So this is definitely a different issue than I ran into.
What I can already tell you is that you need to add an IF statement to cover the cases where the text is deleted. Put a check for IF <WA_S_STXL>-CLUSTD is not initial before the conversion part starts.
The issue here is that for Operation = D (Delete), only the key fields are filled at the time the include executes. CLUSTD is not a key field and therefore is initial and the decompression fails/causes the short dump. I need to update the original post when I get time.
Anyhow - in the short dumps it may show you a clue on which record it failed on. Even the debug procedure I mention won't help you there. If you suspect that the Japanese characters are the issue, once you locate a "problem" record and then attempt to use the READ_TEXT FM in the source system or even write a program that does the same conversion and see if you can simulate the failure.
Not sure if you can use the language field in STXL to help you narrow down that search and prove your hypothesis.
Regards,
Justin