Since it captures numMergedChunksPerRange by reference, if the transaction is ever retried, numMergedChunksPerRange would accumulate elements from previous executions, which later triggers this invariant. numMergedChunksPerRange should be recreated/cleared on every retry.

Retrying a transaction is not a rare occurrence, so hitting this invariant is quite likely. An invariant means a server crash, with the potential unavailability time it carries. There are no further consequences: no data loss or corruption, and no stickiness (just restarting the crashed node solves the issue, and the new primary is very unlikely to hit the same invariant again).