Add test/build: extension-metadata tests + raw SQL syntax checks

[jnasbyupgrade]

Rebase note

Rebased onto sync-pgxntool-2.3.0 (bottom of a 4-branch stack restructuring this PR together with #22/#23; coordinator is handling gh stack link separately). Two things fell out of the rebase:

• Removed the now-redundant explicit make install workaround this branch's CI added before make verify-results -- pgxntool 2.3.0 (the new base) fixes the underlying installcheck-before-install ordering bug upstream (Postgres-Extensions/pgxntool-test#62), so verify-results's own dependency chain now guarantees the ordering itself. Re-reproduced the original bug on this rebased branch before removing the workaround (cleared /usr/share/postgresql/17/extension/'s test_factory files, ran the simplified step alone) to confirm the fix holds standalone.
• Regenerated test/build/expected/syntax.out: the new base (built on current upstream master) already includes upstream PR Drop the original_role temp table from install (#12/#13) #20 ("Drop the original_role temp table from install"), which this branch's original fork-based master predated. That commit shifts every line of the auto-generated sql/test_factory--0.5.0.sql by a constant offset, and syntax.sql's expected output embeds literal \i line numbers. Confirmed the diff was a pure line-number shift, not a content regression, before regenerating from a real run.

---

Summary

pgxntool (vendored at 2.2.0, now 2.3.0 via sync-pgxntool-2.3.0) has had a PGXNTOOL_ENABLE_TEST_BUILD feature for a while, but this repo never had a test/build directory. Quoting pgxntool/base.mk's own purpose comment for that feature:

Validates that extension SQL files are syntactically correct by running files from test/build/ through pg_regress. This provides better error messages than CREATE EXTENSION failures.

This PR addresses a separate, pre-existing gap; it's now the second branch in the stack (base: sync-pgxntool-2.3.0), independent in content from #22/#23.

1. Move install-mechanics tests into test/build/install.sql

test/sql/install.sql and the opening block of test/sql/pgtap.sql tested extension packaging (dependency declarations, clean install/uninstall) rather than test_factory's actual business logic (registering/getting test data). That's exactly what test/build is for: plain SQL + pg_regress's classic expected-output diffing, in an isolated database with none of the main suite's pgTAP/role/schedule machinery.

test/sql/install.sql is deleted (nothing left to test there once this moved), and the leading dependency-check block in test/sql/pgtap.sql is trimmed (the rest of that file — the part that actually exercises tf.tap() — is untouched).

2. Add test/build/syntax.sql: run the raw versioned SQL files directly

Per the repo owner's request: "add simple tests there that just try to run the .sql files. Such tests make it MUCH easier to debug something like a syntax problem in the code, since CREATE EXTENSION tends to completely obscure the real error." This \i's sql/test_factory--0.5.0.sql and sql/test_factory_pgtap--0.1.0.sql directly (not via CREATE EXTENSION).

Known/expected errors baked into its expected output (documented in the file itself, so a future reader isn't confused about which diffs are "expected" vs "something's actually broken"):

• Two pg_extension_config_dump() can only be called from an SQL script executed by CREATE EXTENSION errors — that function always errors when its containing script is \i'd directly instead of run via CREATE/ALTER EXTENSION, regardless of whether the SQL is otherwise valid.
• A SET ROLE "" / "zero-length delimited identifier" error near the end of each file — both files save the caller's role via pg_catalog.set_config(..., true) (transaction-scoped) so they can restore it at the end. A real CREATE EXTENSION runs its whole script as one transaction, so that works; run via plain \i (autocommit), each statement is its own implicit transaction, so the GUC's scope ends immediately and current_setting() returns '' by the time the restore runs. We deliberately do not wrap the \i in an explicit BEGIN/COMMIT to "fix" this — verified locally that doing so makes the first error poison the whole transaction, aborting everything after it instead of just these two harmless statements.

Verified locally (both PG12 and PG17) that a deliberately-introduced syntax typo is immediately and clearly surfaced by this test (syntax error at or near "TYPPE" at the exact line), then reverted.

Also had to dial VERBOSITY down from psql.sql's verbose to default in both new files — verbose mode appends backend C source LOCATION lines that differ across PG major versions/builds, which broke matching a single expected-output file across this repo's PG12–17 matrix.

3. Wire it up

• PGXNTOOL_ENABLE_TEST_BUILD = yes explicit in the Makefile (avoids silently disabling the check if test/build/'s contents were ever accidentally deleted).
• CI gap found and fixed: pgxn-tools' pg-build-test never calls make test, so test-build (wired only into test's dependency chain) would never run in CI. Replaced with a direct make verify-results step (see rebase note above for the latest simplification of that step).

Follow-up needed

Landing this means #22 will need a follow-up rebase to remove now-redundant mode-awareness it added to install.sql/pgtap.sql (not doing that here — just flagging it; being handled as part of the same stack restructuring).

Test plan

• make test passes on PG17 (including the new test-build step)
• make test passes on PG12 (including the new test-build step)
• Simulated CI's exact sequence from a fully clean state — no pre-installed extension, no leftover roles/databases
• Deliberately broke sql/test_factory--0.5.0.sql (typo'd CREATE TYPE → CREATE TYPPE) and confirmed both test/build/install.sql and test/build/syntax.sql surfaced it clearly and immediately; reverted before committing
• test/build/expected/*.out generated from real runs (never hand-authored)
• Rebase: cleared installed extension files and confirmed the simplified CI step alone still passes
• Rebase: confirmed test/build/expected/syntax.out's regenerated diff was a pure line-number shift, not a content regression

🤖 Generated with Claude Code

---

Recreated from #26 (fork-headed) as an upstream-branch PR so it can be part of a formal GitHub stack (gh stack link), which refuses fork PRs. Same content/commits, same CI results.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: f289fe61-5669-4e1b-a31e-35531713ecc3

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}