Investigate honoring proseWrap, and whether wrap-refusal tokenization can shrink the hazard surface

Prettier's markdown plugin defaults to proseWrap preserve because rewrapping prose in markup is semantic; we reflow unconditionally because our admissibility machinery is what makes reflow safe. Two questions worth a bounded investigation:

1. Can we honor proseWrap as users expect (preserve = keep the author's breaks, never = join to one line, always = today's behavior) without a new hazard class? preserve and never both look easier than always: preserve keeps the source partition (always admissible by the AST invariants), and never is a single join whose seams are the only hazards.

2. Does thinking of hazardous shapes as unwrappable TOKENS reduce complexity? A continuation plus, a bare attribute-reference line, a callout marker: if the atom stream glued these to their neighbours up front (noBreakBefore/noBreakAfter at atom construction), some emission-time checks might become unnecessary because the packer could never produce the hazardous line in the first place. This overlaps with the constraint-query architecture work; the investigation should measure whether token-level gluing deletes checks or merely relocates them.

Expected outcome is a short written answer with measurements, not necessarily code. Low expectation of large wins; worth knowing.

Trace the existing proseWrap handling, atom construction, emission-time hazard checks, and constraint-query architecture. Measure whether preserve, never, and always can be supported safely, and whether token-level gluing removes checks or only relocates them; document the findings and measurements as the done condition.