More fun with native RLA on an SQL view -- at the MI level

This is getting downright bizarre. (And it's completely separate from what I posted on MIDRANGE-L, even to coming from two different beta sites; the only things the two incidents have in common is beta test code to handle record filtering in an SQL view.)

I have this sequence of MI declarations and instructions:

> . . .
> CPYBLA .DMIOSW, *ON; /* INDICATE I/O OPERATION */
> CPYBLA CPF5147, 'N';
> DCL EXCM BADSQLEX EXCID(H'5147') BP(BADSQL) CV('CPF) SKP;
> MODEXCPD BADSQLEX, HANDLE, OP2ISCON;
> CALLX .IOEPTR, .OL02008, *;
> B AFBADSQL;
> BADSQL:
> CPYBLA CPF5147, 'Y';
> CPYBLA .F02XSTS, 0; /* RESET ERROR? */
> AFBADSQL:
> . . .

"CPF5147" is a flag I set up, to indicate that the CPF5147 exception was caught by the new BADSQLEX EXCM I recently added, and the operation being done in this particular call to .IOEPTR is a READ, in this case from an SQL view. And much of what's going on in this section of code was there before I'd even heard of MI.

It works just fine, on our V6 box, on "dummy" data that the customer sent us. It also works just fine, on the customer's V6 box, on their "live" data (which they're legally prohibited from sharing with us, and which we don't want to see, anyway). But when the customer uses it on the exact same dummy data that they packed up and sent us, the BADSQLEX EXCM gets ignored, and just before the application crashes, we get an MCH1825 followed by an uncaught CPF5147, with the latter thrown by the very .IOEPTR call that BADSQLEX is supposed to be monitoring.

Any ideas why an EXCM would develop a proverbial "hole in its glove"?

And I've got a request out for the beta tester to do a <HELP> and an <F9> on the MCH1825, in hopes that it might tell me something.

--
JHHL