bundle full schema and constraint catalogs in partition snapshots

[fabracht]

Summary

Closes #57 and #58. After PR #56 surfaced "constraints don't reach all nodes uniformly" as a follow-up, this PR fixes the late-join half of that gap.

Schemas and constraints already broadcast to every alive node on write (replication_ops.rs:81-85 and :345-369 — likely added in #54 after the issues were filed). The remaining hole is the snapshot path: SchemaStore::export_for_partition and ConstraintStore::export_for_partition filter entry.partition() == partition, so a joining node only ever received entries whose schema_partition(entity) happened to hash into a partition it owned. Catalog entries under any other partition were unreachable through snapshots — a node could rebalance into the cluster missing most of the catalog, and a freshly-joined 4th node empirically held 0 constraints in the #57 trace.

Both stores now expose export_all(), which serializes every entry regardless of partition using the same wire format as export_for_partition (so import_schemas / import_constraints are unchanged). StoreManager::export_partition uses these for DB_SCHEMA and DB_CONSTRAINT. Catalogs are small in-memory state and import is idempotent overwrite, so re-applying on every partition snapshot a joining node fetches is safe.

A shared private serialize_entries helper in each store removes the duplicated count-prefixed loop between the per-partition and full-catalog exports.

Test plan

• cargo test -p mqdb-cluster --lib — 3 new tests pass:
  • schema_store::export_all_carries_every_schema_regardless_of_partition
  • constraint_store::export_all_carries_every_constraint_regardless_of_partition
  • partition_io::partition_snapshot_carries_full_schema_and_constraint_catalog — multi-entity setup spanning many partitions; asserts dst has the full catalog after a single-partition snapshot, plus full counts match source.
• partition_io::export_import_roundtrip_covers_all_db_stores updated — previously asserted per-partition isolation for schemas/constraints, now asserts they arrive regardless of partition.
• cargo make clippy — clean (pedantic, all targets + wasm).
• cargo make format-check — clean.
• cargo make test — full workspace clean.
• Pre-commit hook (format-check + clippy) ran on the commit and passed.
• examples/cluster-rebalance-stores/run.sh actually run twice with an enterprise license generated locally:
  • Pre-fix (per constraints not replicated uniformly across cluster nodes #57 trace): schemas/constraints visible only on a subset of nodes; FK cascade left up to 12 of 12 eligible children alive.
  • Post-fix run 1: 10/11 hard assertions pass; all 5 OBSERVATION lines OK (schema get posts/comments via node 4, all duplicates rejected, FK orphan rejected); FK cascade removed 17 of 19 eligible children. The single hard-assertion failure ("18 of 20 extra child comments") is a confirmed-flaky cluster-warmup transient unrelated to schema/constraint replication.
  • Post-fix run 2: 11/11 hard assertions pass; all 5 OBSERVATIONs OK; FK cascade removes all 21 of 21 eligible children via node 4.

Out of scope

Remaining FK-cascade survivors that still show up under hash-unlucky distributions are #22 (cluster cascade does not filter by ownership on remote nodes) — separate follow-up.