release: experimental Flatpak arches can no longer gate Merge/Publish; drop .#windows hydrate from Build (Linux)

founder commented

2026-08-16 10:51:24 +00:00

Owner

KSE-0186 D2 first slice. Splits the Flatpak matrix into build-flatpak-repo (x86_64+aarch64, required; Merge/Publish depend on it alone) and build-flatpak-repo-experimental (riscv64/arm, continue-on-error, nothing depends on it, OFF unless vars.KEYSTORE_FLATPAK_EXPERIMENTAL_ARCHES=true). Merge already picks experimental artifacts up opportunistically via status.env. Also removes the .#windows hydration step from Build (Linux) (46-51 s/run into a discarded per-job cache on the forge). YAML validated (ruby YAML: required matrix [x86_64,aarch64], experimental [riscv64,arm], merge needs [prepare, build-flatpak-repo]).

founder added 1 commit

2026-08-16 10:51:24 +00:00

release: experimental Flatpak arches can no longer gate Merge/Publish; drop the .#windows hydrate from Build (Linux)

Release: Signing Preflight / Release (Signing Preflight) (push) Successful in 56s

Details

Deploy: Web / Deploy (Web) (push) Successful in 2m13s

Details

Deploy: API / Deploy (API) (push) Successful in 2m49s

Details

Release: If Needed / Check (Release Needed) (push) Has been cancelled

Details

2c5cbbb74a

Merge (Flatpak Repo) gated on `needs.build-flatpak-repo.result`, the aggregate
of a four-leg matrix in which the qemu-emulated riscv64/arm legs were
continue-on-error probes. In Forgejo the aggregate still turns failure/
cancelled when such a leg fails or is cancelled (runner death, sweeper,
timeout), so every one of the 95 Merge failures in the last 30 days failed
at step 0 -- 25 of them with both required arches green -- and no Flatpak
repo has merged since 2026-07-27 (run 16732). The two legs have also been
deterministic no-ops since then ("Nothing matches org.gnome.Sdk in remote
flathub"; the manifest only sources x86_64/aarch64) costing 15-55 min of a
16x32 runner each per release (KSE-0186 D2).

- build-flatpak-repo now carries only the required native arches (x86_64,
  aarch64); Merge and Publish depend on it alone.
- build-flatpak-repo-experimental carries riscv64/arm, continue-on-error,
  nothing depends on it (Merge already picks its artifacts up
  opportunistically via status.env), and it is off unless the repo variable
  KEYSTORE_FLATPAK_EXPERIMENTAL_ARCHES=true.
- Build (Linux) no longer hydrates the .#windows closure into a per-job,
  discarded file:// cache (46-51 s per run on the forge runner, which shares
  the host store anyway).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>