Use more context for implicit search only if no default argument

[odersky]

After an "implicit not found", we type additional arguments to get more context which might
give a larger implicit scope to search. With this commit we do that only if there is no
default argument for the implicit.

Based on #23659

This might fix #23610

[WojciechMazur]

Both erikerlandson/coulomb and zio-archive/zio-connect are still failing in the OpenCB with these changes with the same errors as before, but maybe resolving issue mentioned in #23659 (comment) can hopefully fix it.

[odersky]

I think there's nothing more I can do to make coulomb compile.

[odersky]

@som-snytt Since you recently worked in that corner, do you want to take a look for a review?

[EugeneFlesselle]

The changes seem reasonable to minimize impact from 45f7ef6 generally speaking.

For coulomb, I'm wondering if we should try returning with the deepened proto for AmbiguousImplicits even if constrainResult fails when applied to it? I opened #23687 to try it out.

[som-snytt]

LGTM! as they say blithely.

With the caveat that I'm only partly up to speed, I wonder if tightening constraints would work better when the named application is subsequently typed. When defaults are tried, you have the opportunity for additional shenanigans (better constraints for certain failed args that canProfit) possibly without over-constraining successful args (or whatever causes problems).

I'm suspicious of an inference decision depending on whether I have a default arg, since people are already confused about that competition between givens and defaults.

I don't know anything about the mechanics of constraints; and I haven't yet understood concretely what the ctx.typerState.resetTo(saved) does right before the named application is tried to jog the defaults. I know removing it makes the test on my other ticket work and nothing breaks. (In that same area of ignorance, I don't know what ignored proto means or why removing the branch at the end of adaptNoArgsImplicitMethod doesn't break its test.)

On the Pekko ticket, minimal test is at https://github.com/scala/scala3/pull/23556/files and only involves a couple of givens. At bottom, inference breaks when the default is used, where the context bound is in the same param list.

implicit final def marshaller[A: Encoder](implicit printer: Printer = Printer.noSpaces): ToEntityMarshaller[A] =

TIL Pekko is a "Tower of Hanoi" of implicit conversions; with some debug, maybe I'll see what it's doing when the typer reset makes it forget what A is.

[som-snytt]

I cloned coulomb this morning; I'll take another look if it stays broken in a way that is not terribly urgent.

[som-snytt]

Just on the default arg check, I expressed dubiousness. But not regressing is also nice. I will have different opinions when I'm less ignorant.

[odersky]

TIL Pekko is a "Tower of Hanoi" of implicit conversions; with some debug, maybe I'll see what it's doing when the typer reset makes it forget what A is.

The original change in #23532 does change where implicit conversions are inserted. Any propagation of info from the result into a typing does that since it implicitly assumes that the info is not invalidated by an implicit conversion. So I am not terribly surprised that Pekko is affected.

[odersky]

@som-snytt, @odersky Can one of you approve, then we can merge this.