I have stumbled upon what I guess is a wrong choice of parameter M while writing my thesis about a related topic.

You perform rejection sampling based on the Bernoulli sampler with the relevent line being
if (! sampler_ber_exp(&sampler, p.M - norm_v)) {.

While you chose p.M to be P_{max} as in the BLZZRD implementation, this does not work for this rejection sampler. For the Bernoulli sampler p.M needs to be chosen as 2sigma^2ln(M) with the respective values for M given in the BLISS-B paper and the values for sigma given in the BLISS paper.

This is the reason why you needed to add additional bound checks at the end of the signing algorithm. When adapting your p.M, you should be able to ditch those bound checks.