fix: improve checkpoint resume UX with VM state guidance

Show the next step and prompt user to verify VNC matches expected
state before resuming. Default changed to No since fresh start is
the safe choice — resume is only valid after tunnel drops, not
VM reboots.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: abrichr

scripts/record_waa_demos.py

@@ -1678,17 +1678,24 @@ def _hard_reset_task_env() -> bytes:
             print("  Waiting for screen to stabilize...")
             return _wait_for_stable_screen(server)
 
-        # Check for existing checkpoint (resume support)
+        # Check for existing checkpoint (resume support).
+        # Resume is only meaningful if the VM state is intact (e.g. after a
+        # tunnel drop or script crash).  If the VM was rebooted, resuming
+        # would put the recording out of sync with the actual screen.
         checkpoint = _load_checkpoint(task_dir)
         resuming = False
         if checkpoint is not None:
             n_done = len(checkpoint["completed_steps"])
             n_left = len(checkpoint["remaining_steps"])
             ts = checkpoint.get("timestamp", "unknown")
+            next_step = checkpoint["remaining_steps"][0] if checkpoint["remaining_steps"] else "(done)"
             print(f"\n  Checkpoint found: {n_done} step(s) completed, "
                   f"{n_left} remaining (saved {ts})")
-            answer = input("  Resume from checkpoint? [Y/n] ").strip().lower()
-            if answer in ("", "y", "yes"):
+            print(f"  Next step would be: {next_step}")
+            print(f"\n  Check VNC ({vnc_url}) — does the VM match where you left off?")
+            print(f"  Resume is only valid if the VM was NOT rebooted since the checkpoint.")
+            answer = input("  Resume from checkpoint? [y/N] ").strip().lower()
+            if answer in ("y", "yes"):
                 resuming = True
 
         if resuming: