Weight policy to stack/spread instances and "max_placement_results"

Can you give a more detailed explanation/example of this:

"with a number bellow to the existing resources,"

By resources do you mean number of hosts? Or resources as in inventory, e.g. number of available VCPUs?

As for:

"the weight policy will only be applied to the subset of allocation candidates retrieved by placement."

This has always been the case - the weights are applied to the hosts after those hosts are filtered, so even if you had 1000 hosts in a deployment, maybe only 10 apply to a given flavor/image and then those 10 hosts are weighed. Placement is just doing initial filtering.

What am I missing?

This is working as designed. The underlying issues that require CERN to set max_placement_results to such a low number (10 when there are ~200 hosts per cell, and ~14K hosts total in the deployment) are what we need to focus on, like bug 1805984 for example.