nx_arena_types.nx

nx_arena_types.nx -- arena type + constants, zero-syscall. FOUR-PILLAR FIX (2026-05-16): nx_arena.nx hard-imports nx_syscalls.nx because nx_arena_new() calls sys_mmap internally. Any caller that wanted to compose against nx_arena PLUS a different syscall layer (e.g., nx_syscalls_x86_64.nx for native exec) hit linker symbol redefinition -- the F-meta-4-class cross-target portability gap. This file ADDS the inversion-of-control entry point. It carries: - struct NxArena (the type) - NX_ARENA_BYTES (sizeof constant) - nx_arena_align_up (pure math) - nx_arena_alloc / _alloc_zero / _reset / _used / _cap_bytes / _alloc_count / _oom_count (operate on a caller-owned arena) - nx_arena_init(a, base, cap) (init a caller-allocated NxArena struct + caller-allocated backing buffer) What it does NOT carry: nx_arena_new (which mmaps). That stays in nx_arena.nx for backwards-compat + convenience. Callers that want a different syscall ABI can: 1. allocate the NxArena struct + backing buffer themselves 2. call nx_arena_init(a, base, cap) 3. pass `a` to nx_arena_alloc + friends as before Per cardinal user-owns-every-bit: this is the canonical shape. The caller mmaps the bytes (using whichever syscall layer matches the target ABI); substrate just uses them. nx_capability_claims: needs: [pointer_arithmetic, struct_field_assign] provides: [arena_type, arena_init, arena_alloc, arena_reset] safety: [no_unchecked_deref, no_floating_point, no_syscall, bit_equal_reproducible, target_agnostic] verdict: [no_silent_failure (oom returns null + n_oom++)] license: ORIGINAL kind: racing_crew_specialist