According to my analysis, I think the main reasons of the consensus failure of fuxi-3000 are as follows:

• Due the there is not enough voting power at the start of the network, the network has been stucked for a long time. So the double sign evidence has been broadcasted to almost all nodes. Once voting power is enough, then the evidenve will be included in the first block.
• Then each node attempts to apply this block. In beginBlock, the evidence will invoke slashing. However, in this situation, infractionHeight is 1 and ctx.BlockHeight() is 0. Then panic error happened.

Besides, when a validator is new added and it launchs double sign attack in the first block after it become a real validator. Then the double sign evidence happen to be included by the block proposal, then in the beginBlock, we will try to slash a validator which has no sign info. However, the sign info is required for figuring out the jail period. This panic error is just for checking the existence of validator sign info.
But this situation is possible and reasonable, we can't just panic, instead we should check the sign info of a validator before slashing it.