Protect NNUE header parsing against oversized descriptions

[RainRat]

Summary

• cap the NNUE header description length and verify remaining bytes before reading
• ensure the NNUE header reader stops on truncated or otherwise corrupt metadata
• add regression tests that load toy NNUE files through the Stockfish binary to confirm failures

Testing

• make build -j2
• python - <<'PY'
  import pathlib
  import struct
  import subprocess
  import tempfile
  VERSION=0x7AF32F20
  HASH_VALUE=1007697522
  MAX_DESCRIPTION_LENGTH=4096
  root=pathlib.Path('.').resolve()
  stockfish=root/'src'/('stockfish.exe' if subprocess.os.name=='nt' else 'stockfish')
  commands='\n'.join(['uci','setoption name UCI_Variant value chess','setoption name Use NNUE value true','setoption name EvalFile value {path}','isready','go depth 1','quit'])+'\n'
  def run(payload):
  with tempfile.NamedTemporaryFile(prefix='chess-', suffix='.nnue', delete=False) as tmp:
  tmp.write(payload)
  tmp_path=pathlib.Path(tmp.name)
  try:
  result=subprocess.run([str(stockfish)],input=commands.format(path=tmp_path.as_posix()),text=True,capture_output=True,timeout=10)
  return result.returncode,'was not loaded successfully' in result.stdout
  finally:
  tmp_path.unlink(missing_ok=True)
  print('Oversized header check:',run(struct.pack('<III',VERSION,HASH_VALUE,MAX_DESCRIPTION_LENGTH+1)))
  print('Truncated header check:',run(struct.pack('<III',VERSION,HASH_VALUE,16)+b'abcd'))
  PY

---

https://chatgpt.com/codex/tasks/task_e_68ce0e5c2834833096412d6177c0c77e