The scheduler binds a pod A to machine M.
Then it tries pod B to machine M.
In trying this second binding, it uses stale information about A, in particular, that it is not yet bound to a machine.
Sometime later, it "learns" about A being bound to M via watch.

It should assume all bindings it makes happen so as to avoid conflicts.

It will be very common for the scheduler to miss hostPort conflicts once I remove the check.

Also, spreading will work poorly. It may put A and B on same machine.

(Note that optimism is bad for anti-conflicts, but we don't have those I think).