local TCP port from auto_provision_port_start to auto_provision_port_end
  • Create a TcpReceiver and register with TcpServer
  • Create a TrackerInstance with default settings
  • Start the instance (spawns retina-tracker subprocess)
  • Log the auto-provisioning event
• Reject auto-provisioning when port range is exhausted (log error)

Disconnect/reconnect lifecycle

• On node disconnect: TcpReceiver.is_healthy becomes False. TrackerInstance keeps subprocess alive for extended_outage_sec (existing behavior via fetch loop).
• On reconnect: TcpServer routes to existing TcpReceiver. Node's buffered frames drain through. Subprocess continues processing.
• After extended_outage_sec: subprocess killed, instance stays registered in memory.
• Late reconnect: subprocess respawned automatically when frames arrive again.

Graceful shutdown

• Stop accepting new TCP connections
• Close all node connections (nodes reconnect automatically)
• Drain in-progress frames
• Stop all tracker subprocesses (existing TrackerManager.stop() behavior)

Updated: config parsing

• Parse tcp_server block into TcpServerConfig
• Updated config.yaml with tcp_server section

Acceptance criteria

• TcpServer starts when enabled in config
• TCP-mode trackers in config get TcpReceiver-backed instances
• HTTP-mode trackers continue working unchanged
• Unknown nodes with valid tokens get auto-provisioned
• Auto-provisioned nodes get ports from configured range
• Port exhaustion logged and handled gracefully
• Reconnecting nodes reattach to existing instances
• Subprocess stays alive during grace period
• Graceful shutdown closes connections and drains frames
• Unit tests for auto-provisioning and lifecycle