This is showing you that oscillation has some kind of "hump"
to get over, to start, and the damping of the GEAR2 method
might be showing you what will happen if passives' Q or
actives' gain are lower than your models estimate.
PLL designers may include "sweeper" or "kicker" functions
to force their loop to closure. Bandgap designers always
(if they want to stay in the job) need startup circuits for
theirs. Maybe your oscillator should be no different. It's
not always practical to know whether a simulation method
outcomes discrepancy is really. solely about the numerical
goings-on, but you can know that you pounded a spike
through its neck, if you did.