CoefPackedEncoding in BFV with q - abs(value)

[sarojaduality]

In the EncodeVec function of the CoefPackedEncoding, when encoding negative values, we encode t - abs(value), where t is the plaintext modulus. Before, we used q - abs(value), where q is the ciphertext modulus because this has less noise. Why does using q fail with the noise improvement when using Q/t instead of Delta?