atperson: learning-schema compatibility for replay

atperson 81aace8 — learning-schema compatibility for replay across algorithm changes (issue #6).

Every ledger entry records the learning schema it was observed under. Replay now consults a single compatibility predicate, atp_schema_can_replay(version), per entry; snapshots record the schema that produced their state in a new section 8 and load refuses foreign ones.

• ATP_ERR_SCHEMA = 6, distinct from ATP_ERR_FORMAT (corruption): the data is intact, the model generation is the mismatch.
• atp_replay_report.failed_at_id and .failed_schema name the exact entry where a mixed-schema ledger broke.
• Mixed-schema ledgers replay deterministically in id order and fail at the first unreplayable entry — old and new algorithms are never ambiguously mixed in one graph.
• Snapshot section 8 is forward-compatible: v5 readers written before section 8 skip unknown tags, and snapshots without it load as schema 1.
• The C++ layer surfaces actionable errors for both paths: replay names the entry and schema, load points at atperson rebuild or a new model generation.
• Fixtures cover compatible transitions (current-schema ledger replays through the table), incompatible transitions (foreign schema refused), mixed ledgers, and hand-rewritten section 8 bytes with a fixed digest.
• Docs: architecture.md gained a "Learning-schema compatibility" section with the three compatibility classes and review guidance — any PR touching tokenisation, sampling, memory selection, familiarity, or the training equations must include a schema-version decision.

20/20 tests on core, ASan, and network builds.