With a persistent entity cache, this means the following problem is very easy to replicate:
1) Create node A1
2) Translate node A1 as A2
The database is correct for both nodes (both tnids pointing to the nid for A1), but the load cache still contains the untranslated version of A1. If you view the translations for A2 you will see both nodes. If you view the translations for A1 you will see only A1.
If A1 is flushed from the cache at this point, things will resolve themselves, as the database is still correct.
If you edit and save A1 again before the cache is flushed, the untranslated state of A1 will be saved to the database.