Hm, so far I have actually never heard of tasks just vanishing from the BOINC manager - that's really strange.
Anyhow, we use Virtualbox to allow for the writing of checkpoints - which, unfortunately, the original science app (developed by mathematicians for HPCs only) does not support. We were actually the first DC team (which operates multiple own DC projects) suggesting a VM approach for DC apps which can't natively checkpoint at the 2009 Barcelona BOINC Workshop. The LHC@home team at CERN presented their first implementation of this at that same workshop (after us presenting the proposal, however). After years of practice and optimization based on user reports, to me it appears that at present the only error remaining to occasionally occurr in RNA World with these VERY DEMANDING tasks (which you can disable in your settings in case computing a year for a single WU is too ambitious) is that Virtualbox does not keep up with the snapshot writing to disk before a system shuts down. For this reason, we recommend to shut down BOINC, wait until your HD does not appear to write anything relevant anymore and only then shut down your machine.
In any case, after failure, the task should not vanish from your system but "just" restart from scratch (which is anoying enough given the exorbitant runtimes).
A solution to the "premature shutdown issue" could be the keeping of at least TWO snapshots such that in case of the bug described above, the older snapshot could be used to resume work. But this is something that is not in our but in Oracle's hand, I fear.
Unfortunately, the breaking up of these monster tasks into smaller parts is not possible with our current approach. So, I would recommend to just disable the VM tasks in your RNA World settings for now.
Michael.
P.S.: Thanks for reporting the issue to us.
