In the WD, subclause 8.4.3.1.6 "Specification of Intra_Angular (2..9, 11..25, 27..34) prediction mode" the following equation is shown (8-48):
refMain[ x ] = p[ −1, −1+( ( x*invAngle+128 )>>8 ) ], with x=( nS*intraPredAngle ) >>5..−1

For the case of an intra 4x4 block with an intra mode of 25, intraPredAngle is -2 and invAngle is -4096 (table 8-5 and 8-6). Then, the index range goes from (4*-2)>>1 .. -1, which is -1..-1. This implies that the equation is applied once, with x = -1. In that case we index array p with [-1,-1+((-1*-4096+128)>>8)] = [-1,15]. However, array P is specified to be sized from [-1..2*ns-1,-1..2*ns-1], [-1..7,-1..7] in this case.

HM seems to avoid this by just skipping the loop altogether, since it has the following code:

      Int invAngleSum    = 128;       // rounding for (shift by 8)
      for (k=-1; k>blkSize*intraPredAngle>>5; k--)
      {
        invAngleSum += invAngle;
        refMain[k] = refSide[invAngleSum>>8];
      }

In this case, the exit condition is met immediately, since k is not larger than -1 (it is equal to -1). If the HM model followed the WD exactly, the loop exit condition would have been a great-or-equal operator like so:

      for (k=-1; k>=blkSize*intraPredAngle>>5; k--)