Test coverage: pg_upgrade across majors (data preservation)

[ChronicallyJD]

Gap

pg_upgrade has no test coverage anywhere in the suite (no suite, no grep hit). For a table access method that stores data in relation forks plus WAL-logged catalog metadata in pgcolumnar.* heap tables, pg_upgrade (link and copy mode, across majors) is the primary cross-major data-preservation path, and it is exactly the profile where AM + custom-catalog extensions break.

Why this blocks a public release

A user who upgrades their cluster and finds columnar tables unreadable or the extension un-loadable has lost data with no warning. This is acceptable in a caveated alpha only if "no major-version upgrade path is tested yet" is stated explicitly; it is not acceptable for general use.

Suggested acceptance criteria

• A suite that: builds+installs against major N and N+1, creates a USING pgcolumnar table with data and options on N, runs pg_upgrade (both --link and copy), and asserts on N+1 that the extension loads, the table is readable, row data matches, and options survive.
• Run it for at least one adjacent pair in the local matrix; document any modes not covered.

Found in a release-readiness review; part of a set of durability-boundary gaps (see sibling issues on format compatibility, physical replication, and platform coverage).

[jdatcmd]

Closed by #256 (2cf3b81). test/pg_upgrade.sh, verified 15/15 on three pairs: 17→18 copy, 17→18 link, 15→16 copy. Asserts the extension loads, the access method survives, content matches by hash, per-table options survive, the index answers the same range as before, and the table still accepts writes.

Two things worth carrying forward. PostgreSQL 18 changed the initdb default to enable data checksums, so pg_upgrade refuses a default-initdb 17→18 pair with old cluster does not use data checksums but the new one does — users hit that before this extension is even considered. And per-table options survive pg_upgrade but not pg_dump (#248, fixed in #258), because pg_upgrade migrates the catalog physically; that narrowed #248 to pg_dump emission rather than option storage.

Not in run_all_versions.sh, and excluded in harness_selftest, for the same reason run_san is: it needs two majors at once.