It's not behaving like a saturating amplifier; it's folding back down
to zero output as the light input increases. And, apparently, much
sooner than would be expected from the data sheet.
The pfet current mirror model explains the effect, but not why it does
it at low output currents.
It's not the output current - as such - that's the problem, but the
light level.
The P-channel MOSFET current mirror model explains the effect, but
without data on the specific P-channel MOSFET involved it can't say
anything about the current level where the tail MOSFET switches from
being a resistor to being a constant current source.
All we know is that the P-channel MOSFET's in the production batch
from which Marco Trapanese bought his parts switched at a rather lower
current than the batch the manufacturer tested to generate the data-
sheet.
Whoever wrote the data sheet has got egg all over their face, but
that's human error, and it happens from time to time.