[assistant-ngalaiko]

Cause

The latest master workflow failed in Wait for the node, not in flake checks or
the Linux image build:
https://github.com/ngalaiko/computer/actions/runs/40002731857

Every probe received a pong through DERP(fra), then exited with direct connection not established. The default tailscale ping --until-direct=true made the loop
require direct UDP connectivity and time out after 30 minutes, despite a reachable
peer. Activation never ran.

Fix

Set --until-direct=false on the existing one-probe readiness command. Accept
either direct or relayed reachability while retaining retries and the existing
timeout when the peer is genuinely unreachable. The SSH activation step still
enforces its own connection/authentication checks.

This changes only the reusable deployment workflow. No Dasha account, environment,
credentials, services, permissions, or runtime configuration is modified.

Verification

• Reproduced the old failure with the installed Tailscale CLI against an isolated
  LocalAPI Unix socket: relay pong => exit 1, direct => exit 0, unreachable => exit 1.
• Reran the same scenarios with the exact updated workflow command: relay and
  direct => exit 0; unreachable => exit 1. Every scenario sent exactly one probe.
• Checked the action-selected Tailscale v1.82.0 source: the flag is supported and
  its default/direct-only failure matches the observed CI logs. Local CLI tests
  use installed Tailscale v1.98.10; they simulate peer responses, not a live CI link.
• nix flake check --all-systems --no-update-lock-file passed.
• Explicit checks.x86_64-linux.formatting build passed; formatter made no changes.
• Both live runner PIDs and start times stayed unchanged.

No manual deployment or CI rerun was triggered. Post-merge activation remains a
separate workflow outcome and is not claimed by these local checks.