Very impressive result, Tom! Congratulations!
Now I believe 85's exist.
The problem is with these thousands of core-hours.
Do you have canonicalization procedure for pencilmark subgrid?
If no, IMO investing some effort there would reduce the times for vicinity search on the cost of huge files containing sorted subgrids that are already checked at {+n}.
Another idea is to generalize the unavoidable set concept to pencilmarks (i.e. representing a regular UA as a list of generalized UA, each corresponding to a particular permutation of givens). Then, with some luck, we may achieve full scan of a chosen solution grid using McGuire's method within reasonable core-hours. Both the whole search space, and, on the other hand, the sets requiring N constrains to be resolved (being composed of mutually disjoint generalized UA sets) grow enormously and I have no idea up to which degree they compensate each other. Also I don't know the reasonable minimum of constrains to start from, but for initial investigations it could be far less than 85.