Self-healing codes: How stable neural populations can track continually reconfiguring neural representations