Both Proc and Method respond_to?(:call) and respond_to? inline caches in TruffleRuby, so there is no need to check explicit types here.

Given it's so simple after that, I would skip the helper and have the logic inline raise unless ...

Hm, I thought that anyway kind_of? is a bit faster than respond_to? even if the method is inlined. Or it is not so significant?

I would think both compile down to a single check on the metaclass.
You can verify with IGV or https://github.com/Shopify/seafoam if you like, or just write a benchmark using benchmark-ips :)

require 'benchmark/ips'

VALUE = -> {}

Benchmark.ips do |x|
  x.report 'kind_of?' do
    VALUE.kind_of?(Proc)
  end
  
  x.report 'respond_to?' do
    VALUE.respond_to?(:call)
  end
end

ruby bench_kind_of_respond_to.rb
Warming up --------------------------------------
            kind_of?   415.314M i/100ms
         respond_to?   434.209M i/100ms
Calculating -------------------------------------
            kind_of?      4.246B (± 1.7%) i/s -     21.596B in   5.088181s
         respond_to?      4.224B (± 0.4%) i/s -     21.276B in   5.036533s

So they both seem to optimize away.

Also more checks would be bad if the value is actually not a Proc or Method, and different value classes have been seen, then those would just be extra needless checks on the fast path.

Thanks to the benchmarks! I have no more arguments - I pushed the version with respond_to? only

Given it's so simple after that, I would skip the helper and have the logic inline raise unless ...

I did it; however, in MRI we have 5 places with this check and maybe we'll return this later; I guess additional method is not affect anything, just readability