bin/release: the version check scans every ## header, so an ## Unreleased section makes it re-register the last release

[1-Bort-1]

TL;DR

bin/release read the changelog version with an awk scan that only stopped once it had found a v[0-9] field, so a top section without one — ## Unreleased — fell through to the next ## header and returned the previous release. The exit now fires on the first ## line whatever it holds, so an unversioned top header yields an empty version and the guard refuses.

What it did on this tree

main at fb1eea7 carried version = "5.0.0" in Project.toml and ## Unreleased above the ## VortexStepMethod v5.0.0 2026-09-07 section, which is exactly the shape that defeats the scan — the fallthrough finds v5.0.0, it matches Project.toml, and the guard passes:

$ ./bin/release --dry-run
Repository:   OpenSourceAWE/VortexStepMethod.jl
Version:      v5.0.0
Commit:       fb1eea7d9965980b0893dbc8d90e6bbc6f57788c

Release notes:
### Added
- `solve!(...; throw_on_fail=true)` throws a `SolveFailure` ...
### Fixed
- The `NONLIN` solver backtracks along each Newton step ...
exit=0

That is a run of ./bin/release posting its registration comment for v5.0.0 — already tagged and registered — carrying the notes of the version that has not gone out yet. The same tree after the fix:

$ ./bin/release --dry-run
Version mismatch: Project.toml is 5.0.0 but CHANGELOG top is  (expected v5.0.0).
Update CHANGELOG.md so its first '## ' header matches the package version.
exit=1

The message the guard already prints — "Update CHANGELOG.md so its first ## header matches the package version" — is what the scan was supposed to be checking, so this is the code catching up with its own error message rather than a new rule. The empty slot it leaves in that first line ("CHANGELOG top is (expected ...)") reads a little bare; I left it alone rather than add a branch for the no-version case, since the sentence still says the true thing.

The test

test/bin/test_release.jl builds a throwaway repository — Project.toml at 5.0.0, a two-section CHANGELOG.md, git init and a commit, a stub gh first on PATH so nothing reaches GitHub — and runs bash bin/release --dry-run in it.

The fixture has to put the package's own version in the second section for the bug to bite: a first draft with v4.3.1 down there passed against the unfixed script, because the fallthrough then returned a version that mismatched anyway. The test that reproduces is the pair ## Unreleased over ## Fixture v5.0.0, asserting a non-zero exit and Version mismatch on stderr; the second testset is the control, a top header naming the package version, which must still exit 0 with nothing on stderr — without it the first would pass on any failure at all, gh not being installed included.

It is registered under Sys.isunix() in test/runtests.jl: the script is bash, and the Windows runner in the matrix has no business running it.

No CHANGELOG.md entry. bin/install and bin/run_julia changes are logged there because users run them; bin/release is the maintainer's, and its top section is what #302 is turning into v5.1.0 right now — a bullet from me would land in the middle of a release in flight for a line no package user reads. Say the word and I will add one.

Why there is a Register Failed comment on this thread

JuliaRegistrator answered this pull request here. Describing what bin/release posts meant writing the trigger phrase, and backticks do not hide it from the bot — it reads the PR body, found the phrase, and tried to register. It failed only because the author's org membership is not public. So the phrase must not appear in any comment or PR body on this repo, and nothing above spells it out. Worth knowing while #302 is in flight: it is the same accident this issue is about, arriving by a different door.

The red check is #300, not this branch

Test end-user and developer setup is red on ea6680877 (34645746964). It is ArgumentError: reducing over an empty collection at src/airfoil_aero/kulfan.jl:71 — argmin on an empty x — through fit_kulfan_parameters → generated_slices → plot_slices_3d from examples/ram_air_kite.jl:79, ending the job at run_all failed for 1 example(s): ram_air_kite.jl, 15 passed / 1 failed. That is #300 down to the same stacktrace, already written up there and with the fix open as #301: write_section_aero writes an all-NaN _d1.dat, read_dat_coordinates drops every NaN NaN row, and the reader hands fit_kulfan_parameters nothing.

Nothing on this branch can reach it — the diff is bin/release, a new test file and one line of test/runtests.jl, none of which the setup workflow's examples touch. It is landing on whatever is in the queue: #306's branch, whose diff is solver code only, failed the same job with the identical signature nine minutes before mine (34645270956) and passed it an hour earlier (34639515312); release/v5.1.0 (34628824619) and #292's branch (34637447090) failed it too. Which XFoil deflections converge moves run to run — #300's own finding, and #291's. It is not falling my way, though: the previous commit 0524692 hit it on both attempts of its run, and the merge of main has now hit it again.

No new issue: #300 has it, with more detail than I could add.

Verification

• Reproduced first: ./bin/release --dry-run on fb1eea7 printed Version: v5.0.0 with the Unreleased notes and exited 0
• test/bin/test_release.jl red before, green after (juliaserver): before, "refuses an unversioned top section" failed 2/2 while the control passed 2/2; after, 4/4 pass — re-run on the merged head ea66808, still 4/4
• test/runtests.jl bin/test_release.jl through the suite runner: 4/4 PASS
• Local full suite on 0524692: PASS (7 min, Julia 1.12.7, one matrix cell) — the agent:ci-local comment on the thread. Not re-run on the merged head: the box killed it twice for memory with other agents on it, and GitHub CI covers ea66808 on seven cells, so I left the box alone rather than take a third run
• GitHub CI on ea66808: all seven cells of 34645746941 green — Documentation, 1.12 macOS/ubuntu/windows, 1.11 ubuntu, and both coverage jobs — plus codecov/patch. Test end-user and developer setup red on kulfan.jl:71 — write_section_aero writes an all-NaN deflected .dat, and read_section_ #300, as above
• No new public symbols, so no docs page · no REUSE setup in this repo
• Merged origin/main (ea66808); no conflict — main's side is src/solver.jl, src/settings.jl, CHANGELOG.md and test files, none of them this branch's
• Benchmark: n/a
• Risk: the fixture's stub gh answers only gh repo view; a future bin/release that shells out to gh for something new before the version guard would make the control testset fail rather than the case it is guarding.

Scope

+73 / -1 across 3 files, of which 70 lines are the new test. bin/release is one line: the exit moves out of the if. Not stacked — #302 touches only CHANGELOG.md and Project.toml. CI.yml's fail-fast: false (gate item 10) is #279's, already queued.

Closes #280 · task VortexStepMethod.jl-280

[JuliaRegistrator]

Error while trying to register: Register Failed
@1-Bort-1, it looks like you are not a publicly listed member/owner in the parent organization (OpenSourceAWE).
If you are a member/owner, you will need to change your membership to public. See GitHub Help

[1-Bort-1]

Local full suite: PASS (7 min, Julia 1.12.7, one cell of the matrix)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[1-Bort-1]

CI: Test end-user and developer setup failed on ea6680877.

https://github.com/OpenSourceAWE/VortexStepMethod.jl/actions/runs/34645746964/job/103415986216