datalake_fdw: Builtin catalog

Start by reading the B2 and B3 dependencies, #2008, and #1683 §4, then inspect src/meta/engine_builtin/ and the engine interface documentation. Use MinIO and the agent to run the DDL and full conformance suite, including rollback and concurrent-writer cases. Done means Spark opens the table by path, rollback preserves HEAD, and only one concurrent writer commits.

Part of #2008. Letters (A, B0–B7, C, D, E) are the PRs listed there; this is B4.

Scope

The type-unset case of #1683 §4: a lake table with no catalog service, only object storage. The first catalog to work end to end, so C and D develop and test against MinIO and the agent alone.

• src/meta/engine_builtin/: each table's current metadata_location in an extension-owned table on the coordinator; commit is a conditional update under a row lock inside the transaction — atomic with the statement, and across tables.
• Metadata work delegated to B3 by location; the agent holds no state for these tables.
• Other engines open the table by path (HadoopCatalog / FileIO).
• The two commit regimes (builtin: atomic with the transaction; external catalogs: an in-doubt window) written into the engine interface docs.

Out of scope

Sharing a builtin table between two clusters.

Depends on

B2, B3.

Acceptance

• DDL and the full conformance suite green with only MinIO and the agent running; rollback leaves the HEAD unchanged.
• Spark opens the table by path and sees Cloudberry's metadata.
• Two concurrent writers: one commits, the other errors; no lost update.