javac apparently does some partial checking of lambda bodies in nested invocations during overload resolution, and uses the results to consider some methods inapplicable.  This is incorrect -- the nested lambda body should not influence applicability at all.

<T> T apply(UnaryOperator<T> op, T arg) { return op.apply(arg); }

class B { B x() { return this; } }
class C {}

void m(B arg) { System.out.println("B"); }
void m(C arg) { System.out.println("C"); }

m(apply(arg -> arg.x(), null)); // ambiguous, as expected
m(apply(arg -> new B(), null)); // expected: ambiguous; actual: chooses m(B)

(I believe this is a holdover from earlier overload resolution prototypes, which were designed to make use of partial information in the lambda body.)