[Bug][template alignment] SIFT matcher giving different matches in different environments?

[Udayraj123]

Describe the bug
The value of file_level_fallback_threshold is different on a local machine vs in CI

To Reproduce
Steps to reproduce the behavior:

1. Use sample '2-template-shifts/alignment'
2. Command(s) used 'uv run main.py -i samples/experimental/2-template-shifts/alignment --debug'
3. See the value of file_level_fallback_threshold

CI Value

2025-08-10T23:31:34.4986281Z            DEBUG    Thresholding:                      interpretation_pass.py:62
2025-08-10T23:31:34.4986415Z                     file_level_fallback_threshold:                              
2025-08-10T23:31:34.4986531Z                     155.9   global_std_THR: 31.36                               
2025-08-10T23:31:34.4986672Z            INFO     Read Response:                                  entry.py:341
2025-08-10T23:31:34.4986789Z                     {'Booklet_No': '04223',

Local Value

    DEBUG    Thresholding:    file_level_fallback_threshold: 134.27  global_std_THR: 10                                            interpretation_pass.py:62
           INFO     Read Response:                                                                                                                     entry.py:341
                    {'Booklet_No': '046233',

sample run: link

Expected behavior
The value of the threshold should be identical, effectively the Booklet_No should match and the test should pass

Screenshots

Desktop (please complete the following information):

• OS: MacOS 15.5 (24F74)
• Python version: 3.13
• OpenCV version: 4.11.0

Additional context
Add any other context about the problem here.

Error Stack trace. Sample images used, etc

[Udayraj123]

The culprit could potentially be in the hashing method used for @functools.total_ordering

        sorted_bubble_means_and_refs = sorted(
            bubble_means_and_refs,
        )

edit: actually the mean values are not accurate themselves - actually the issue would be with running a all folders together vs single folder.

[Udayraj123]

FLANN is based on a randomized k-d tree algorithm that approximates the nearest neighbor. Its aim is to find a fast approximation with acceptable declines in accuracy. As a consequence, the results may not always be the same.

[Udayraj123]

Looks like the input images are incorrect to sift, the issue may be in scan zones