Because the message often has almost no content of its own. "What about software?" contains no mention of refunds or time windows, so a search engine matching on meaning or on words has nothing useful to match. The meaning lived in the previous turn, which the search engine never saw. A human fills that gap automatically, which is why the bug is easy to miss when you test by hand.

Because older turns are usually about different topics, and the rewriter cannot always tell which parts still matter. Given a long history it tends to blend old subjects into the new question, producing something that matches nothing. Recent turns hold everything needed to resolve "it" or "that one", so more history adds cost and noise without adding accuracy.

The message is already complete, so it should go to search unchanged, and the earlier history should be dropped. This is a deliberate topic change. Rewriting it against the previous subject would attach context the user has explicitly abandoned, and the search would then look for office hours mixed with refunds. Detecting these resets is as important as doing the rewriting.

Because users refer back to earlier answers, and by then those chunks have left the context window. If the model is asked "where did that come from?" at turn 8 about something said at turn 2, it has nothing to work from and will invent a source. Storing the sources used with each answer means the reference can be looked up rather than guessed.