Thanks everybody for your suggestions.

It looks like the concensus is the reset circuit.

My design uses a CR network but this feeds a HC14 Schmitt trigger inverter which then drives the microcontroller RST pin. The RST signal rises with the power supply and is a clean pulse of 50ms duration. The power supply rises in less than 10ms.

Still, I think it will be worthwhile grafting a proper supervisor chip onto one of the failing boards
to see what effect this has.

Two other pieces of information. (1) All the failing devices have a date code of 0110 (2001 week 10). (2) When the problem occurs we have observed that the ALE output signal is sometimes at half its normal frequency. This suggests to me that the device has gone into 12 clock mode.

Regards,
Paul Griffith