I believe you are misreading 9.2.3.2.4: where does it say that a bit has to be read? The "last bit" refers to the most recently read bit, not one that you have to read additionally.

I think this may still be an issue.

In our parser, which is written in a domain-specific language very similar to the syntax tables in the text, code looks more-or-less exactly like that in chapter 7 of the text. There is one exception in the syntax for coding unit:

if (pcm_flag[x0][y0)]) {
    pcm_alignment_one_bit()         f(1)     <-- not in the text
    while (!byte_aligned())
	pcm_alignment_zero_bit()    f(1)

This allows pcm_flag to be handled by the exact same termination routine as elements end_of_slice_segment and end_of_subset_one_bit. The latter two elements are always followed by rbsp_stop_one_bit and alignment_bit_equal_to_one respectively. I can't see why pcm_flag doesn't also get a "one bit" in the text.

Unless someone can explain how pcm_flag is special (in which case I apologise for reopening a can of worms), the coding unit syntax could be incorrect.

I am looking at the published edition 1.0 specification Rec. ITU-T H.265 (04/2013) and see this.

if( pcm_flag[ x0 ][ y0 ] ) {
    while( !byte_aligned( ) )
        pcm_alignment_zero_bit f(1)
    pcm_sample( x0, y0, log2CbSize )
} else {

This is not that same as what you quoted.

Yes, that is the problem. I think the text may be incorrect and so proposed a correction. Your published excerpt is different to my proposed. My theory is that there should be a pcm_alignment_one_bit in the text (all versions). Since this text has been reviewed so much now, I am hesitant to believe that this theory could be correct. But Parabola may be the only company that has "compiled" and run the syntax tables in the text of the standard..?

For end_of_slice_segment_flag the ‘one bit’ corresponds with rbsp_stop_one_bit – part of subclause 7.3.2.11 rbsp_trailing_bits(), part of subclause 7.3.2.10 rbsp_slice_segment_trailing_bits(), which follows slice_segment_data, which terminates with a end_of_slice_segment_flag equal to one, so again a redundant ‘one bit’ is avoided.

For end_of_subset_one_bit the ‘one bit’ corresponds with alignment_bit_equal_to_one – part of subclause 7.3.2.12 byte_alignment(), which follows end_of_subset_one_bit so avoids a redundant ‘one bit’.

In both cases, a one-valued flag decoded at a lower level syntax table is effectively reused as a flag required to be one at a higher level syntax table(where CABAC decoding is terminated).
When terminating CABAC due to pcm_flag equal to one, zero or more pcm_alignment_zero_bit follow pcm_flag but the coding unit syntax table continues after the PCM block. As such, there is no redundant ‘one bit’ introduced by a higher level syntax table that can be avoided by also re-interpreting the pcm_flag (as is done in the other two cases).