Reinstantiate restriction to transparent inline methods

[odersky]

Reverts parts of #19922.

Fixes #20342
Fixes #20297

The logic that we should ignore declared result types of inline methods really only applies to transparent inlines.

[odersky]

It seems this breaks more in the CB. Not sure what to do, I think I'll give this up for now.

[noti0na1]

I added a test for #20342, and it passes locally as well.
So, LGTM now.

[odersky]

@WojciechMazur So, this test fixes two regressions and probably opens two others that were fixed by #19922. Nevertheless, from a point of the program logic this is the right thing to do. Should we merge this (which would revert to the 3.4 and LTS behavior) or test the open CB before that?

[WojciechMazur]

I've run the OpenCB, it does in fact reintroduce the 2 regressions mentioned above. Seems to also introduce a new regression in softwaremill/tapir (although it released a new version in meantime, so the sources are slightly different). I can try to provide a minimization of this new problem.

Project	Version	Build URL	Notes
kalin-rudnicki/harness	4.1.3 -> 4.2.5	Open CB logs	#19415
softwaremill/tapir	1.10.0 -> 1.10.6	Open CB logs	Cannot find a codec between types: String and T - exists already in main #20420
apache/pekko-connectors	1.0.2	Open CB logs	#19479

I can't really tell should it be merged or not, as it seems like in each some projects would fail to compile.

[WojciechMazur]

Reproduction for softwaremill/tapir:
Edit: This issue is already failing on main, see: #20420

final class StrictEqual[V]
final class Less[V]
type LessEqual[V] = Less[V] | StrictEqual[V]

object TapirCodecIron:
  trait ValidatorForPredicate[Value, Predicate]
  trait PrimitiveValidatorForPredicate[Value, Predicate]
      extends ValidatorForPredicate[Value, Predicate]

  given validatorForLessEqual[N: Numeric, NM <: N](using
      ValueOf[NM]
  ): PrimitiveValidatorForPredicate[N, LessEqual[NM]] = ???
  given validatorForDescribedOr[N, P](using
      IsDescription[P]
  ): ValidatorForPredicate[N, P] = ???

  trait IsDescription[A]
  object IsDescription:
    given derived[A]: IsDescription[A] = ???

@main def Test = {
  import TapirCodecIron.{*, given}
  type IntConstraint = LessEqual[3]
  summon[ValidatorForPredicate[Int, IntConstraint]]
}

Without -source flag it compiles without any warning
Under -source:3.5 it yields warning:

-- Warning: /Users/wmazur/projects/sandbox/tapir.repro.scala:24:51 -------------
24 |  summon[ValidatorForPredicate[Int, IntConstraint]]
   |                                                   ^
   |Given search preference for TapirCodecIron.ValidatorForPredicate[Int, IntConstraint] between alternatives (TapirCodecIron.validatorForDescribedOr :
   |  [N, P]
   |    (using x$1: TapirCodecIron.IsDescription[P]):
   |      TapirCodecIron.ValidatorForPredicate[N, P]
   |) and (TapirCodecIron.validatorForLessEqual :
   |  [N, NM <: N]
   |    (implicit evidence$1: Numeric[N], x$1: ValueOf[NM]):
   |      TapirCodecIron.PrimitiveValidatorForPredicate[N, LessEqual[NM]]
   |) will change
   |Current choice           : the second alternative
   |New choice from Scala 3.6: none - it's ambiguous
1 warning found

Under 3.5-migration it fails with compilation errror:

-- [E172] Type Error: /Users/wmazur/projects/sandbox/tapir.repro.scala:24:51 ---
24 |  summon[ValidatorForPredicate[Int, IntConstraint]]
   |                                                   ^
   |Ambiguous given instances: both given instance validatorForDescribedOr in object TapirCodecIron and given instance validatorForLessEqual in object TapirCodecIron match type TapirCodecIron.ValidatorForPredicate[Int, IntConstraint] of parameter x of method summon in object Predef

Seems like the 3.5-migration flag is treated as the 3.6

[SethTisue]

If the broken projects need only add a transparent here and there, that’s plausibly a reasonable thing to ask, but I think we'd need the project maintainers to confirm that it's an acceptable change, without unwanted other effects, in the context of their codebases.

[odersky]

@SethTisue One of the failing tests (#19415) could be fixed in its minimization by refining the inference scheme. The other is new code (#19479), using inline givens and could trivially be changed by adding a transparent. So I think we are good to merge.