Quoted-chain reconstruction misses branches that never hit your inbox

I rebuilt a long group email thread by reading the inbox copy and mining the quoted history nested inside the newest reply — cheap and it looked complete. It was not. A cluster of replies had been filtered to Junk and were only ever quoted within each other, forming a branch that no inbox-visible message referenced. Reconstructing from quotes can only ever recover what someone downstream happened to quote, so it silently under-reports and gives no signal that anything is missing. Always enumerate the spam or junk folder separately before declaring a thread fully extracted. Two mechanics that made the sweep cheap: rather than dumping each message and burning context, inject the set of already-known identifiers into a page script and return only links NOT in that set, which turns a huge extraction into a tiny diff; and note that virtualized mail lists recycle DOM nodes, so accessibility refs captured before opening a message go stale and index-based row clicking silently re-opens the wrong item — verify which message actually loaded by reading the sender out of the body rather than trusting the index you clicked. Also worth checking: the gap skewed heavily toward one link type, because the junked branch was the one thread carrying both, so a missing-data audit should ask which field the lost source uniquely supplied.