HUMMINGBIRD TARGET

What "build for Hummingbird" means in this repository, stated once with the measurements behind it, so the same facts stop being re-derived from version strings. Written 2026-09-02 after comparing this factory with projectbluefin/utah-packages, which built GNOME 51 for Hummingbird in a fraction of the time this repository has spent on it. Every number below came out of a real primary.xml; the reproduction is scripts/check-hummingbird-installability.py and scripts/gap_engine.py --catalog manifests/hummingbird-desktops.yaml.

Hummingbird's own documentation is authoritative over anything here. When the two disagree, theirs wins and this file is the bug (https://hummingbird-project.io/docs/, https://gitlab.com/redhat/hummingbird).

1. What Hummingbird is, by ABI

Indexes read 2026-09-02, x86_64 (primary.xml sha256 prefixes: public-hummingbird 5c75eb01881b, Fedora 44 c48e47563bbf, Rawhide 8daf5868c675):

package	hummingbird	fedora 44	rawhide
glibc	2.43-8.4.hum1	2.43-2.fc44	2.44.9000-1.fc46
openssl-libs	3.5.8-0.1.hum1 (libcrypto.so.3)	3.5.5-1.fc44 (.so.3)	4.0.2-1.fc46 (.so.4)
python3	3.14.7-1.hum1	3.14.3-2.fc44	3.15.0~rc1-1.fc45
perl-libs	5.42.2-525.hum1	5.42.1-523.fc44	5.44.0-527.fc45
libxml2	2.15.3-0.1.4.hum1	2.12.10-6.fc44	2.13.9-4.fc45
glib2	2.89.3-1.hum1	2.88.0-1.fc44	2.89.4-1.fc46
harfbuzz	14.3.1-1.hum1	12.3.2-1.fc44	14.4.0-1.fc46

Two things are true at once, and the 2026-08-06 measurement (docs/hummingbird-desktop-gap.md, Finding 1) saw only the first:

• The packages Hummingbird rebuilds (glib2, harfbuzz, gcc, rust, meson) are at Rawhide's versions. That is why "30 of 38 core packages match Rawhide" was a true sentence.
• The packages that define the ABI -- glibc, openssl, python, perl -- are Fedora 44's. Hummingbird is an overlay of rebuilt packages on a pinned Fedora release; its own container repository maps the "rawhide" variant to fedora-44.repo (redhat/hummingbird/containers, images/variables.yml), and its mock/mock.cfg composes the build root from that release's repositories with the Hummingbird Pulp repos shadowing them by priority.

So the build root is Fedora 44 plus public-hummingbird at higher priority. That is what Hummingbird builds in, what utah-packages builds in, and -- since 2026-09-02 -- what mock/hummingbird-ci*.cfg builds in.

2. What the Rawhide root cost

The Rawhide root was not a harmless approximation. Read off our own published prefix, repo.tunaos.org/hummingbird/20251124-x86_64 (index dacef1b3b4c9…, 8738 binaries):

libm.so.6(GLIBC_2.44)(64bit)   needed by gstreamer1-plugins-good, libavutil-free, zvbi

Those three were built, signed, published, and cannot install on the target they were built for: Hummingbird's glibc is 2.43. utah-packages hit the same class one soname over (gnome-shell built on Rawhide needing libcrypto.so.4) and documented it as the reason for the Fedora 44 root. The three interpreter pins the old config carried ([fedora-44-python], [fedora-44-perl], [fedora-44-mpich], each an includepkgs list) were patching Fedora 44 back in one namespace at a time, and each said "remove when Hummingbird catches up with Rawhide". It was never going to.

