This behavior is not reported by anyone else. Please clarify the following.
1- What is the mechanism to measure 20-sec oscillator stop? - probing Xin/Xout pin or observing time lag in the RTC time update?
2- C1 (C65) value in the above schematic shows 18 pf, which is 6 pF higher than the datasheet recommended value (12 pF). If this mismatch exists on the PC board as well, then please change the C1 value to 12 pf and retest. C1 and C2 loads across the RTC crystal are very sensitive for clock accuracy and sustained oscillation.
3- If possible please capture all 16 RTC registers value when failing.