fix: hold the changelog to its released changesets, and waive no checksum #32

Merged
james.maes merged 1 commits from fix/changelog-append-only into develop 2026-11-19 11:12:33 +00:00
Owner

What this corrects

0.1.0-RC.3 (#30) took runAlways off the released changeset 1.0.0-guard-existing-tables and gave it validCheckSum: ANY. This changelog's rule is that a released changeset never changes. Nobody else reviewed that change; this one makes it safe, minimal and enforced.

What changes

  • The checksum waiver is removed. It was not needed: only an execution attribute changed, the changeset's changes are what they were, so the checksum an existing install recorded is still its checksum. ANY switched off checksum protection for that changeset for good.
  • A real upgrade test. CmsChangelogReleasedTest migrates a database with the 0.1.0-RC.2 changelog exactly as tagged (twice, as its guard ran every time), then with the current changelog, with nothing waived, and asserts that every history row the first release recorded is untouched and that a further migration changes nothing. The earlier test only wrote an invented checksum into the history table.
  • The rule is enforced by the build. The same test fails when a released changeset differs from src/test/resources/released/qbit-cms.changelog.released.yaml, when any changeset runs on every migration, or when one waives its checksum. When a release is cut, the changelog is copied over that file in the same change (AGENTS.md).
  • The one exception is written down where the rule is: the changelog's header, AGENTS.md, CLAUDE.md and the CHANGELOG's schema notes.

What stays as released in 0.1.0-RC.3

The guard still runs once and no longer on every migration. That is the defect being fixed (it rewrote its history row on every run, so a host with repeatable migrations could not include the changelog), and it cannot be fixed without touching that attribute. What is lost is only the guard's message at a later migration: a feature switched on later still cannot adopt a foreign table, because creating a table that exists fails.

Checks run locally

./mvnw -B verify: 42 tests, 0 failures, Checkstyle 0, guardrail and coverage checks pass.

🤖 Generated with Claude Code

## What this corrects `0.1.0-RC.3` (#30) took `runAlways` off the released changeset `1.0.0-guard-existing-tables` and gave it `validCheckSum: ANY`. This changelog's rule is that a released changeset never changes. Nobody else reviewed that change; this one makes it safe, minimal and enforced. ## What changes - **The checksum waiver is removed.** It was not needed: only an execution attribute changed, the changeset's changes are what they were, so the checksum an existing install recorded is still its checksum. `ANY` switched off checksum protection for that changeset for good. - **A real upgrade test.** `CmsChangelogReleasedTest` migrates a database with the `0.1.0-RC.2` changelog exactly as tagged (twice, as its guard ran every time), then with the current changelog, with nothing waived, and asserts that every history row the first release recorded is untouched and that a further migration changes nothing. The earlier test only wrote an invented checksum into the history table. - **The rule is enforced by the build.** The same test fails when a released changeset differs from `src/test/resources/released/qbit-cms.changelog.released.yaml`, when any changeset runs on every migration, or when one waives its checksum. When a release is cut, the changelog is copied over that file in the same change (`AGENTS.md`). - **The one exception is written down** where the rule is: the changelog's header, `AGENTS.md`, `CLAUDE.md` and the CHANGELOG's schema notes. ## What stays as released in 0.1.0-RC.3 The guard still runs once and no longer on every migration. That is the defect being fixed (it rewrote its history row on every run, so a host with repeatable migrations could not include the changelog), and it cannot be fixed without touching that attribute. What is lost is only the guard's message at a later migration: a feature switched on later still cannot adopt a foreign table, because creating a table that exists fails. ## Checks run locally `./mvnw -B verify`: 42 tests, 0 failures, Checkstyle 0, guardrail and coverage checks pass. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
james.maes added 1 commit 2026-11-19 11:10:56 +00:00
fix: hold the changelog to its released changesets, and waive no checksum
maven / build (push) Successful in 1m37s
build / maven (push) Successful in 1m37s
d7774fd734
0.1.0-RC.3 took runAlways off the released changeset
1.0.0-guard-existing-tables and gave it validCheckSum: ANY. The rule of this
changelog is that a released changeset never changes.

- the waiver is removed: it was not needed (the changeset's changes are what
  they were, so the checksum existing installs recorded is still valid) and
  it switched off checksum protection for that changeset for good
- CmsChangelogReleasedTest migrates an install made by the 0.1.0-RC.2
  changelog as tagged with the current one and asserts that its history
  rows are untouched
- the same test fails the build when a released changeset differs from the
  recorded release, runs on every migration, or waives its checksum
- the one exception (runAlways removed from the guard) is recorded where
  the rule is written

Tests: 42 run, 0 failures, Checkstyle 0.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
james.maes merged commit 99f385a869 into develop 2026-11-19 11:12:32 +00:00
Sign in to join this conversation.