The publish job had never run. It is the part that produces the apt and dnf
repositories, which is to say it is the whole update mechanism, and finding
out whether it works when a tag is already pushed is the wrong time.
A dry run now builds both repositories with a key generated on the spot,
verifies the three signatures it wrote, and then installs the packages back
out of them - apt on the runner, dnf in a Fedora container - so the thing
being tested is the thing that runs. Nothing is pushed and no release is
created.
The runners force actions/checkout@v4 and the artifact actions onto Node 24
already and warn about it on every run. Current majors instead, so the
annotation goes away and the pin says what is actually running.
The code works on those distributions now; there was still nothing to
install. This adds the two packages and, more to the point, somewhere for
them to live that hands out updates - a package a user has to notice a new
version of and download again is not much better than a checkout.
Both are built from `make install` and nothing else. A packaging script that
lists the installed files a second time is a second description of the
layout, and the two drift the first time a file moves; here the Makefile
stays the only place that says where anything goes. The .deb is staged and
wrapped with dpkg-deb, the .rpm goes through a spec whose %install is the
same make invocation. Both are architecture-independent, so one file each
covers Debian, Ubuntu and their derivatives on one side and Fedora, RHEL and
openSUSE on the other.
The version is not written down twice either. The Makefile has it, the
control file and the spec take it as a placeholder, and check-version.sh
refuses a tag that disagrees - otherwise a v1.0.4 release quietly ships a
program that reports 1.0.3.
The release workflow builds both in a Debian and a Fedora container, signs
the RPM where there is a native rpm-sign, attaches both to the GitHub
release, and then adds them to an APT and a DNF repository on gh-pages,
regenerating the indexes over every version ever published so that pinning
and going back to one still work. Missing the signing key is not an error:
it builds, it says in the log that the repositories were left alone, and the
packages are still on the release.
There is also a check workflow, which is the first time shellcheck actually
runs on this.
Neither package carries a maintainer script. The units are enabled per user
by the program itself, so there is nothing for a package to do as root - and
nothing it could do about the changes in a user's home either, which is why
both descriptions say to run `disable` before removing.
packaging/README.md has the two things that cannot be automated: making the
signing key, and pointing Pages at the branch.
Felitendo