There was a bit of confusion between individual patterns and lists of
patterns, and index mismatches linked to that. This introduces a
vocabulary of "pattern-stacks" to provide a clearer mental model of what
is happening. This also adds examples.

This is equivalent to the previous code in terms of performance. The
expensive path is clearly identical. The fast path is also the same,
because in both cases we loop until we get a non-empty `refined_ctors`,
and then stop there. So the new code doesn't compute anything more
than the previous did.