feat: Machine capability routing — K8s-style node affinity for agent workloads #599

Problem

Multi-machine squad deployments waste rounds when agents pick up work they can't complete (e.g., browser automation on a headless server, corp-network access from a cloud runner).

Proposal

K8s node affinity pattern for Squad:

Machine side: ~/.squad/machine-capabilities.json

{
  "machineId": "MY-MACHINE",
  "capabilities": ["browser", "personal-gh", "emu-gh", "teams-mcp"]
}

Workload side: GitHub issue labels with needs:* prefix (e.g., needs:browser, needs:kusto, needs:corp-network)

Matching: Before Ralph picks up any issue, check: does this machine have all required capabilities? If not, skip silently — another Ralph on a capable machine will handle it.

Multi-Ralph coordination: Claim issues via GitHub assignees before working. Branch names include machineId to prevent collisions. Heartbeat files for stale claim detection.

Prior Art

Running in production with 2 machines (local dev + DevBox). Successfully routes WhatsApp monitoring to local machine, Kusto queries to DevBox, and GitHub work to either.

Deliverables

• machine-capabilities.json schema + squad init auto-detection
• Test-MachineCapability function for ralph-watch
• Docs: multi-machine setup guide
• needs:* label convention in routing docs