The problem was that manualy reexcuting the query, didn’t show the messed up plan, it took the right plan !!!
During two days I tried to reproduce the issue but i couldn’t.
i tried it again with some random values when the load was higher
and now could in a consistent way reproduce it.
By comparing the CBO traces is saw that the cpu_count was different .
this caused the optimizer to take the wrong plan when more than 11 cpu’s were assigned….
Case closed 😉