Not enough information - ie you have more that you could have given but didn't.
MOSFET IRF9321 datasheet here
Reference diode LM285 datasheet here
Darlington PMBTA13 datasheet here
I'll assume that you are testing this with Vin = 4V2 (as per diagram).
You don't say and you should as if I am correct about the problem then it may work OK on eg 5V.
What is Vin when testing?
What value is Vbat?
Your diagram has committed an Olin fail, with text over-writing important information.
Bigger and still munted version here
R30/R31 are 178k/100k to make a 2.78:1 divider.
If output is meant to be 3V3 that implies Q11 and Q12 bases should be 3.3/2.78 = 1.187V = 1.2V.
Closest regulator = 1.235V.
As you are using 1% resistors you could have got nominally if not actually closer than that. No problem though.
Fig 10 page 4 shows that gate threshold voltage increases with falling temperature.
The graph shows 2V at 0 C.
That's probably 'typical'.
Page 2 shows Vgsth as -1.3 / -1.8 / -2.4 min / typ / max, so you MIGHTY expect it to need as much as about -2.6V to get 50 uA.
The darlington transistors in the long tailed can saturate at 1 Vbe minimum.
The datasheet says Vsat = 1.5V but the current is "rather higher" than you are using.
A single transistor can saturate with Vce approaching zero, and this notionally applies to the upper transistor in the darlington pair. However, as it steals current from the lower transistor base as it saturate you won't get upper collector all the way down to lower collector. So Vsat will be 1 Vbe + some (probably :-) ).
You lose some headroom in the drop across R32 in the "tail" but as it's voltage is 1.235 - 2 x Vbe it is not much voltage - I'd have thought it was almost dangerously low as 1.235 - 2 x Vbe is about ill conditioned.
Overall you appear to have close to zero headroom at best, and a small increase in Vgsth could drive you against one limit. Whether you have any depends on the factors above - what is Vgsth actual?, what does the darlington saturate at actually?, does the long tailed pair behave if V_R32 is very low (and how low is it?).
What ... ?
Easy check - operate with Vin = say 6V.
If it works OK with temperature at 265 K, reduce Vin until it faults.
Report.
IF headroom is the problem, changing the darlingtons for single bipolars will give you at least a Vbe more at the loss of gain.