nx_hw.nx

nx_hw.nx -- runtime hardware probes (CPU count, cache line, page). SOVEREIGN: every probe goes through a real syscall or CPUID-class instruction. No /proc parsing, no libc sysconf. Same code paths on RV64 + x86_64 because the syscalls are number-portable. FOUNDATION for [[feedback-dynamic-hw-sizing-no-hardcoded-thread-counts]]: every threading / parallel / pool / partition primitive in the substrate must derive its sizing from THESE functions, not from hardcoded literals. Syscall numbers used: 123 = sched_getaffinity (rv64/asm-generic) -- cpu count. The compiler's x86ctx_rv64_to_x86_64_syscall table translates 123 -> x86_64 204 (row added + blessed 2026-07-07). HISTORY: the original constant 122 was WRONG twice over -- rv64 122 is sched_SETaffinity, and 122 was not in the swap table so it passed through to x86_64 setfsgid, which SUCCEEDS, leaves the mask all-zero, popcounts to 0, and made this probe silently report 1 CPU forever (every auto-sized pool ran single-threaded; caught by nx_conv2d_mt_gate check 10, 2026-07-07, interim-fixed with raw 204 pass-through, then made portable via the table row the same day). Cache-line probe is currently a sealed-enum default (64 B, the near-universal value across x86 / ARM / RISC-V); CPUID-based probing on x86 lands when we add an inline-asm intrinsic. Page size is fixed at 4 KiB (Linux RV64 + x86_64 default). Larger pages (2 MiB hugepages) handled by future nx_hugepage.nx.