Comment 1 for bug 415508

The associated branch has a potential fix.

Basically, stop paying attention to 'content_summary[...]' to be accurate about the committed text size.

This means that for files with content filtering, we will read them off disk, apply the filtering, and then see that nothing has changed.

I expect performance will suck a little bit, as we'll be re-reading all the files that have a content filter. But at least it won't cause bogus data to be inserted into the repository on every merge commit.