Avoid widening into unreducible types when inferring types

This is a general improvement, independent of named tuples.

Author: odersky

compiler/src/dotty/tools/dotc/core/TypeComparer.scala

@@ -2360,7 +2360,7 @@ class TypeComparer(@constructorOnly initctx: Context) extends ConstraintHandling
    *  @param canConstrain  If true, new constraints might be added to simplify the lub.
    *  @param isSoft        If the lub is a union, this determines whether it's a soft union.
    */
-  def lub(tp1: Type, tp2: Type, canConstrain: Boolean = false, isSoft: Boolean = true): Type = /*>|>*/ trace(s"lub(${tp1.show}, ${tp2.show}, canConstrain=$canConstrain, isSoft=$isSoft)", subtyping, show = true) /*<|<*/ {
+  def lub(tp1: Type, tp2: Type, canConstrain: Boolean = false, isSoft: Boolean = true): Type = trace(s"lub(${tp1.show}, ${tp2.show}, canConstrain=$canConstrain, isSoft=$isSoft)", subtyping, show = true) {
     if (tp1 eq tp2) tp1
     else if !tp1.exists || (tp2 eq WildcardType) then tp1
     else if !tp2.exists || (tp1 eq WildcardType) then tp2

compiler/src/dotty/tools/dotc/core/TypeOps.scala

@@ -390,7 +390,12 @@ object TypeOps:
         (tp.tp1.dealias, tp.tp2.dealias) match
           case (tp1 @ AppliedType(tycon1, args1), tp2 @ AppliedType(tycon2, args2))
           if tycon1.typeSymbol == tycon2.typeSymbol && (tycon1 =:= tycon2) =>
-            mergeRefinedOrApplied(tp1, tp2)
+            mergeRefinedOrApplied(tp1, tp2) match
+              case tp: AppliedType if tp.isUnreducibleWild =>
+                // fall back to or-dominators rather tahn inferring a type that would
+                // caue an unreducible type error later.
+                approximateOr(tp1, tp2)
+              case tp => tp
           case (tp1, tp2) =>
             approximateOr(tp1, tp2)
       case _ =>

tests/pos/named-tuple-widen.scala

@@ -0,0 +1,9 @@
+import language.experimental.namedTuples
+
+class A
+class B
+val y1: (a1: A, b1:  B) = ???
+val y2: (a2: A, b2: B) = ???
+var z1 = if ??? then y1 else y2 // -- what is the type of z2
+var z2: NamedTuple.AnyNamedTuple = z1
+val _ = z1 = z2