The TM document never defined the DCT directly. Appendix A had a single sentence referring to the video Working Draft (which of course only defines the Inverse Discrete Cosine Transform). But like many things, the knowledge was too well known and considered too insignificant in the context of far more difficult problems by the readers and writers of the TM and was therefore left out of the document. (TM was justifiably written from the perspective of an insider for an insider to meet the straightforward goals of the time).
It is important that the encoder use an accurate forward DCT, and a fully compliant IDCT in the decode subloop. Also, heed the warning in IS 13818-2 Section 7.4.4 Note 2.
During the bitstream exchange phase (April 1993 - March 1994), double floating point (64-bit IEEE) as used in the IEEE 1180 reference IDCT model, was strongly recommended so as everybody could reproduce identical frames in the decoder.