Limits

Smaller losses in a round trip. A code block's trailing blank lines, a heading's leading space, the words after a fence's language (```ts title), and a line break inside a setext heading do not survive; a list item that is only a rule reads back as a rule; a row with fewer cells than the widest is padded. Each is reported by nothing yet.

A table's header. GFM's header is the first row, so that row is printed as the header whether or not the document marks one; a row marked as the header anywhere else is reported, because GFM cannot place it. A cell's blocks are inlined with a space between them.

Inline atoms. The model has no inline image or break, so an Image is a block and prints as its own line — which a parser reads back as a paragraph holding an image.

A link with no href prints as its label, with an UnsupportedMark diagnostic.

A mark on whitespace at a run's edge. **bold ** is not emphasis in CommonMark, so a marked run's leading and trailing whitespace is printed outside its delimiters: the text survives, and the space reads back unmarked.

A bare URL in text. GFM links https://…, www.…, and email addresses written as plain text, and no escape stops it, so such text reads back as a link.

A hard line break has no shape inside a block, so it ends the paragraph and the rest starts a new one, with a diagnostic. Raw HTML and link definitions are reported and skipped: a document holds neither. Footnotes are too, unless parse is given footnotes: true; a footnote label GFM cannot spell (one with a space) prints as text, reported.

No profile yet. Custom syntax needs both directions at once, so it arrives with the first application that declares one.