SERVER-65260 introduces a change to avoid IXSCAN for regex predicates if the index has collation arguing that bounds cannot be tightened and COLLSCAN performs better. We've discovered a number of scenarios in which the unbounded IXSCAN is better than COLLSCAN (e.g. HELP-60129) so we want to (see HELP-60129 for discussion and details):

1. stick to SERVER-65260's described change (ignore regex type)
2. setup a query knob to disable the behaviour for customers whose workloads experience regressions as a result of this.