LAPJV is not giving the correct result.

I followed the instructions and installed the package using pip3 install lapjv. However, when I test it with a simple matrix, the results are incorrect. I compared the results with scipy linear_sum_assignment.

from lapjv import lapjv
from scipy.optimize import linear_sum_assignment as lsa
import numpy as np
# cost_matrix = np.array([[9.0, 7.6, 7.5,100.0],[3.5, 8.5, 5.5,100.0],[12.5, 9.5, 9.0,100.0],[4.5, 11.0, 9.5,100.0]])
cost_matrix = np.array([[9.0, 7.6, 7.5,7.0],[3.5, 8.5, 5.5,6.5],[12.5, 9.5, 9.0,10.5],[4.5, 11.0, 9.5,11.5]])
row_ind, col_ind, _ = lapjv(cost_matrix)
cost_tot = 0
print("===LAPJV sol: ")
for id in range (len(row_ind)):
    i = row_ind[id]
    j = col_ind[id]
    print (f"track {i} match det {j} cost {cost_matrix[i,j]}")
    cost_tot += cost_matrix[i,j]
print (f"cost total = {cost_tot}")

row_ind1, col_ind1= lsa(cost_matrix)
print("===LSA sol: ")
cost_tot = 0
for id in range (len(row_ind)):
    i = row_ind1[id]
    j = col_ind1[id]
    print (f"track {i} match det {j} cost {cost_matrix[i,j]}")


    cost_tot += cost_matrix[i,j]
print (f"cost total = {cost_tot}")

I get

===LAPJV sol: 
track 3 match det 3 cost 11.5
track 2 match det 2 cost 9.0
track 1 match det 1 cost 8.5
track 0 match det 0 cost 9.0
cost total = 38.0
===LSA sol: 
track 0 match det 3 cost 7.0
track 1 match det 2 cost 5.5
track 2 match det 1 cost 9.5
track 3 match det 0 cost 4.5
cost total = 26.5

I also took the cpp code in the repo created a main. The project is here lapjv-cpp.zip and run the same squared matrix and I get this

9.000000 7.600000 7.500000 7.000000 
3.500000 8.500000 5.500000 6.500000 
12.500000 9.500000 9.000000 10.500000 
4.500000 11.000000 9.500000 11.500000 
start
Beginning lapjv method.
AVX2: enabled
lapjv: COLUMN REDUCTION finished
lapjv: REDUCTION TRANSFER finished
lapjv: AUGMENTING ROW REDUCTION 1 / 2
lapjv: AUGMENTING ROW REDUCTION 2 / 2
lapjv: AUGMENT SOLUTION finished
lapjv: optimal cost calculated
Track 2 assigned to detect 3.  Cost: 10.5
Track 1 assigned to detect 1.  Cost: 8.5
Track 3 assigned to detect 0.  Cost: 4.5
Track 0 assigned to detect 2.  Cost: 7.5
Total cost = 31


dim     4 - lap cost  3.500 - runtime  0.017 ms

Something is up with this code. I would appreciate any help with it. Am I missing something?

Start by running the Python example from the issue and the attached C++ main, then compare both assignments and total costs with scipy.optimize.linear_sum_assignment. Trace the lapjv entry point and the native implementation used by the package; done means the Python and C++ results agree with the reference assignment for this matrix.