Unfortunately a DECODE has the same drawback as a CASE... the assumption that only one correction has to be made to a value. In my example from above,
DMGD INVTY MATL has 3 corrections to be applied at once. With the decode, only the first one in the list would be applied.
It looks like it's definitely a buffer issue. I tried sorting the input by the length of the field, so the shortest records would be modified first and longest last. It was fairly effective, except in the few cases where the modification shortened the record. Then the next record had an extra unwanted character added at the end.