3. Does what we publish resolve on Hummingbird?

The question a consumer image asks, and the one utah-packages gates its publish on (it runs the consumer transaction inside the bootc-os image with only Hummingbird and its factory enabled). Static answer, 2026-09-02, roots from manifests/hummingbird-desktops.yaml, closure over public-hummingbird ∪ the named repository, nothing pre-installed:

repositories enabled	gnome roots absent	closure	unresolved capabilities
hummingbird alone	53 of 58	94	5
hummingbird + utah-packages (Pages mirror, 421 binaries)	33	208	185
hummingbird + tunaos-hummingbird (8738 binaries)	3	602	30
hummingbird + Fedora 44 (if a consumer enabled it)	0	695	0

Per desktop, against our prefix: gnome 30 unresolved, kde 25, cosmic 9, niri 28, xfce 18 -- no desktop is resolvable today, and the blockers are concrete (scripts/check-hummingbird-installability.py prints them):

• the GLIBC_2.44 leak above (gnome, kde, cosmic all reach it);
• gdm requires gnome50-el10-compat, an EL10-only package, because src/gnome-51 is a byte copy of src/gnome-50 (#537);
• gtk4 requires libgstplay-1.0.so.0 and nothing ships gstreamer1-plugins-bad-free (#540);
• libraries the closure needs and the chain has not produced: libcanberra, evolution-data-server (libecal, libedataserver), webkitgtk6.0, libosinfo, samba (libsmbclient), udisks2, libvorbis, poppler-data;
• roots in neither repository: fprintd-pam, gnome-disk-utility, gnome-user-share.

Two readings of that table matter more than the counts. First, the utah-packages row is not evidence that utah's approach fails: its Pages mirror lags the OCI image utah consumes, and its consumer contract is Bluefin's list, not this catalog's 58 roots. Second, the Fedora 44 row is the shape of the shortcut: a Hummingbird image that enabled Fedora 44 would resolve GNOME 50 entirely, modulo the name conflicts the walk cannot see (libxml2 2.12 vs 2.15 under one name). This factory exists because consumer images do not enable Fedora 44; that row is the ceiling, not the plan.

scripts/check-hummingbird-installability.py runs daily (.github/workflows/hummingbird-installability.yml), advisory, per arch. It becomes a gate the day a desktop first resolves.

4. What utah-packages does differently, and what this repository took

	utah-packages	tunaos-packages before	after 2026-09-02
build root	Fedora 44 + public-hummingbird (priority)	Rawhide + hummingbird + three F44 pins	Fedora 44 + public-hummingbird + our prefix
recipes	Rawhide dist-git, pinned by commit	Rawhide dist-git, imported at build time	unchanged
sources	upstream release tarballs, SHA-512 locked	dist-git lookaside	unchanged (see §6)
what gets built	a hand-curated, dependency-first list (193 sources)	the measured runtime closure (670 sources, 18 tiers)	unchanged (see §5)
staging	5 stages, each a local repo for the next	tiers over one topological order, [local-build]	unchanged
disttag	.hum1.bfin (sorts above .hum1)	.bfin1 (sorts below .hum1)	unchanged, deliberate: a rebuild never shadows the base
publish	OCI image, consumed with COPY --from by digest	R2 prefix behind a Worker	unchanged
installability gate	dnf transaction inside the bootc-os image before publish	none	static checker, daily; container gate is the next step
preflight	resolves every recipe's BuildRequires in the real root, as a worklist	scripts/preflight-buildrequires.py exists	unchanged

The root and the gate are the two that decide whether output is usable. The rest is scope.

5. Why the gap did not shrink when the reference moved

An obvious hypothesis was that measuring the closure against Rawhide inflated it. Measured, same day, same roots:

reference	gnome sources to build	kde	cosmic	niri	xfce
Rawhide (committed)	301	375	175	297	268
Fedora 44	311	378	152	306	256

No. The closure is what the roots need that Hummingbird does not ship, and Hummingbird ships no desktop stack at all; every cairo, gdk-pixbuf and pipewire is a gap whichever Fedora you read the Requires: from. The reference stays Rawhide, because the recipes are Rawhide's and Rawhide's binary index is the closest proxy for what those recipes produce.

What utah-packages did about the size was not a measurement trick. It built a narrower contract -- one desktop, Bluefin's list -- and it built it in the right root, so what came out installed. This catalog declares five desktops; that is a scope decision for the maintainer, and §7 makes the case.

