When dealing with unschedulable pods, the intent was to deallocate only claims
which are allocated and use delayed allocation. That if check wasn't handled
correctly, causing also claims with immediate allocation to be considered as
candidates.
Found during code reading, probably has never occurred in practice yet.