Probabilistic disconnections could make it quite hard to debug protocol
implementations and increases the risk of flaky behaviour in the wild
significantly. I don't see why a simpler solution isn't better.

The most likely failure mode of this is not an attack but the same as
previous breakages - scaling or legitimate version skew that causes problems
as the network evolves.

Agree with Luke that non-standard transactions should not be considered an
attack.

If you stay with the scoring system I'd be tempted to have a flag (defaults
to 100) that sets a minimum threshold for the badness scores and ignores any
below that. Attacks based on sending transactions that aren't syntactically
valid don't seem likely to me, this isn't a good way to DoS somebody because
discarding them is so cheap. If it turns out later there is a problem,
people under attack could flip the flag until a new version is released.

The formula for the DoS score in the case of invalid signatures/merkle roots
seem unnecessarily elaborate. An invalid signature should never occur and
could always result in immediate disconnection.

Treating a block with too many sigops as invalid means legitimate relayers
might be treated as an attacker if/when the constant changes in future. I'd
suggest not treating this as an attacking situation at all.

Why use a mutable field with a const setter?

Unit tests that rely on sleeps like this can be flaky because the OS delay
isn't always precise, not to mention slow/irritating to run. It's better if
tests can override the clock, eg, if GetTime() did something like

   if (nMockTime) { return nMockTime} else { ... }

then unit tests could reliably modify and advance the clock in a
fast/efficient manner.