How does ZFS Block Level Deduplication fit with Variable Block Size?

ZFS deduplication works on blocks (recordlength) it does not know/care about files. Each block is checksummed using sha256 (by default changeable). If the checksum matches an other block it will just reference the same record and no new data will be written. One problem of deduplication with ZFS is that checksums are kept in memory so large pools will require a lot of memory. So you should only apply reduplication when using large record length

Assuming recordlength 128k

If I a randomly filled file of 1GB, then I write a second file that is the same except half way through, I change one of the bytes. Will that file be deduplicated (all except for the changed byte's block?)

Yes only one block will not be duplicated.

If I write a single byte file, will it take a whole 128 kilobytes? If not, will the blocks get larger in the event the file gets longer?

128k will be allocated, if the file size grows above 128k more blocks will be allocated as needed.

If a file takes two 64kilobyte blocks (would this ever happen?), then would an identical file get deduped after taking a single 128 kilobyte block

A file will take 128k the same file will be deduplicated

If a file is shortened, then part of its block would have been ignored, perhaps the data would not be reset to 0x00 bytes. Would a half used block get deduced?

If the exact same block is found yes

The variable sized blocks of ZFS are as Yavor mentioned already should not be confused randomized variable sized chunking, also called content-defined chunking or Rabin fingerprinting. Here is a small talk describing the differences.

ZFS used static, but configurable block sizes.