[Tracking] datalake_fdw: Iceberg data path and the metadata agent

[MisterRaindrop]

Everything between the DDL skeleton (#1842) and a lake table that can be written and read from SQL. Design: #1683. This issue is the execution view.

Landed: #1842, the skeleton (AM, catalog/volume FDWs, stub engine, DDL guards); #1951, the Parquet format layer (Arrow, field ids, tracked memory pool, type rules).

Twelve PRs, three waves

 wave 1                           wave 2                            wave 3
 A  S3 storage ────────────────────────────────────────┐
 B0 proto ─┬► B1 agent framework ─┬► B3 Iceberg ops ─► B4 Builtin ─┼► C write ─┬► E merge-on-read + DML
           └► B2 engine client ───┘        ├► B5 Polaris            ├► D read ──┘
                                           ├► B6 Hive Metastore
                                           └► B7 Hadoop

C and D need A, B2 and B4; with the Builtin catalog they run in CI against MinIO and the agent alone. B5–B7 need B3 and run alongside C and D.

• A datalake_fdw: storage I/O over object storage (S3) #2009 Storage I/O over object storage (S3) — first
• B0 datalake_agent: the gRPC contract (proto files) #2010 The gRPC contract: proto files
• B1 datalake_agent: service framework (Java) #2011 datalake_agent: service framework (Java)
• B2 datalake_fdw: the agent engine client (meta/engine_agent) #2012 datalake_fdw: the agent engine client
• B3 datalake_agent: Iceberg operations by metadata location #2013 datalake_agent: Iceberg operations by metadata location
• B4 datalake_fdw: Builtin catalog #2014 datalake_fdw: Builtin catalog
• B5 datalake_agent: Polaris (REST) catalog #2015 datalake_agent: Polaris (REST) catalog
• B6 datalake_agent: Hive Metastore catalog #2016 datalake_agent: Hive Metastore catalog
• B7 datalake_agent: Hadoop (filesystem) catalog #2017 datalake_agent: Hadoop (filesystem) catalog
• C datalake_fdw: write path — INSERT to data files and a committed snapshot #2018 Write path: INSERT to data files and a committed snapshot
• D datalake_fdw: read path — planned fragments, CustomScan, projection and pruning #2019 Read path: planned fragments, CustomScan, projection and pruning
• E datalake_fdw: merge-on-read and DML #2020 Merge-on-read and DML

Attached to the PR they belong with: #1988 NUMERIC (C), #1990 timestamp units (C), #1989 name mapping (D). Later, blocking nothing: the optional datalake_proxy bgworker (#1683 §5.4).

Rules

• PRs target main and are squash-merged; the description is the commit message (datalake_fdw: read and write Parquet through Arrow #1951 layout).
• Every PR leaves both components shippable: an unfinished path refuses with a clear error. The agent answers UNIMPLEMENTED for RPCs it does not carry yet, as the stub engine does.
• The proto files and the shared headers (format/format.h, meta/iceberg_meta_engine.h, common/*.h) change only in their own small PR.
• Nothing generated is committed; protoc runs at build time on both sides.
• At most two PRs in review at once; the next stays a draft.
• Each PR brings its tests: a smoke category on the C side, mvn -B verify on the Java side, a CI job for any service it needs.
• Milestone datalake_fdw: Iceberg read/write MVP, label datalake.

Out of scope: the FDW raw-file path (#1683 §2.3), the CI build image (devops repo), HDFS storage, execution-engine work.