Running neural networks in a custom kernel#

Context#

I copied my neural network code into the nirOS directory.

integrate the NN code into the kernel build
remove userspace and STL dependencies that do not belong in a freestanding kernel context
make math, allocation, randomness, and data loading work with kernel-safe paths

The neural network sources now have kernel-oriented build and runtime shims in place instead of depending on ordinary userspace facilities.

copied NN code into the nirOS directory
added CPPFILES discovery and .cpp objects / deps in GNUmakefile
added NN include path to CPPFLAGS in GNUmakefile
added or kept a kernel-safe .cpp compile rule with -fno-exceptions -fno-rtti plus freestanding-safe extras
added a math.c special compile path to avoid -mgeneral-regs-only for floating-point code
added kernel libc support files
added C++-safe extern "C" guards in libc headers
updated the NN shim to use kernel libc math and allocation
replaced STL RNG with xorshift32 plus Box-Muller
updated dense-layer init to use kernel-safe RNG and math with no <random> or <cmath>
rewrote the CSV loader to use TarFS plus raw buffer parsing with no std::string, ifstream, or STL
added tarfs_open / tarfs_read API and used it in the CSV path

Current proof captured in this note is structural only:

the build system changes are listed
kernel-safe math, allocation, RNG, and file access replacements were added
this note does not yet include a successful build log, boot log, or runtime inference output

userspace or STL dependencies getting reintroduced into freestanding kernel code
compile flags that silently break floating-point or C++ integration in kernel space
TarFS parsing behaviour diverging from the original userspace CSV loader
allocator or math shim behaviour differing enough to invalidate model assumptions