Validate assistant's troubleshooting hypotheses against actual system state before acting on them

Across multiple sessions involving Tailscale networking issues, the user consistently pushes the assistant to ground troubleshooting in empirical system state — fetching actual config files, checking real network interfaces, inspecting live tailnet/control state, and confirming port bindings — rather than relying on doc-fetching or theoretical reasoning. When doc-fetching fails repeatedly (wrong pages returned), pivot immediately to direct policy/config inspection: ask user to paste relevant sections of the actual policy file or check admin console directly (e.g., https://login.tailscale.com/admin/services). Pattern: prefer hypotheses confirmed by running commands and inspecting current state; treat doc-fetching as a fallback only when local investigation is exhausted.