Fix #24824 FunProtoTyped preservation in wildApprox and map to avoid compiler crash

[dancewithheart]

Fixes #24824: preserve FunProtoTyped in wildApprox and FunProtoTyped.map to avoid assert from collectParts that leading to a crash

How much have you relied on LLM-based tools in this contribution?

moderately: I feed LLM with relevant issues, PRs; reiterated over proposed solutions,
then switched to work done by @Alex1005a as he had more mature solution

How was the solution tested?

Tested locally added reproduction case:

sbt --client "testCompilation 24824"

and nearby cases (that failed during earlier iterations):

sbt --client "testCompilation i12456"
sbt --client "testCompilation i24553"
sbt --client "testCompilation i9328"
sbt --client "testCompilation i16116"
sbt --client "testCompilation i9299"

Not all tests are passing locally:

sbt --error "testOnly dotty.tools.dotc.CompilationTests"

but errors seems to be environment-related (e.g. subprocess interruptions or JDK-related issues).
I hope CI will clarify this.

Additional notes

My first attempt was to restrict when assert in collectParts is triggered

scala3/compiler/src/dotty/tools/dotc/typer/Implicits.scala

Lines 684 to 685 in 7c54aad

	assert(!ctx.typerState.constraint.contains(t), i"`wildApprox` failed to remove uninstantiated $t")
	traverse(t.underlying)

by:

case t: TypeParamRef =>
  def isTrivial(bounds: TypeBounds) =
    bounds.lo.isRef(defn.NothingClass) && bounds.hi.isRef(defn.AnyClass)

  ctx.typerState.constraint.entry(t) match
    case bounds: TypeBounds if ctx.typerState.constraint.contains(t) && !isTrivial(bounds) =>
      assert(false, i"`wildApprox` failed to remove uninstantiated $t")
    case _ =>
      traverse(t.underlying)

It worked for repro sample.

Similar approach was proposed @andrzejressel in #17305 (comment).
Following review remark oh his PR: #17471 (comment) I was investigating why we end up in collectparts in assert.

I tried modifying other parts of wildApprox and related pattern matches,
without luck.

From debugging (it can change the behavior of compiler!) it looked like issue is not in a match cases,
but in how FunProtoTyped is handled in error-recovery paths, where wildApprox reconstructs protos.
Preserving FunProtoTyped avoids breaking invariants about already-typed arguments.

Since the same code was changed #23020 I switched to extract changes around FunProtoTyped as Alex1005a proposed in #23020#issuecomment-3993039055 to check if it works.

[dancewithheart]

Hi @Alex1005a, I would like to give you credit - your work on FunProtoTyped handling was very insightful. I arrived at a similar direction after many failed experiments, and in the end I used your changes - kudos 🥇 for this 🙂

I cherry-picked these commits:

• b287832 — add map and fold for FunProtoTyped
• 48194df — remove funproto optimization from wildApprox

and used them in my PR.

I’m planning to move the PR from draft to ready for review so it can hopefully be merged - if that’s OK with you 🙂