Add POSIX shared-memory support

[dg1sbg]

Adds POSIX named shared memory to Clasp as a thin syscall layer plus a small Lisp ergonomics layer, exposed through CLASP-POSIX.

What this adds

src/core/shmem.cc — thin CL_DEFUN wrappers over shm_open, shm_unlink, ftruncate, mmap, munmap, mprotect, msync, mlock, munlock, getpagesize, strerror. Every wrapper returns (values result errno) with errno 0 on success, so no C++ exceptions cross the FFI boundary and there is no errno race. core__sys_shm_constants returns an alist of the O_* / PROT_* / MAP_* / MS_* values so the Lisp side never hardcodes platform numbers.

src/lisp/kernel/lsp/shmem.lisp — flag handling that accepts a keyword, a list of keywords, or a raw integer; a syscall-error condition carrying errno and the call name with a strerror-based report; a single %checked macro that turns (values result errno) into either the result or a signalled condition; a shm-mapping struct; and the high-level open-shared-memory, close-shared-memory, and with-shared-memory.

src/lisp/regression-tests/shmem.lisp — 10 tests: round-trip through a mapping, two mappings of one fd aliasing the same page, error on a missing object, munmap idempotence, page-size sanity, cross-process IPC via fork, plus the four regression tests described below.

Notes for review

SHM-MAPPING, not MAPPING. CORE::MAPPING is already an exported Clasp class — Mapping_O in hashTable.h, the abstract base of StrongMapping_O and WeakKeyMapping_O. A defstruct mapping inside (in-package #:core) redefines it as a STRUCTURE-CLASS, and the image then fails to load with "When redefining a class, the metaclass can not change". The struct is therefore named shm-mapping, with (:conc-name mapping-) so the accessors read naturally (mapping-address, mapping-size, mappingp). Happy to rename the accessors too if you'd prefer they carry the shm- prefix.

Mappings are not GC-reclaimed, by design. mapping-pointer hands out a ForeignData that holds no reference back to the shm-mapping, so a GC finalizer that unmaps would be free to run while that pointer is still live — the finalizer table is weak-keyed, so nothing keeps the mapping alive. Rather than hand out a pointer that can be invalidated underneath the caller, there is no finalizer: lifetime is explicit via munmap / close-shared-memory, with with-shared-memory for the scoped case. This mirrors the resolution taken for foreign-alloc in #1792.

Only a creator may size a shm object. ftruncate on an already-existing POSIX shm object fails EINVAL regardless of the fd's access mode, and even when the requested size is unchanged. open-shared-memory therefore attempts O_CREAT|O_EXCL|O_RDWR first and sizes the object only when that succeeds, falling back to a plain open on EEXIST. Without this, any call against an already-existing region fails, in every :direction — which is the normal case for the second and every later participant.

Testing

macOS arm64, native build, LLVM 22.1.8, rebased onto current main (da70d6ff1):

GC variant	tree tested	successes	unexpected failures	shmem tests
boehm	this branch alone	1971	none	10/10
boehmprecise (clean build from scratch)	this branch + #1810	1973	none	10/10

The four expected failures (SBCL-CROSS-COMPILE-4, INCLUDE-LEVEL-2B, INCLUDE-LEVEL-3, TYPES-CLASSES-10) are pre-existing and unrelated. The count differs between variants only because boehmprecise also runs the variant-specific snapshot tests.

boehmprecise is the interesting one here: shmem.cc compiles and links under precise stack scanning with no GC layout descriptor, confirming it introduces no managed C++ class. The cross-process fork IPC test passes, as does a probe that forces GC while holding only the raw mapping pointer — the failure mode the removed finalizer would have caused.

For disclosure, the boehmprecise run came from a tree that also carried #1810 on top of this branch; #1810 touches only clos/slot-value.lisp and no shmem code.