Test install foundation: TEST_LOAD_SOURCE modes, dependency guard

[jnasbyupgrade]

Summary

• Adds test/install/load.sql (pgxntool's test/install feature): installs extension_drop ONCE, committed, before the pgTAP suite runs, instead of every test file reinstalling it via test/deps.sql.
• TEST_LOAD_SOURCE (Makefile) selects fresh (default) / update / existing install mode, propagated to load.sql via a GUC over PGOPTIONS.
• Adds test/sql/dependency_guard.sql: proves non-CASCADE DROP EXTENSION is blocked by the dependency guard.
• test/deps.sql simplified: no longer creates a schema or installs the extension (that's load.sql's job now).
• test/sql/simple.sql: removed a SET search_path line that was dead code once installs moved to the ambient/public path (search_path always includes public, so it couldn't prove anything test/sql/schema.sql doesn't already cover for the custom-schema path).
• test/sql/schema.sql: added the minimal preamble (drop the committed install, recreate in its own schema) needed because load.sql now pre-installs into public before this file runs.

Previously this PR also included a schema-name mixed-case/quoting-robustness rename to test/sql/schema.sql -- that's unrelated to this PR's actual purpose (the test/install foundation) and has been split out to #34, which this PR is now rebased on top of.

Test plan

• make test passes (load, dependency_guard, schema, simple, build)
• make lint clean

[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: 22edf953-5274-4098-8ded-9383100dc240

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.}

[jnasbyupgrade]

@claude review

[claude]

Claude finished @jnasbyupgrade's task in 3m 12s —— View job

---

Claude is reviewing this PR

• Gather context (diff, changed files, repo CLAUDE.md)
• Review Makefile TEST_LOAD_SOURCE changes
• Review test/install/load.sql
• Review test/sql/dependency_guard.sql
• Review test/deps.sql / test/sql/schema.sql / test/sql/simple.sql changes
• Post review feedback
  · branch test-install-foundation

---

Review

Traced the full execution path (pgxntool/base.mk's existing test/install schedule feature → Makefile → load.sql → each test/sql/*.sql file) rather than just reading files in isolation, since the whole point of this PR is that state now persists across files. It holds together correctly:

• Sequencing is sound. REGRESS is built from a plain sorted file list (pgxntool/base.mk:72), not a schedule with multiple names per line, so test/sql/*.sql runs strictly sequentially: dependency_guard.sql → schema.sql → simple.sql. Each file's pgTAP transaction is rolled back on exit (finish.sql's # TRANSACTION INTENTIONALLY LEFT OPEN!), so schema.sql's non-CASCADE DROP EXTENSION extension_drop; (test/sql/schema.sql:22) never collides with dependency_guard.sql's guard view — that view never escapes its own rolled-back transaction.
• test/deps.sql simplification is safe. Confirmed no other test file reads :TEST_SCHEMA/:"TEST_SCHEMA" expecting deps.sql to have set it — schema.sql sets its own copy locally (test/sql/schema.sql:2). Removing the old schema-creation/install block from deps.sql doesn't strand anything.
• Makefile TEST_LOAD_SOURCE validation logic is correct — $(filter $(TEST_LOAD_SOURCE),fresh update existing) works here (empty for a typo, non-empty for a valid mode) because TEST_LOAD_SOURCE is a single word with no % wildcard, so argument order doesn't change the result. One nit: this reverses the conventional Make idiom $(filter <patterns>,<text>) (normally the static list comes first), which could read as backwards/confusing to a future maintainer even though it's correct today. Not worth a fix, just flagging.
• load.sql's three modes (fresh/update/existing) are internally consistent, the existing-mode assertions (extension present + version matches default_version) actually protect against a silent no-op, and the PR is upfront that update mode has no real historical version to exercise yet (documented in both the Makefile and load.sql, and enforced — TEST_UPDATE_FROM is required, no unsafe default).
• dependency_guard.sql proves a real property (non-CASCADE DROP EXTENSION fails via 2BP01) rather than just asserting the SQL exists, and correctly uses the 4-arg throws_ok form to check the actual SQLSTATE rather than string-matching the message.
• CI is green on this PR (make test + make lint).

No functional issues found. This is a well-reasoned, well-documented change — the inline comments carry real rationale (e.g. why dependency_guard.sql's view types on extension_drop__commands's row rather than a column, why test/install/.gitignore excludes load.out) rather than restating the code.

[github-actions]

Claude finished @jnasbyupgrade's task in 4m 3s —— View job

---

Claude finished reviewing PR #19

• Gather context (diff, changed files, repo CLAUDE.md)
• Re-verify PR content against CLAUDE.md's schema-independence rules
• Post review feedback

· branch test-install-foundation

---

Review

This PR was already reviewed once in this thread and merged. I independently re-pulled and re-read the actual changed content (Makefile, test/install/load.sql, test/sql/dependency_guard.sql, test/deps.sql, test/sql/schema.sql, test/sql/simple.sql) rather than assuming the prior pass — findings below confirm the earlier review and add one more check specific to this repo's CLAUDE.md.

Sequencing / mechanics — same conclusion as before, re-verified: test/install/load.sql installs extension_drop once, committed, before the pgTAP suite runs; each test/sql/*.sql file's transaction rolls back on exit, so schema.sql's non-CASCADE DROP EXTENSION never collides with dependency_guard.sql's guard view. test/deps.sql no longer creates a schema or installs the extension — confirmed no other file depends on it having done so.

dependency_guard.sql correctly uses the 4-arg throws_ok(sql, sqlstate, message, description) form to check 2BP01 rather than string-matching, and asserts both that the drop is blocked and that the extension/guard view survive the attempt — a real behavioral proof, not just "the SQL exists."

CLAUDE.md schema-independence check — this repo's CLAUDE.md sets a specific, easy-to-get-wrong bar: at least one tested leg must install into a schema verifiably absent from search_path at assertion time, via CREATE EXTENSION ... SCHEMA, with a mixed-case/quoting-sensitive name. Reading test/sql/schema.sql end to end:

• TEST_SCHEMA_2 (_Test_Ed_2, mixed-case → requires quoting) is installed via CREATE EXTENSION extension_drop SCHEMA %I while TEST_SCHEMA is still in search_path — not by mutating search_path first.
• search_path is then explicitly reset to exclude TEST_SCHEMA_2 ("Ensure test schema 2 isn't in search_path") before extension_drop__update()/extension_drop__get() are exercised against it via fully-qualified %I.function() calls.

That satisfies both conditions in CLAUDE.md (verifiable absence at assertion time + quoting-sensitive name) for this leg. The minimal preamble this PR itself adds to schema.sql (drop the committed load.sql install, recreate locally) is consistent with that existing structure and doesn't disturb it.

Makefile TEST_LOAD_SOURCE validation — confirmed still correct: $(filter $(TEST_LOAD_SOURCE),fresh update existing) works because TEST_LOAD_SOURCE is a single word with no % wildcard. Same nit as before: this reverses the conventional $(filter <patterns>,<text>) idiom (static list normally comes first), which could read as backwards to a future maintainer even though it's correct today. Not worth a fix.

load.sql's three modes (fresh/update/existing) remain internally consistent; existing mode's assertions (presence + version match) genuinely guard against a silent no-op; update mode is honestly documented as wired-up-but-unexercised since extension_drop has no real second released version yet, and TEST_UPDATE_FROM has no unsafe default.

No functional issues found on re-verification. The CLAUDE.md schema-independence property this repo cares about specifically is upheld by the existing test/sql/schema.sql structure that this PR builds on top of.