Fold one piece's slice sequence into Markov transition counts over every CONSISTENT interpretation path.

A cadence is a movement onto a target chord, so an edge is built from a slice TRIPLE (s1, s2, s3): the cadence a -> b followed by the cadence b -> c, where the middle interpretation b is shared by both sides. Each such edge accumulates weight w_a * w_b * w_c from the (per-slice normalised) interpretation weights, so the total mass contributed by each slice triple is exactly 1.

Sharing the middle reading is what keeps the expansion honest: every reasonable reading of an ambiguous vertical informs the model (recovering movement that triad reduction trims away), while two alternative readings of the SAME moment can never be counted as a transition between moments. Genuine corpus pedals (roughly a quarter of YCACL slices repeat their predecessor) produce self-edges via the same triple rule as every other transition.

Pieces with fewer than three slices contribute no edges (no complete triple exists).

Sum buildTransitionCounts over a composer's pieces. Each piece is counted independently, so no transition can be invented across the boundary between the last slices of one piece and the first of the next (the flattened pipeline leaked exactly such edges).

Weighted triad interpretations of one slice, most consonant first. The top three candidates (by Hindemith dissonance, Harmonic.Evaluation.Scoring.Dissonance) carry a [3,2,1] preference profile, NORMALISED so every slice's interpretation weights sum to 1 — an ambiguous 7-PC vertical contributes exactly the same total probability mass as an unambiguous triad, instead of shouting several times louder purely because it was ambiguous. Ties keep possibleTriads enumeration order (sortOn is stable).

Falls back to a single flat triad at weight 1 if ranking yields no candidates (e.g. the exporter filtered the slice below 3 unique PCs).