Arch: single registry, broad aliases, x86_64 canonical

[lacraig2]

Fourth PR in the config-reshape series. Targets config-reshape-patches (#820) (it overlaps arch.py, utils.get_arch_subdir, and templating.build_context). Merge #818 → #820 → this.

Makes architecture naming consistent and forgiving: one source of truth, accept any common spelling, and flip the odd-one-out intel64 to x86_64.

One registry

src/penguin/arch_registry.py (pure stdlib) defines one ArchSpec per arch holding every per-namespace name — arch_subdir, dylib_subdir, kmod_subdir, qemu/panda arch, machine, cpu, kconf, kernel_fmt/kernel_whole, serial (major,minor), console fixup, endianness — plus accepted aliases. normalize_arch()/spec()/accessors are alias-tolerant. A unit test asserts the schema Literal equals arch_registry.all_names(), so it can't drift.

Broad aliases, x86_64 canonical

You can now write x86_64, intel64, amd64, arm64, ppc64le, powerpc64el, … — all normalize to one canonical name at load. x86_64 is now canonical (intel64 is an accepted alias), matching the on-disk asset layout where x86_64 was always the odd one out.

Consolidation (deletes the duplicates)

Refactored every consumer to read the registry: utils.get_arch_subdir/get_driver_kmod_path, arch.get_dylib_subdir, dropin_compile.DYLIB_DIRS, q_config.qemu_configs, abi_info, the nvram2 dylib-override copy, and the inline serial/console branches. Bugs fixed along the way:

• q_config load_q_config("powerpc64le") no longer KeyErrors (it keyed the stale powerpc64el), and it now returns a fresh dict instead of mutating a shared module dict.
• dropin_compile.DYLIB_DIRS was missing intel64 (would look in dylibs/intel64) — gone.

Safe transition (no x86 breakage)

The only two assets named by the config arch are guesthopper.<arch> and sysroots/<arch>. utils.resolve_arch_asset prefers the canonical (x86_64) name and falls back to an alias filename that actually exists, so x86 rehosting keeps working against the current guesthopper.intel64 artifact. The in-repo Dockerfile sysroots stage is renamed intel64→x86_64.

Cross-repo follow-up (deferred, non-breaking): rename the guesthopper sibling artifact guesthopper.intel64→guesthopper.x86_64. The resolver covers the interim.

Left intentionally

arch.arch_filter/arch_end and static_analyses keep emitting intel64 — that's the ELF-identifier namespace, distinct from config arch; normalization happens at the identifier→config boundary (set_arch_info).

Known pre-existing smell (not changed)

penguin_prep.py keeps its own ARCH_ABI_INFO copy (drifted from abi_info.py on powerpc64/powerpc64le); I rekeyed/normalized it but didn't merge the duplicate to avoid behavior change. Flagged for a follow-up.

Testing

tests/unit_tests/test_config.py: 75 passing on host — alias normalization (all arches, case-insensitive, unknown raises), Literal↔registry sync, subdir/dylib/kmod parity vs the OLD tables, load_config with intel64 and x86_64 producing identical realized configs, q_config powerpc64le + fresh-dict, abi rekey, dropin dedupe, templating alias→canonical. Recommend tests/comprehensive/test.sh across the arch matrix in-container before merge.