6. Open, and deliberately not asserted

• Rawhide recipes on a Fedora 44 root will hit version floors. Rawhide specs BuildRequire what Rawhide has. The toolchain floors are covered -- Hummingbird itself ships gcc 16, rust 1.97, meson 1.11 at priority 10 -- but library floors (utah met pango >= 1.58, wayland-protocols >= 1.48, gsettings-desktop-schemas >= 51.alpha) will surface as preflight-buildrequires.py findings and become build-order entries. That is the expected cost, and it is the cost utah paid.
• The three GLIBC_2.44 packages must be rebuilt, and the prefix should be re-verified for any other symbol-version leak (GLIBC_2.44, OPENSSL_3.5+, libpython3.15). The checker's needer_from: published column is the list.
• Name masking is the overlay's inherent cost. dnf's priority filter drops a name from every lower-priority repo, so a Fedora 44 package that needs a library the base ships under the same name at a different soname cannot resolve: F44 fontforge wants libxml2.so.2, the base's libxml2 2.15 provides .so.16, and F44's libxml2 is never a candidate. The old includepkgs list handled four such names by keeping them out; the answer per package is the same either way -- exclude the Fedora name or build the package against the base -- and it is a per-package decision, not a global list. utah-packages carries exactly this list as HB_EXCLUDE (libicu, ruby-default-gems, gpgme, qt6-qtbase).
• Upstream sources. utah's direct-source model (fetch the release tarball, verify SHA-512, fail closed) is the Hummingbird project's own freshness model; this repository still builds from dist-git lookaside. Not changed here; it is orthogonal to whether output installs.
• The container gate. The static checker cannot judge version constraints or conflicts. The next step is utah's step verbatim: run the catalog's roots through dnf --assumeno install inside quay.io/hummingbird-community/bootc-os with only public-hummingbird and our prefix enabled, in verify-package-factory-cell.sh, and refuse to publish a wave that regresses it.

7. The decision this measurement asks for

tunaOS's hummingbird images consume ghcr.io/projectbluefin/common and ghcr.io/ublue-os/brew by digest today. ghcr.io/projectbluefin/utah-packages is published the same way, is Bluefin's GNOME 51 for Hummingbird, and is built by the project tunaOS is derived from. Building a second GNOME 51 for the same base, in this repository, on a smaller team, is the part of the plan that has not been paying back.

Proposed split, for the maintainer to decide (tracked in the issue that accompanies this document):

• GNOME on hummingbird: consume utah-packages. tunaOS's manifests/desktops/gnome.yaml hummingbird section points at the OCI repository by digest; src/gnome-51 here stops being a second copy.
• This factory builds what utah does not: the KDE, COSMIC, Niri and XFCE stacks for Hummingbird, in the Fedora 44 root, gated on installability -- the same shape, one desktop at a time, dependency-first.
• Both factories share the base-image audit (components: in the catalog) so a package neither builds is a named gap, not a surprise.

That is the model utah-packages proved: right root, one contract, a gate that asks the consumer's question. Not a bigger build order.

8. Taken, 2026-09-02, and what it measured

The split in §7 is implemented, on both sides:

• tunaOS (hummingbird:gnome): image-versions.yaml pins ghcr.io/projectbluefin/utah-packages by digest, Containerfile.el10 bind-mounts its /repository at /run/utah-packages for the gnome stage's install, and manifests/desktops/gnome.yaml's hummingbird section enables it as a file:// repo at priority 4, above the tunaos-hummingbird prefix (5).
• This factory: the same digest is a consumed_indexes entry on the hummingbird target (manifests/package-factory.yaml), read straight out of the registry by scripts/oci_repository.py (token, manifest, one streamed layer; only repository/repodata/* is kept). The gap engine counts everything utah ships as already had; the installability walk enables it; the container gate (scripts/check-hummingbird-installability-container.sh, utah's dnf --assumeno step verbatim in shape, inside the pinned bootc-os image) materialises it from the digest and asks dnf. The catalog's gnome desktop is consumed_from: utah-packages; its roots stay declared so the audit and both gates still cover them.

