Might not be the issue here, but recommend to enable "use unbuffered I/O" on every mountpath of MagLibs in order to avoid FS Cache loss issues (especially on clusters when active node dies). Caching doesn't help when writing large data streams, so it's useless anyway for backup to my experience.
Also, avoid mounting SAN-based mount paths to an existing file system (if applicable). Instead, I'd prefer to mount to a drive letter as this will provide more transparancy and additionally get around some OS issues with that in 2008 (if used).
Appart from that, the only cases where I saw chunks disappearing were related to virus scanning. Although you said mount paths are excluded from scanning, it might be that the backup/writer process in RAM is not, so there might be a small chance that data is already been quarantined in RAM before it gets to the mount path or even to the writer. Do you use a virus scanner that allows for excluding active processes from scanning?