[SakuraMathcraft]

Fix native release failures during vendor Python extraction, Qt resource inventory, frozen application validation, shared model transfer, and DMG creation. Linux offline installers now stage exactly the declared components and distribution manifest, so first installation passes the deployment contract.

• Extract the verified vendor Python tar with the standard data filter, prune unused terminfo, and remove Qt aliases whose targets were pruned.
• Initialize WebEngine before QApplication for frozen validation; write failure reports and exit without a blocking traceback dialog.
• Normalize model file permissions before inventorying and transferring shared artifacts; create APFS disk images on macOS ARM.
• Reduce compression overhead with fast gzip for Linux carriers, two LZMA2 block threads for Windows, uncompressed Actions transport for compressed artifacts, and uncompressed temporary Debian transaction packages. Optional missing validation reports no longer produce upload warnings.

Validation: local pytest, Pyright, Ruff and actionlint passed. Actions run 39103755312 succeeded for all three desktop builds and native installation, same-version model repair, edition replacement, CPU formula/text/mixed recognition, and uninstall. GPU editions passed installation and repair on Windows/Linux; real CUDA hardware checks were not selected. Publishing remained disabled.

Measured packaging times: Linux 10m36s to 5m43s; Windows GPU installer compilation 8m43s to 6m34s. Faster carrier compression can modestly increase installer size.