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.