v2.4.1

Fixed

• Re-embedding to a custom provider no longer breaks the store.
  iai-mcp migrate --reembed-to-configured-provider rebuilt the records table
  from a schema that dropped the primary-key constraint, so records captured
  afterward stored a null internal label and the daemon crash-looped on the next
  restart with memory reported unreachable. The migration now preserves the
  table's canonical constraints; a null label left by an affected earlier run is
  repaired automatically on startup; and a genuinely null label now fails with a
  clear, actionable error instead of an opaque one. If you hit this on 2.4.0 or
  2.3.1, upgrading and restarting is enough — no manual repair needed.
• doctor no longer reports a false failure while the daemon is winding down.
  Check (e) rejected the TRANSITIONING state the daemon legitimately writes on
  the way to drowsy, showing a spurious failed check on a healthy install. It now
  accepts that state.