As shown by the tools/regen-transformTests.sh script, every time we reparse a large page (like [[en:Barack_Obama]]) we generate different ID values for the about attributes.
This seems to be because the token pipeline is async and thus non-deterministic, and we don't have a final DOM post-pass to ensure a deterministic numbering.
See also T87556: Thoughts on element IDs, sections, incremental parsing and fast section editing.