Two-marker setup: anchor closer to persona mimicry first

The two-marker chunk setup in #281 (train persona1 on <A> answer <B>, train persona2 on <A> alone, check whether persona2 emits <B>) currently sits further from the persona-mimicry baseline than it could. Before drawing conclusions from the no-transfer result, anchor the experimental design closer to the standard persona-mimicry setup so we can tell whether the absence of transfer is real or an artifact of the divergent recipe.

Concretely: identify what's different from the persona-mimicry baseline in #281's training and inference setup (data format, loss masking, marker tokenization, end-of-sentence handling), and run a sweep that progressively moves the two-marker design toward persona-mimicry while preserving the chunk structure. If transfer appears under a persona-mimicry-adjacent setup, the original no-transfer claim was design-dependent.

Source comment (mentor update, 2026-05-11):

Make more similar to persona mimicry first

From mentor update on #281 — Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start.