Measured on the regeneration that landed with it (target primary 3f8c4b86…, utah primary 38da5357…, 421 binaries counted as had):

desktop	sources to build, before	after	Δ
gnome	301	248	−53
kde	375	353	−22
cosmic	175	115	−60
niri	297	252	−45
xfce	268	115	−153
bluefin (parity set)	392	186	−206
build order, packages	778	664	−114

Two readings. The GNOME row is the smallest shrink, not the largest: utah's 421 binaries are Bluefin's contract, and this catalog's 58 GNOME roots reach past it (the fedora list's extras -- gnome-initial-setup, ptyxis, gjs, nautilus and five more are still built here from src/gnome-51). The other four desktops shrink because utah ships the platform every desktop stands on -- gtk4, glib2, pipewire, mesa, the portals -- and that is the part of the closure this factory no longer has to build twice. The reference is unchanged (Rawhide, §5).

What was NOT cut, because measurement said otherwise (§7's "scope decision"): openSUSE Tumbleweed serves a resolvable index on both architectures (rpm/opensuse-tumbleweed/{x86_64,aarch64}, 2026-09-02), so it stays declared -- with the note that no tunaOS manifest consumes it yet (sailfin's niri reads Tumbleweed's own repositories); the flat apt repositories serve amd64 and arm64 debs, so the deb targets keep both; EL10, Fedora and Hummingbird have aarch64 legs. The only architecture that fails the bar is arch's aarch64: no leg, no index, no consumer. Narrowing it is decided but ships separately: editing architectures schedules every Arch cell, and the Arch engine cannot yet build a cell whose runtime dependencies are sibling recipes (dms-cli needs dms and quickshell; makepkg has no repository to find them in), so that PR waits on same-wave sibling staging for the Arch engine. GNOME-on-DEB was not removed either: it is its own measured chain now (manifests/gnome51-deb.yaml, docs/gnome51-deb-gap.json), not the 1-of-9 recipe gap the July readiness note described.

9. What the first legs on this root failed, and why it was not the root

Run 33597624514 (this branch, both architectures, 2026-09-02) was the first full build on the Fedora 44 root. Both legs hit the four-hour deadline with the chain healthy -- x86_64 built 132 packages and deferred 646, aarch64 built 189 and deferred 622 -- and failed three:

package	leg	what dnf or rpmbuild said	on main's Rawhide root
langtable	both	%check: xmllint: command not found	identical (runs 33580878734, both legs)
python-dbus-next	aarch64	builddep: python3-pytest-asyncio-1.1.0-3.fc44 requires python3.14dist(pytest) < 9, cannot install both python3-pytest-9.1.1-1.1.hum1 and python3-pytest-8.4.2-3.2.hum1	never reached (deferred)
python-aiohappyeyeballs	aarch64	same	never reached (deferred)

langtable is Fedora's spec calling a tool it never names: Koji's buildroot carries libxml2, Hummingbird's does not. It is now a vendored spec in src/hummingbird/langtable/ differing from Rawhide's by one BuildRequires: libxml2, and the build order no longer re-imports it.

pytest-asyncio is the shape of failure this root will keep producing and the one to recognise: a BuildRequires-only tool that Fedora 44 ships in a version window Hummingbird's own packages have moved out of. Hummingbird publishes pytest 9.1.1 (with 8.4.2 and 9.0.3 beside it in the same Pulp repository); F44's pytest-asyncio wants < 9; mock's best=1 will not settle for the older pytest, so the transaction fails rather than degrades. Three answers were possible:

• best=0, which is what utah-packages effectively runs (plain dnf builddep in a container, where best defaults off). It fixes this and every case like it silently -- including the ones where "the older candidate" is a C library's -devel and the produced binary then wants a soname the image no longer ships. Rejected: this file's own rule is fail closed.
• an excludepkgs= for the newer pytest, the by-name precedent mock/hummingbird-ci.cfg already uses for docutils/sphinx. Rejected: it would build every consumer against a pytest the target no longer installs.
• build Rawhide's pytest-asyncio 1.4.0 (pytest < 10) first, so local-build (priority 1) masks Fedora's by name. Taken: bootstrap-04 in manifests/hummingbird-desktops.yaml, next to the PEP-517 backends the measurement cannot discover for the same reason (they appear only in BuildRequires).

The rule that falls out: a BuildRequires-only tool whose Fedora 44 copy cannot install beside the target's own packages is built here from the Rawhide reference. tests/test_the_first_fedora44_legs_taught_two_buildroot_gaps.py pins both fixes to the files the chain reads.