Well, another day working on smudge filters, or unlocked files as the feature will be known when it's ready. Got both git annex get and git annex drop working for these files today.

Get was the easy part; it just has to hard link or copy the object to the work tree file(s) that point to it.

Handling dropping was hard. If the user drops a file, but it's unlocked and modified, it shouldn't reset it to the pointer file. For this, I reused the InodeCache stuff that was built for direct mode. So the sqlite database tracks the InodeCaches of unlocked files, and when a key is dropped it can check if the file is modified.

But that's not a complete solution, because when git uses a clean filter, it will write the file itself, and git-annex won't have an InodeCache for it. To handle this case, git-annex will fall back to verifying the content of the file when dropping it if its InodeCache isn't known. Bit of a shame to need an expensive checksum to drop an unlocked file; maybe the git clean filter interface will eventually be improved to let git-annex use it more efficiently.

Anyway, smudged aka unlocked files are working now well enough to be a proof of concept. I have several missing safety checks that need to be added to get the implementation to be really correct, and quite a lot of polishing still to do, including making unlock, lock, fsck, and merge handle them, and finishing repository upgrade code.