Fix GraphFlow state corruption on interrupt/resume

[veeceey]

Summary

Fixes #7043

When a GraphFlow workflow is interrupted during agent transitions, the state can become corrupted with work remaining but an empty ready queue. This causes resume attempts to fail with "Digraph execution is complete" even though agents haven't finished executing.

Problem

The root cause is that the ready queue population happens during message processing (in update_message_thread), but if the workflow is interrupted between when an agent completes and when the next agent is enqueued, the state is saved with:

• remaining shows work left to do
• ready queue is empty (no agents enqueued)
• Resume fails because _apply_termination_condition sees empty ready queue and terminates

Solution

Added state validation and repair logic:

• New _validate_and_repair_state() method detects corrupted states on load
• Checks if work remains but ready queue is empty (corrupted state indicator)
• Reconstructs ready queue by finding nodes that should be ready:
  • For "all" activation: nodes with remaining count == 0
  • For "any" activation: nodes that were previously enqueued
• Automatically called in load_state() to repair on every resume

Testing

Manual unit test confirms the fix works:

# Simulate corrupted state: A has run, B should be ready but queue is empty
manager._remaining['B']['B'] = 0  # B's dependency satisfied
manager._ready = deque([])  # Corrupted: empty despite B being ready

manager._validate_and_repair_state()

# After repair: B is correctly added to ready queue
assert 'B' in manager._ready  # ✓ Passes

This ensures workflows can always resume from saved state, even if interrupted mid-transition.

[veeceey]

All checks passing, DCO signed, ready for merge

[veeceey]

Friendly ping - any chance someone could take a look at this when they get a chance? Happy to make any changes if needed.