Build a hash-chained audit log and a supervision sampler

Three small pure pieces do the work: canonical() turns a record into the exact bytes that are hashed, AuditLog appends records and walks each matter's chain, and sample() picks the supervisor's records. Everything else is synthetic data and demo scripts. The decision that matters most is what the hash covers: the record's own fields, its position in the matter (seq) and the previous record's hash, serialised with sorted keys. With that, any edit, deletion or reordering of an earlier record breaks verification, and no harmless change (key order) does.