Fix nrLDPCEncode(algo="sionna") for valid BG2/K=640 inputs

[ThereWasAYang]

Related Issue

Fixes #78

Summary

This PR fixes a bug in nrLDPCEncode(..., algo="sionna") for valid BG2 low-rate inputs such as:

• tbs = 552
• R = 120 / 1024
• BGN = 2
• K = 640
• Zc = 64

Previously, this case failed with a matrix dimension mismatch during encoding.

Root cause

In the current implementation, k_b is used for two different purposes:

1. selecting the lifting-size set (i_ls)
2. selecting the number of systematic base-graph columns passed into _gen_submat(...)

For BG2 and K=640, the code correctly selects:

• k_b = 9

but this value is only intended for lifting-size selection.

When constructing the encoding submatrices, _gen_submat(...) should still use the base-graph systematic width:

• nsys = 10 for BG2

Using k_b=9 here causes the base graph to be sliced incorrectly, which leads to a dimension mismatch later in _encode(...).

Fix

Keep k_b for lifting-size selection, but use nsys when calling _gen_submat(...).

Before

pcm_a, pcm_b_inv, pcm_c1, pcm_c2 = _gen_submat(bm, k_b, Zc, bgn)

After

pcm_a, pcm_b_inv, pcm_c1, pcm_c2 = _gen_submat(bm, nsys, Zc, bgn)

I also removed the unused pcm = _lift_basegraph(bm, Zc) line from this branch.

Regression test

This PR adds a regression test covering a valid BG2/K=640 case:

• tbs = 552
• R = 120 / 1024

The test verifies that nrLDPCEncode(..., algo="sionna") no longer crashes and returns an output with the expected shape.

Observed error before this fix

For the valid input above, the encoder failed with:

ValueError: matmul: dimension mismatch with signature (n,k=576),(k=640,1?)->(n,1?)

Why this is correct

For BG2, the reduced k_b values (10/9/8/6) are part of the lifting-size selection procedure in 38.212. They do not change the actual number of systematic base-graph columns, which remains fixed by the base graph itself.

So:

• k_b should affect i_ls selection
• nsys should affect _gen_submat(...) slicing