SANDOGASA ADAPTATIONS

slopfest/sandogasa is a Rust workspace of Fedora/CentOS/Debian packaging tools (Apache-2.0 OR MIT). Several of its tools solve, in mature form, problems this factory had been hitting reactively — each of the adaptations below is pinned to a factory incident that predates it. The ideas and algorithms were reimplemented in this repository's Python, not linked as binaries: the algorithms are small, the factory already parses its own repo metadata, and a Rust toolchain dependency for CI-side checks would cost more than it saves.

Licensing: sandogasa is dual-licensed Apache-2.0 OR MIT, which permits reimplementation and adaptation here; the per-file docstrings record the origin. Test vectors for the version comparator are carried over verbatim.

What was adapted, and from where

Factory piece	Adapted from	The incident it pins
scripts/rpm_vercmp.py — librpm's rpmvercmp with ~/^, EVR compare, constraint check	sandogasa-rpmvercmp crate (vectors carried over)	FACTORY-STATUS.md measures presence, not freshness; every version-aware check below needs this primitive
scripts/preflight-buildrequires.py version_blocked	ebranch BlockedByBase	libnotify >= 0.8.7 unsatisfiable on both arches, found by mock 2.5 h in, twice (#480)
scripts/preflight-buildrequires.py runtime_unsatisfied	ebranch check_installability	gtkgreet → greetd, xfce4-pulseaudio-plugin → pulseaudio: runtime holes found at clean-install time after a 53-minute build (#480)
scripts/check-published-hygiene.py	hs-relmon dupe-subpkgs + file-conflicts	the createrepo_c --update stale-entry class (#358); 107 same-NEVRA pairs with differing checksums on xfce/10-stream (#471)
scripts/check-reverse-deps.py + the NEVER BREAK RDEPS gate in publish-rpm-wave.sh	ebranch check-update	the gnome50 bootstrap glib2 Obsoletes hijacking AppStream (run 32405815822); publishing a libnotify the factory could no longer rebuild
scripts/extract-buildroot-manifest.py + scripts/diff-buildroots.py, recorded per package by build-chain.sh into artifacts/buildroots/	koji-diff	the #480 libnotify buildroot diagnosis, reconstructed from issue comments when mock's root.log had the answer
scripts/collect-cell-throughput.py	koji-lag's measure-from-metadata approach	docs/hummingbird-throughput.md was a one-off hand scrape (it found concurrency 1.0 with --jobs 2); now re-runnable against any cell log

One level of support across targets

sandogasa itself is Fedora/CentOS-centric; adapting it naively would have produced an EL-focused toolset bolted onto a multi-target factory. The adaptations are therefore built on a format-neutral layer, scripts/repo_index.py: one index shape for every reader (rpm-md primary.xml, flat-APT Packages, pacman .db) and one version comparator per format — rpm_vercmp, deb_version (validated against real dpkg --compare-versions, 900/900 pairs), and libalpm's variant in pacman_db. The comparators provably disagree on real versions, so no format is ever judged with another's ruler. tests/test_target_tooling_parity.py is the enforcement: a format declared in manifests/package-factory.yaml without a reader, a comparator, a publish gate, and a buildroot record is a red test.

Capability	rpm (el10/fedora/hummingbird/tumbleweed)	deb (ubuntu/debian)	pkg.tar.zst (arch)
index reader + version comparator	✅	✅	✅
served-index hygiene	✅	✅ (no file lists in flat APT → file-conflict check inert, said in-tool)	✅ once a published_index is declared
reverse-dep publish gate	✅ publish-rpm-wave.sh	✅ publish-tideforge-debs.yml (old-vs-new Packages)	✅ publish-tideforge-arch.yml (old-vs-new .db)
buildroot manifests + differ	✅ mock root.log / installed_pkgs	✅ dpkg-query after build-dep	— no arch build chain exists yet; the differ already parses any manifest a future one writes
build/version/runtime preflight	✅ (gap-engine manifests)	measured by the deb backport gap engine (measure-deb-backport-gap.py, pre-existing)	— no chain to preflight yet
throughput timers	✅ mock's own timers	— the deb chain logs no per-package timer; add one there before extending the tool	—

Where each runs

• Preflight (preflight-buildrequires.py): manual gate before dispatching a chain; now answers build-time satisfiability, version constraints, and runtime installability in one run.
• Hygiene (check-published-hygiene.py): ad hoc or scheduled; reads the same published_index contract every buildroot reads — every format, through repo_index — so a clean report covers the combination of prefixes a buildroot sees. First live run: 8 findings on el10 (gtk-layer-shell and xfconf families served identically from both prefixes), hummingbird and both deb targets clean.
• Reverse-dep gates: every publish path refuses a publish that breaks what is already served, entirely locally. rpm gates inside publish-rpm-wave.sh (staged repodata vs the synced-down tree); deb and arch regenerate their whole index in place, so they gate old-vs-new (check-index-regression.py) in their publishers, with native semantics — apt's highest-version candidate and | alternatives on deb, libalpm ordering on arch. All are differential by design: only what a publish newly breaks counts, so distro-archive dependencies outside the view are never noise, and the blind spots all lean lenient (documented in each script).
• Buildroot manifests: the rpm chains record mock's resolved buildroot (opt-in via BUILDROOT_MANIFESTS, switched on by the cell runner, into artifacts/buildroots/); the deb chain records a dpkg-query snapshot after every build-dep into its uploaded buildroots/. One differ reads both conventions.
• Throughput: run by hand against a downloaded cell job log when the 6-hour-ceiling work needs numbers.

What was considered and NOT adapted

Recorded so the next reader does not re-survey the same ground:

• The forge/bureaucracy tooling (Bodhi, Bugzilla, FESCo, Pagure ACLs, meetbot, activity reporting) has no counterpart in this factory's problem space.
• dbranch assumes Debian's dist-git/PPA workflow; the deb side here is the backport chain (backport-deb-chain.yml), a different shape.
• ebranch's fedrq shell-outs were not carried over: the factory already parses primary.xml itself, and a run-time dependency on a Fedora-packaged query tool would not work in the deb and arch legs.
• koji-lag's SQLite store is the right shape if throughput collection ever becomes scheduled; for on-demand use, stateless parsing of one log is enough and simpler.