The obvious thing to be suspicious of I suppose is the implementation of GroupCompressVersionedFiles.insert_record_stream, and its interaction with _GCGraphIndex.add_records (i.e. perhaps essentially the same bug as bug 437626, but in the groupcompress code rather than the knit code). At a glance I don't see any obvious flaw, but it's also not obviously correct to me either.
The obvious thing to be suspicious of I suppose is the implementation of GroupCompressVe rsionedFiles. insert_ record_ stream, and its interaction with _GCGraphIndex. add_records (i.e. perhaps essentially the same bug as bug 437626, but in the groupcompress code rather than the knit code). At a glance I don't see any obvious flaw, but it's also not obviously correct to me either.