Adding an XBee carrier board raised the 3V3 rail from about 36 uF to 47 uF, and a slower-collapsing rail gave a pull-up longer to act at power-off, so worst-case latch margin fell to about 0.03 V. One board change ate the other board margin, and the fix is a 1 uF cap on the same pads, checked by a sim script at the harder case.