You call it a constant gm reference but are criticizing the
raw current value.
What can you say about the gm (of which species?) over
the corners set?
What makes you think that only the three simpleminded-est
"digital" corners are what an analog function ought to be
criticized, across?
Have you tried at all to debug this for yourself, such as by
looking at the difference between TT, TS, ST (is it the
PMOS, or the NMOS? Or is it neither, but rather that the
(say) NWell or whatever you use for the current setting
element, a "fellow traveler" to one or the other FETs? Do
you have any idea just what changes (or, is modeled to
change) between the corner cases?
Could be simply headroom related, with more current
in feedback to establish the same loop closure. Check that
by supply corners, I'd say.