TextMergeWithOffsets can merge text nodes that are not actually adjacent in the source document.

The default Parser parses an indented code block as a series of text events, one per line. TextMergeWithOffsets merges all of these into a single event; but the text for that event is borrowed and contains the extra indentation characters in between those nodes. (These are usually spaces, but I think they could even contain > characters if the code block is in a block quote.)

I think this must be a bug. It's surprising and hard to work around, and I don't think TextMergeStream behaves this way.

The code with the bug is here.