Add pgoutput logical decoding plugin and make it the default

[teknogeek0]

Summary

Adds the pgoutput logical decoding plugin and makes it the default.

pgoutput ships with Postgres core, so CDC works without installing an
extension on the source server. It also sends a compact binary protocol: on a
mixed INSERT/UPDATE/DELETE workload it used 4.5x less network volume and about
4x less CPU on the source than wal2json.

test_decoding and wal2json remain supported via --plugin.

How it works

The receive step decodes the pgoutput binary protocol and writes a
wal2json-shaped JSON message, so the transform and apply steps are unchanged.

pgoutput sends only the tables of a publication. Without --publication,
pgcopydb builds one from the --filters table list and drops it during
pgcopydb stream cleanup, so the source server does the filtering.

Usage

pgcopydb clone --follow --plugin pgoutput --filters filters.ini

Multi-step migrations pass the filters at the snapshot step, where the
publication is created:

pgcopydb snapshot --follow --plugin pgoutput --filters filters.ini
pgcopydb stream setup
pgcopydb clone --filters filters.ini

New options

• --publication — use an existing publication instead of letting pgcopydb
  manage one
• --filters on pgcopydb snapshot

Tests

New suites cdc-pgoutput, cdc-filtering-pgoutput and follow-pgoutput, and
cdc-low-level now runs on pgoutput. All are in the CI matrix.

Coverage includes INSERT/UPDATE/DELETE/TRUNCATE, REPLICA IDENTITY DEFAULT and
FULL, unchanged TOAST values, multi-relation TRUNCATE, and both include-only
and exclude filters. Data fidelity is checked with source/target checksums over
float8 and float4 (including denormals, Infinity and NaN, and float8 used as a
WHERE-clause key), bigint and numeric ranges, text needing escaping, quoted
identifiers with embedded double quotes, and uuid, date, time, interval, inet,
arrays and jsonb.

tests/ci-parity.sh runs the CI suite list and Postgres versions locally.

Full matrix: 33 suites on PostgreSQL 16, 17 and 18.

[robdaly93]

Reviewed from the migration/field side, and ran the pgoutput path against PostgreSQL 18.6 locally.

Approach looks right to me. Decoding pgoutput into the existing wal2json JSON shape so transform and apply stay untouched keeps the change well contained, and emitting values as strings rather than JSON numbers is actually a fidelity improvement over the wal2json path for numeric. The 'K' vs 'O' identity handling and the synthesized old-key case when no old tuple is sent both look correct to me.

I spent most of my time on the new publication behavior, since pgcopydb performing DDL on the source is new. Noting what I confirmed, for anyone reading this later:

• A table in the publication with no replica identity gets UPDATE/DELETE rejected on the source: ERROR: cannot update table "..." because it does not have a replica identity and publishes updates. This is the prerequisite in the "Replica Identity and lack of Primary Keys" docs section doing its job — non-PK tables need REPLICA IDENTITY FULL or USING INDEX for CDC under any plugin.
• A --filters-scoped publication only covers the filtered tables, so excluded tables are unaffected.
• pgcopydb stream cleanup correctly drops the publication it created and leaves a user-supplied --publication alone.

One small non-blocking thing you may want to look at: the publication builder applies exclude-schema and exclude-table, but not exclude-table-data. So a table excluded only from the data copy still joins the publication if someone used that on a large non-PK log table they'd hit the error above. Might be worth aligning the filter handling or adding a doc note.

LGTM.