sounds good! i'll need to find some time to look into it. want to run a community build on it early?

@lrytz thanks, before late bedtime I almost gave you a heads up to ask if you'd like me to continue the PR.

Before spending electrons on CB, I'll try some projects locally. I expect to discover more behaviors.

Edit: the cycle will need fixing first, I hit it right away. I'll leave the PR open out of sheer optimism.

Maybe there's nothing better to do than catch & ignore.

This is current behavior, regular type check fails in "assign type to tree" because asking if it's a refinement (typedDefDef) means checking all overrides. I see dotty has a <refinement> artifact. My bigger idea was that lookupSymbol need not force in lookupInPrefix, since it just needs to search parents for the name to detect that there is a symbol with that name; actually getting the (possibly overloaded) symbol is not necessary when probing. typedIdent tries a few strategies; I hoped gentler probing might help with cycles, but maybe wishful thinking.

[log typer] collect local implicits List(primary constructor $anon, type T, getter tTag, value tTag )
[log typer] collect local implicits List(<$anon: <empty>.this.Extractors>)
[log typer] collect local implicits List(type skolem S&0, value a, value evidence$1)
[log typer] collect local implicits List()
[log typer] collect local implicits List(primary constructor C, method extractorMatch)
huh.scala:20: tTag is not a valid implicit value for reflect.this.ClassTag[$anon.this.T] because:
illegal cyclic reference involving getter tTag
    (new Extractors { type T = S; val tTag = classTag[T] })(a)
                                                     ^

I haven't looked at what problem is solved by "pre-filtering". typedImplicit already says we're checking shadowing to reduce ambiguity, and typedIdent is basically just the lookup. It's nice to think that filtering is efficient, but not at the cost of correctness; and maybe it had more benefit in ancient days.