As secondary indexes are stored individually on each node what you're suggesting sounds exactly like a consistency issue. the fact that you read 0 cells on one query implies the node that got the query did not have any data for the row. The reason you would sometimes see different behaviours is likely because of read repairs. The fact that the repair guides the issue pretty much guarantees it's a consistency issue.
You should check for dropped mutations in tpstats/logs and if they are occurring try and stop that from happening (probably load related). You could also try performing reads and writes at LOCAL_QUORUM for stronger consistency, however note this has a performance/latency impact.
