FWIW, I was just looking into weirdly missing unmounts on the Elipsa, and, yeah, the lazy unmount is really sweeping a lot of stuff under the rug, stuff that might actually be problematic.
As in, quite a few Nickel threads *still* have fds open'ed on (at least?) book files, which means the filesystem is never actually unmounted. Which is obviously how Really Bad Stuff Happens (TM).
I happened to be testing a CBZ here, and hwmon, QProcessManager, eventengine, QEvdevTouch, QTBearerThread, imageprovider and QtHTTPThread all had a couple of fds opened on the last book.
Sidebar: I haven't looked at this stuff in a *long* time, so, this may not be specific to the Elipsa, so this isn't excluding the possibility that the Elipsa's storage solution might just be more prone to this kind of badness

.
(for instance, I can almost reliably trash my ZSH history (which lives in onboard) if I *properly* soft-reboot the Elipsa with the shell open. That never happens anywhere else).
EDIT: This is fairly easy to double-check with ForceWifi and a real lsof binary. fds from MNT_DETACHED fs end up relocated to / (with the rest of their path, relative to the original mountpoint, preserved) in the listing.
(I happened to notice the missing unmounts because I track 'em in KFMon & NanoClock ^^).
EDIT²: On the upside, quickly tested with a notebook, a PDF, an ePub, a KePub, and a KePub w/ annotations, and I can only reproduce a consistent failure with a CBZ

.
EDIT³: Issue persists as long as a CBZ has been open'ed during that boot session, though :/.