Unles the verification routine (eg. garage) keeps some state somewhere it has no way to prevent replays.

i did not say replay was impossible. i said the statement in the article was overly broad, which it is. If it said "a single result could be hard-coded", it would be fine. but it said the function is constant, which it need not be.

But even that need not be true. here is how you could design that dongle to require no persistent state on RX and still not allow simple replay.

TX code:

   static u8 counter = 0;
   tmp = randU12() * 16 + (counter++);
   tmp2 = sha256(secret + tmp) & 0xffff;
   send32((tmp << 16) + tmp2)

RX code: static u16 prevSeenCodes[16]; static u8 idx = 0;

   tmp = recv32();
   tmp2 = tmp & 0xffff;
   tmp >>= 16;
   if (sha256(secret + tmp) & 0xffff != tmp2)
      fail();
   if (tmp in prevSeenCodes)
      fail();
   prevSeenCodes[15 & idx++] = tmp;

now you need to replay at least a sequence of 16 codes

that array is in ram and need not persist across program runs. dongle can be powered off or can be left on. all will work