Write a 1byte and a 10GB blob in a test database. If length() takes the same time for both blobs, the blob's length is probably accessed. Otherwise the blob is probably read.
OR: download the source code and debug through it: http://www.sqlite.org/download.html. These are some relevant bits:
/*
** Implementation of the length() function
*/
static void lengthFunc(
  sqlite3_context *context,
  int argc,
  sqlite3_value **argv
){
  int len;
  assert( argc==1 );
  UNUSED_PARAMETER(argc);
  switch( sqlite3_value_type(argv[0]) ){
    case SQLITE_BLOB:
    case SQLITE_INTEGER:
    case SQLITE_FLOAT: {
      sqlite3_result_int(context, sqlite3_value_bytes(argv[0]));
      break;
    }
    case SQLITE_TEXT: {
      const unsigned char *z = sqlite3_value_text(argv[0]);
      if( z==0 ) return;
      len = 0;
      while( *z ){
        len++;
        SQLITE_SKIP_UTF8(z);
      }
      sqlite3_result_int(context, len);
      break;
    }
    default: {
      sqlite3_result_null(context);
      break;
    }
  }
}
and then
/*
** Return the number of bytes in the sqlite3_value object assuming
** that it uses the encoding "enc"
*/
SQLITE_PRIVATE int sqlite3ValueBytes(sqlite3_value *pVal, u8 enc){
  Mem *p = (Mem*)pVal;
  if( (p->flags & MEM_Blob)!=0 || sqlite3ValueText(pVal, enc) ){
    if( p->flags & MEM_Zero ){
      return p->n + p->u.nZero;
    }else{
      return p->n;
    }
  }
  return 0;
}
You can see that the length of text data is calculated on the fly. That of blobs... well, I'm not fluent enough in C... :-)