Streamline desktop publishing and reuse packaged components

[SakuraMathcraft]

The desktop release run took 51m15s and failed after publishing an incomplete release because the Linux GPU carrier exceeded GitHub's 2 GiB asset limit. Its final job spent 14m05s downloading build artifacts that were already available on the platform runners.

• Upload installers, the manual, and required update payloads directly from their producing runners. Only the small distribution manifests travel to the final catalog job.
• Publish a draft only after the selected jobs finish. Keep only referenced update resources and user downloads; remove unreferenced sidecars and stale assets when completing a release.
• Remove mandatory native installation/repair/uninstall acceptance from the default publishing path. The standalone acceptance CLI remains available; CUDA hardware checks remain explicitly opt-in.
• Cache verified component ZIPs and reuse them without recompression. Remove the overlapping raw-runtime/download caches, overlap CPU/GPU packaging, and avoid unsigned-installer artifact uploads when SignPath is unavailable.
• Use XZ for Linux offline carriers and reject oversized assets before upload. Document the current release path and Linux archive format in the authoritative guides.

Validation: 545 pytest tests and 57 subtests passed; Pyright, Ruff, formatting, compileall and actionlint passed. Temporary production-code checks exercised archive cache restore/reuse, corrupt-cache rejection, Linux offline payload round trips, upload selection, and metadata-only finalization/asset cleanup. No persistent tests or new release validation jobs were added.

Full desktop installers have not been rebuilt with this change. Actual Linux GPU carrier size and new end-to-end timings must be confirmed by the next Actions build.