But in ideal world we don't want to have partially executed scripts. The
text in 3028bis is describing real world and is much more liberal than I
would like it to be.
For debugging of a Sieve script it is much better to know in which state
the saved message is. So using the original message (before any
partially applied header changes) seems to be the best.
I agree with this analysis. Our implementation could have handled the
attachment of header information in the event of an error either way without
too much trouble, but I think what the draft describes is the correct approach.
Ned