Maybe not "normal" but not unheard of. You can run into
numerical issues in high gain circuits, varying by integration
method.
TRAP has the problem of being underdamped / undamped.
If you look close enough there's always a tiny bit of
"sawtoothing" timepoint-to-timepoint. A minuscule error
that keeps repeating and reversing, can be gained up
by the circuit behavior to make significant error behaviors.
GEAR (variable-order) can be better or worse. I don't know
too much about GEAR. I used to use the default TRAPGEAR2
back in Cadence days, unless I saw kooky stuff like you
describe.
When stuff gets weird I like to fall back to Euler, which
is inherently damped (don't ask me why - math, mumble
mumble). The downside of Euler is that it can also damp
out real problem behaviors (like local oscillation / ringing).