In HM 6.0, split_transform_flag uses a depth-based context initValue table as follows:

INIT_TRANS_SUBDIV_FLAG[3][NUM_TRANS_SUBDIV_FLAG_CTX] =
{
  { CNU,  224,  167,  122,  CNU,  CNU,  CNU,  CNU,  CNU,  CNU, }, 
  { CNU,  124,  138,   94,  CNU,  CNU,  CNU,  CNU,  CNU,  CNU, }, 
  { CNU,  153,  138,  138,  CNU,  CNU,  CNU,  CNU,  CNU,  CNU, }, 
};

The context initValue for depth zero is equal to CNU (context model not used). This is true when LCU size is 64x64 since the split_transform_flag is never coded for depth zero. But when LCU size is less than 64x64, e.g., 16x16, split_transform_size may be coded for depth zero and the context initValue is undefined. Besides that, the context initValue in the spec (Table 9-23) has not been updated yet.

the encoder crahes for all the configurations when 4x4 transform is disabled.

Solution:

8.4.2.1.7 Derivation process for motion vector predictor candidates

Replace

4. When isScaledFlagLX is equal to 0, availableFlagLXB is set equal to 0 and for ( xBk, yBk ) from ( xB0, yB0 ) to ( xB2, yB2 ) where xB0 = xP +nPSW, xB1 = xB0 - MinPuSize , and xB2 = xP - MinPuSize, the following applies repeatedly until availableFlagLXB is equal to 1:

With

4. When isScaledFlagLX is equal to 0 and availableFlagLXB is equal to 1, mvLXA is set equal to mvLXB and refIdxA is set equal to refIdxB and availableFlagLXA is set equal to 1.
5. When isScaledFlagLX is equal to 0, availableFlagLXB is set equal to 0 and for ( xBk, yBk ) from ( xB0, yB0 ) to ( xB2, yB2 ) where xB0 = xP +nPSW, xB1 = xB0 - MinPuSize , and xB2 = xP - MinPuSize, the following applies repeatedly until availableFlagLXB is equal to 1: