That's exactly how you get today's abysmally slow software though, because if you chain enough components you end up running into 1% cases all the time.

(In that specific case, with that specific very skewed data distribution, yes. But not necessarily in the general case. So "they" would be better off with that solution, but not a generalized "you").

Then how do they learn

My guess is that they would be more likely to “fix” and “improve” Ken’s algorithm than understand it (Jon Bentley mentioned, in passing, a similar case where this was the outcome.)

Oh! "Educating the newbies/grunts" is the reason for #$#@&#?

I'd say that an org has a limited number of pennies to spend on non-trivial code. Let's spend them wisely.

They can learn finite automata from computer science text books.

True, i suppose, but theres simply no substitute for contextual, on the job learning. imo