None-lambda doesn't mean that it never checks its arguments.
And if it's a lambda, it doesn't match the arity value.

As for the implementation detail, there is no room for arity in ifuncs.

JRuby implements Hash#to_proc as:

class Hash
  def to_proc
    method(:[]).to_proc
  end
end

This allows us to present the proc as a lambda with correct arity:

$ jruby -e "pr = {}.to_proc; puts pr.arity; puts pr.lambda?"
1
true

It works for MRI too:

$ ruby23 -e "class Hash; def to_proc; method(:[]).to_proc; end; end; pr = {}.to_proc; puts pr.arity; puts pr.lambda?"
1
true

I think this is more representative of this proc's behavior. Can MRI do it this way?

This ticket was briefly discussed at today's dev meeting, and matz said that the result of Hash#to_proc should be a lambda.

Looking at the commit, should Hash#to_proc use rb_func_lambda_new() instead of rb_func_proc_new()? (and then rb_func_proc_new is unused)
Also it seems confusing that rb_func_proc_new() creates a lambda, so using rb_func_lambda_new() seems better here.
Code changed quite a bit though around there.

Proposed fix (simple as it gets): https://.com/ruby/ruby/pull/3370