[Ethereum] How are Ethereum transaction costs calculated

Ethereum-based tokens are intrinsically bound to Ethereum. Any token transaction would be in the form of a valid Ethereum transaction, in a block on the Ethereum chain. (see footnote)

It's entirely possible, as you've seen, to reward miners in a token for including a token-related transaction. This would indeed probably require modifying mining software, although in a future where tokens are common this might become a default setting. Either way, a token miner would also be an Ethereum miner. There'd be no way to mine only the tokens--nor, really, would there be any reason when you could get 5+ ETH per mined block for no extra effort.

At the moment, token transactions are payed for in ETH, which any miner would accept. The problem with paying solely in tokens is that even if someone will include the transaction, no unmodified miner would. That's why, as far as I know, there are no token-based fees currently in common use.

Footnote: An Ethereum smart contract could host a sidechain, and then the sidechain would have its own kind of miner. That said, sidechains have less use in Ethereum, since you can write contracts to have most sidechain features.

I can reproduce this with compiler version 0.4.12+commit.194ff033.Emscripten.clang (but, in my case, the other way round - file1.sol has a deployment cost of 68990, file2.sol has a deployment cost of 68926.)

There is a 64 gas difference in the cost between the two deployments. The reason for this is that, although the deployment code is exactly the same length (82 bytes), one of the sets of bytecodes has a zero byte in a place where the other has a non-zero byte.

When transmitting a transaction to the blockchain, the gas cost per byte of calldata is 68 for non-zero bytes, and 4 for zero bytes (Appendix G of the Yellow paper, G_txdatanonzero vs. G_txdatazero). Hence the 68 - 4 = 64 gas difference between your codes.

The different byte actually occurs in the appended contract metadata section, not in the deployment code or the deployed code itself which are identical for both your examples as you would expect. The following are the metadata sections for the contracts - the metadata is generated automatically by the compiler and comprises a Swarm hash of the contents of the source file plus some other bits and pieces:

a165627a7a72305820176535c2c0a4a3ac4a26111bf3f9aa4ab0fa5db14d592f8574bf15f1f42f46420029
a165627a7a723058204e0a27613ab151031db65041c0e2ff84bf88842fd7b2cc00163bc9af9acd05230029
                          zero/non-zero byte difference is here ^^

So, with two near identical source files you have some chance of there being a difference in deployment cost due to this, and it is compiler version dependent since the compiler version is embedded in the data used to create the Swarm hash.