Since the light client work introduced in v0.33 it appears full nodes
are no longer fully verifying commit signatures during block execution -
they stop after +2/3. See in VerifyCommit:
https://github.com/tendermint/tendermint/blob/0c7fd316eb006c0afc13996c00ac8bde1078b32c/types/validator_set.go#L700-L703

This means proposers can propose blocks that contain valid +2/3
signatures and then the rest of the signatures can be whatever they
want. They can claim that all the other validators signed just by
including a CommitSig with arbitrary signature data. While this doesn't
seem to impact safety of Tendermint per se, it means that Commits may
contain a lot of invalid data. This is already true of blocks, since
they can include invalid txs filled with garbage, but in that case the
application knows they they are invalid and can punish the proposer. But
since applications dont verify commit signatures directly (they trust
tendermint to do that), they won't be able to detect it.

This can impact incentivization logic in the application that depends on
the LastCommitInfo sent in BeginBlock, which includes which validators
signed. For instance, Gaia incentivizes proposers with a bonus for
including more than +2/3 of the signatures. But a proposer can now claim
that bonus just by including arbitrary data for the final -1/3 of
validators without actually waiting for their signatures. There may be
other tricks that can be played because of this.

In general, the full node should be a fully verifying machine. While
it's true that the light client can avoid verifying all signatures by
stopping after +2/3, the full node can not. Thus the light client and
full node should use distinct VerifyCommit functions if one is going to
stop after +2/3 or otherwise perform less validation (for instance light
clients can also skip verifying votes for nil while full nodes can not).

See a commit with a bad signature that verifies here: 56367fd. From what
I can tell, Tendermint will go on to think this commit is valid and
forward this data to the app, so the app will think the second validator
actually signed when it clearly did not.

Closes #4926

The dump consensus state had this:

      "last_commit": {
        "votes": [
          "Vote{0:04CBBF43CA3E 385085/00/2(Precommit) 1B73DA9FC4C8 42C97B86D89D @ 2020-05-27T06:46:51.042392895Z}",
          "Vote{1:055799E028FA 385085/00/2(Precommit) 652B08AD61EA 0D507D7FA3AB @ 2020-06-28T04:57:29.20793209Z}",
          "Vote{2:056024CFA910 385085/00/2(Precommit) 652B08AD61EA C8E95532A4C3 @ 2020-06-28T04:57:29.452696998Z}",
          "Vote{3:0741C95814DA 385085/00/2(Precommit) 652B08AD61EA 36D567615F7C @ 2020-06-28T04:57:29.279788593Z}",

Note there's a precommit in there from the first val from May (2020-05-27) while the rest are from today (2020-06-28). It suggests there's a validator from an old instance of the network at this height (they're using the same chain-id!). Obviously a single bad validator shouldn't be an issue. But the Commit refactor work introduced a bug.

When we propose a block, we get the block.LastCommit by calling MakeCommit on the set of precommits we saw for the last height. This set may include precommits for a different block, and hence the block.LastCommit we propose may include precommits that aren't actually for the last block (but of course +2/3 will be). Before v0.33, we just skipped over these precommits during verification. But in v0.33, we expect all signatures for a blockID to be for the same block ID! Thus we end up proposing a block that we can't verify.