I think you mean "Thompson NFA construction" and "NFA->DFA."

Regardless though, this is not what the OP is pointing out. 'egrep' (or just GNU grep these days) is doing something more clever (emphasis mine):

> Al Aho expected his deterministic regular-expression recognizer would beat Ken's classic nondeterministic recognizer. Unfortunately, for single-shot use on complex regular expressions, Ken's could finish while egrep was still busy building a deterministic automaton. To finally gain the prize, Al sidestepped the curse of the automaton's exponentially big state table by inventing a way to build on the fly only the table entries that are actually visited during recognition.

Russ Cox talks about this a bit in part 3 of his articles on regex matching[1]. Its implementation in RE2 is here: https://github.com/google/re2/blob/master/re2/dfa.cc

[1] - https://swtch.com/~rsc/regexp/regexp3.html

Yep, only noticed it later, then left it in to see who's paying attention :)

I know a number of projects that generate their roff by using pandoc. They don't actually know, or have the inclination to learn, exactly how g/roff works.

My (very recent) university education was unfortunately quite light on UNIX folklore, but this was converted in our formal automata course as we traversed the Chomsky hierarchy.