use global recoding vectors (dragoon/komodo!20)

this MR combines the encoding and the recoding processes into a single one, expressed as the shard linear combination.

this allows to not recompute the encoding matrix each time, whether it's explicit during decoding or implicit during shard verification.