NeXTSTEP CD-ROM builder, old-tar import fixes, RISC-V CI + GUI fixes

[danifunker]

Summary

Adds rb-cli optical new next-ufs (alias nextstep) and a TUI wizard target that build NeXTSTEP / OPENSTEP CD-ROM images. A NeXT "ISO" is not ISO 9660: every NeXTSTEP 3.x CISC distribution disc is a dlV3 NeXT disk label in 2048-byte sectors wrapping one 4.3BSD UFS partition, and that is the shape this writes.

Building real discs from a NeXT software archive exposed several import bugs, fixed here too.

NeXT CD builder

• partition::next_cd_builder streams the label + UFS (8 KiB blocks / 2 KiB fragments, fs_fsbtodb 0, 32×64 geometry, 80-sector porch). For a disc the size of NeXT's own NeXTTIME CD, every derived superblock word and the label's slot 0 match the real disc byte for byte (pinned in unit tests).
• ufs_format::Bsd43Geometry makes the 4.3BSD track geometry / rps / maxcontig / maxbpg configurable; hard-disk output is unchanged.
• Bug fix: NeXT's DIRBLKSIZ is a compile-time 1024 even on 2048-byte media. The reader derived 2048 from fsize >> fsbtodb, so editing any NeXT CD wrote directory chunks a NeXT kernel would reject. Verified against all 3168 directories on the NeXTSTEP 3.3 CISC disc.
• NextPartitionSpec::cpg, boot-block numbers scaled to d_secsize.
• CLI flags mirror sgi-efs (--size/auto, --from-dir, --expand-archives, --flatten-folders, ...) plus --expand-gunzip and --bytes-per-inode. TUI: new CD-ROM target with an extra Gunzip toggle.

Import

• --expand-gunzip (also on rb-cli import): strip one gzip layer only (x.tar.gz -> x.tar, f.gz -> f).
• Pre-POSIX (v7) tar sniffing, and typeflag-less name/ entries read as directories.
• Tar hard links land as copies of their target (spooled, not loaded into RAM). They used to be dropped, e.g. Opener.app's gunzip, gcc's cross-driver names.
• GNU tar 1.11 @@MaNgLeD.N long names restored from the trailing N entry.
• Old NeXT tar puts a hard link's target size in the link header with no data behind it; a small header normalizer zeroes it, as GNU tar ignores it.
• A member repeated within one archive: the later copy wins, as with GNU tar.

Raw file-name bytes (NeXT only)

NeXTSTEP names are raw NeXTSTEP-encoded or EUC-JP bytes, not UTF-8. fs::raw_name carries undecodable bytes through the &str API as Private Use Area-B placeholders and writes them back byte for byte. Gated by EditableFilesystem::stores_raw_names, which only NeXT UFS volumes report; every other filesystem and UFS flavour keeps its previous behaviour (non-NeXT tar targets get %XX).

RISC-V CI and GUI

• CI (Build Linux RISC-V 64 GUI): glib is Multi-Arch: same, so the riscv64 sysroot needs the exact glib version of the runner's amd64 install. The old pin to the release pocket broke whenever the runner image's glib moved on (intermittent on main since 2026-09-10). The install step now tries each glib version available for riscv64 dev packages and amd64, newest first, pins it on both arches, and keeps the first one apt-get install -s resolves (today 2.80.0-6ubuntu3.9). Green on 1c662201.
• GUI segfault on RISC-V hardware: on an Orange Pi RV2 (PowerVR BXE-2-32, Ubuntu 24.04) eframe picked wgpu, wgpu picked the PowerVR Vulkan driver, and the driver segfaulted; a segfault bypasses the wgpu -> glow -> software fallback. Linux on aarch64 / arm / riscv64 now starts on OpenGL (glow) first (GL_FIRST), which also targets the slow wgpu path reported on the Raspberry Pi 5. Other platforms are unchanged.
• Renderer override through elevation: pkexec strips the environment and the elevated relaunch did not pass RUSTY_BACKUP_RENDERER / WGPU_BACKEND, so "Unlock Physical Devices" dropped a working override. Both now pass through. The override is documented in the README install steps.

Also

• fix(os): device_open_error no longer compiles on Windows, where it has had no caller since c85fa9b, so clippy -D warnings passes on a Windows host again.
• Docs: README, cli-examples ("Build a NeXTSTEP CD-ROM"), OPEN-WORK (§3.3 follow-ups, §6.2 GUI path, §10 audit).

Scope cuts

• Data discs only: no boot blocks. RISC (HP-PA/SPARC hybrid) discs are not produced. No GUI path yet (OPEN-WORK §6.2).
• The UFS writer still gives each file tail a whole 8 KiB block (NeXT packs tails into fragments); --size auto budgets for it. OPEN-WORK §3.3.

Testing

• scripts/preflight.sh green: cargo test --release, MiSTer feature set, Rust 1.73 vintage floor, doc parity, rb-regress.
• New unit tests: geometry / superblock / label pinned to the NeXTTIME CD, reader + fsck round trip, v7 / hard-link / mangled-name / duplicate-member / non-UTF-8 tar cases, raw-name round trip including the on-disk byte. CLI integration: tests/cli_suite/cli_next_cdrom.rs.
• Built nine 640 MiB discs from a NeXTSTEP software archive with --expand-archives; all pass our fsck and an independent Python UFS walk (no directory-chunk violations).
• RISC-V GUI: on the Orange Pi RV2 the old default segfaults every time; WGPU_BACKEND=gl and RUSTY_BACKUP_RENDERER=glow run clean. This PR's build (v2026-09-25-14-54), launched with no overrides, logs Using glow, draws the full window, and runs until closed. Raspberry Pi 5 not tested here.
• Real NeXT: under the Previous emulator (NeXTstation, 68040, v66 ROM, NeXTSTEP 3.3) the disc auto-mounts in Workspace on boot, as root and as a normal user. The Developer disc also reads on the MiSTer NeXT core. Launching apps from the disc under Previous is not yet confirmed (keyboard-only driving through WSLg).

🤖 Generated with Claude Code