In your scenario, you are using

  66% ..3..........2....1.3.....................2......2.X
  33% ..3..........2....1........1...1...................X

but that's closer to a 50%-50% split of the miners, not a 66%-33%.

The problem is that 2 and 3 will produce the double of blocks. For e while 1 may tie or get ahead, but after 30 blocks the the 23 fork will almost always be like 10 blocks ahead of the only 1 fork.

Ignoring most of the . to make the graph shorter

  66% 3213...223.3223..23.23.323..3.2223
  33% 321.11....1....11..1..1...11.1....

or looking at each chain

  66% 32132233223232332332223
  33% 3211111111111

After a while, 1 can

A) Give up and start mining censored blocks

B) Start a new fork, from the current 2-3 chain with many censored blocks

C) Declare that they are a fork and convince the exchanges to give them a new moniker, like ETC of BXC

D) Convince the exchanges that they are the real chain in spite they are shorter, and convince the exchanges to give the other chain a new moniker.