A second adversarial review found that the case-operation vectors added in the previous commit fired without ever reaching the consequence they were justified by. All 11 changes were tier 4 to tier 2 -- high confidence to high confidence, same entity. The boundary that matters is the one between a high-confidence tier, which resolves silently, and everything below it, which stops and asks the owner. No vector crossed it. They all resolved through the slug tier, because a candidate's id is the slug of its own name and transliteration already erases the distinction. Giving a candidate an opaque id removes that path and leaves the case operation alone deciding between no match at all and a high-confidence match. Three such vectors now cross the boundary; measured by replaying all 65 vectors against a build with the substitution applied. Eight of the previously added vectors were inert. Three are kept, relabelled as controls with the reason: Cherokee case agrees under both operations so it cannot diverge here at all, Turkish dotted I is neutral outside the Turkic profile, and Greek FINAL sigma is neutral where the medial form is not. They were not a spelling error, as the review supposed -- checked character by character -- but a reader who assumed they covered Cherokee would have been wrong, which is why they now say so. Tiers 7 and 8 had zero vectors and survived only inside inline tests, so a move that hollowed those out would keep every count and drop two tiers. Both now have vectors; all eight tiers are covered. Also corrects the divergence fixtures' own prose. The backfill obligation is 62 assigned codepoints, not the 21 letters -- the rest are C1 controls, tone bars, a currency sign and a ligature, and scoping from the letter count misses two thirds of it. And both files now say what to do when they redden: rebuild both implementations and rewrite wholesale, never hand-edit one entry to make a digest agree.