When chasing after something else a couple days ago I discovered that the existing hs26 test does not cover the cases where the initial on-disk values are globally visible. Since the RLE caching code does a global visibility check, this is a noticeable oversight. I've extended the test accordingly.
It turns out that it all works fine. The changeset also adds comments to the code that explain the critical point regarding this, on the grounds that since I wrote the code, if I missed this point a few months later it likely isn't obvious to anyone else either.