The only place I could find lparam being used as a signed value is in SBM_SETRANGE. However it is also cast to int, so even if the parameter was not sign extended in the wow64 thunk, it gets sign extended there by the LPARAM -> int conversion.
I think it's probably safe to do this, and the issues we need to be checking for regressions are cases where lparam is used as a signed 64bit value on the unix side, but it's probably unlikely.