building upon #3714, this (almost) implements the second half of #3302.
The main effect is that we now get a better error message when `rfl`
fails. For
```lean
example : n+1+m = n + (1+m) := by rfl
```
instead of the wall of text
```
The rfl tactic failed. Possible reasons:
- The goal is not a reflexive relation (neither `=` nor a relation with a @[refl] lemma).
- The arguments of the relation are not equal.
Try using the reflexivity lemma for your relation explicitly, e.g. `exact Eq.refl _` or
`exact HEq.rfl` etc.
n m : Nat
⊢ n + 1 + m = n + (1 + m)
```
we now get
```
error: tactic 'rfl' failed, the left-hand side
n + 1 + m
is not definitionally equal to the right-hand side
n + (1 + m)
n m : Nat
⊢ n + 1 + m = n + (1 + m)
```
Unfortunately, because of very subtle differences in semantics (which
transparency setting is used when reducing the goal and whether the
“implicit lambda” feature applies) I could not make this simply the only
`rfl` implementation. So `rfl` remains a macro and is still expanded to
`eq_refl` (difference transparency setting) and `exact Iff.rfl` and
`exact HEq.rfl` (implicit lambda) to not break existing code. This can
be revised later, so this still closes: #3302.
A user might still be puzzled *why* to terms are not defeq. Explaining
that better (“reduced to… and reduces to… etc.”) would also be great,
but that’s not specific to `rfl`, so better left for some other time.
Sebastian mentioned that the use of the kernel defeq was to work around
a performance issue that was fixed since. Let's see if we can do
without.
This is also a semantic change: Ground terms (no free vars, no mvars)
are reduced at
“all” transparency even if the the transparency setting is default. This
was the case
even before 03f6b87647 switched to the
kernel defeq
checking for performance. It seems that this is rather surprising
behavior from the user
point of view. The fallout on batteries and mathlib is rather limited,
only a few
`rfl` proofs seem to have (inadvertently or not) have relied on this.
The speedcenter reports no significant regressions on core or mathlib.
When the `decide` tactic fails, it can try to give hints about the
failure:
- It tells you which `Decidable` instances it unfolded, by making use of
the diagnostics feature.
- If it encounters `Eq.rec`, it gives you a hint that one of these
instances was likely defined using tactics.
- If it encounters `Classical.choice`, it hints that you might have
classical instances in scope.
- During this, it tries to process `Decidable.rec`s and matchers to pin
blame on a particular instance that failed to reduce.
This idea comes from discussion with Heather Macbeth [on
Zulip](https://leanprover.zulipchat.com/#narrow/stream/270676-lean4/topic/Decidable.20with.20structures/near/449409870).
