The join between a block document and the chart it compiled to (ADR-0004 decision 7).

Document identity hashes what an author wrote; chart identity hashes what the compiler generated; decision 6's determinism is the function between them:

document bytes + palette + compiler version  --compile-->  SCXML bytes
        |                                                       |
     sha256                                                  sha256
        v                                                       v
 document identity                                      chart identity

Neither is derived from the other, and neither should be. Deriving chart identity from the document hash would be hashing the wrong bytes: two documents differing only in metadata generate identical SCXML and must get the same chart identity, or a metadata edit would break resume for every running session. Deriving document identity from chart identity is not even a function - the mapping is many-to-one in exactly that case.

Hashes do not invert, so the compiler emits the join as a fact. The record is the artifact's primary key: given a running session, which names a chart identity, look up by chart_identity and get back the document, the revision, and (once sb-qz0 lands it) the provenance map that explains it.

chart_name carries the document id and chart_version stays nil

Statifier.Machine.Identity.matches?/2 is struct equality across all three fields, and st-ADR-0060's resume refuses on {:identity_mismatch, expected, actual}. So putting the document revision in chart_version would break resume on every save, and putting the document hash there would break it on a metadata-only edit - precisely the case decision 6 established as a non-event. The document id is constant across revisions and is safe; the revision is not, and lives in this record instead, where nothing compares it.

palette_hash is a hygiene aid, not a commitment

It digests the sorted {type_name, module, current_version} triples of the entries the compile actually resolved. It is not a cryptographic commitment to those modules' behaviour - nothing short of hashing compiled beam would be - and this module does not pretend otherwise. It exists so the common cause of a surprising recompile, a host adding or swapping a palette entry, is visible in the record instead of invisible. A host that changes an emit/2 without bumping current_version/0 has moved a compile input without moving the record; that is a palette-hygiene obligation on the host, stated here so it is not discovered later.