I would suggest bounding the number of abstract states of PDBs that are kept in memory. This is a lot more direct than (for example) bounding the number of patterns, as a pattern can have 2 states or millions of states.

In issue1007, Malte observed that when improving the runtime of the hill climbing algorithm, it could sometimes run out of memory instead of stopping before this happens. We want to investigate better ways of controlling its memory consumption. Possible ways of doing so:
- add a parameter max generated patterns (see msg10275)
- stop keeping all ever generated PDBs as candidates during hill climbing (see10272)
While we are at it, we could also make the time control more fine-grained.