fix: To fix DEM canonicalisation

[kaiqiy-nv]

This PR tries to fix [B] 6234203 .

It also strengthen the test cases to expose the bug and test further:

• Specify merge_strategy='independent' in test_pymatching_decode_to_observable_surface_code_dem, since correct DEMs can contain parallel matching edges.
• Use the external Stim’s DEM results to compare with the cudaq-qec's results.
• Stim tests use pytest.importorskip("stim") so unsupported platforms get skipped.

Thank you for looking at this PR.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[bmhowe23]

/ok to test c374d76

[bmhowe23]

Thanks, @kaiqiy-nv. Could you please investigate the errors and warnings in

• https://github.com/NVIDIA/cudaqx/actions/runs/26766866582/job/78895569832?pr=568#step:16:962
• https://github.com/NVIDIA/cudaqx/actions/runs/26766866582/job/78895569832?pr=568#step:16:932

For the warning, could we please cap it at ~10 messages per function invocation so as to not spam the user's console with a short distance code?

[kaiqiy-nv]

Hi @bmhowe23 ,

To cap warnings at 10 messages per function invocation -- DONE in the new push

Cause of the CI failure above here:

1. We updated canonicalize_for_rounds to keep same-H diff-O signatures, which gives parallel edges in matching-graph terms.
2. z_dem_from_memory_circuit calls canonicalize_for_rounds, so dem.detector_error_matrix now contains those parallel-edge columns.
3. The failed test case test_sliding_window_1 builds the full decoder first qec.get_decoder(decoder_name, dem.detector_error_matrix). For pymatching, plain construction uses the default merge_strategy = disallow without observables provided, so it rejects the parallel edges, giving ValueError:.

Potential need to update sliding_window.cpp:

Setting merge_strategy="independent" alone does not fix the test logic because it only stops the construction crash, not the assertion. Without O, PyMatching decodes in column space and the full vs. windowed decoders pick different representative columns, so exact-equality still fails.

We may need further updates in sliding_window.cpp (slice and forward the observables matrix O per window, and decode/commit in observable space) to enable a genuinely observable-aware sliding-window decode — i.e., a faithful full-vs-window equivalence for matching decoders.

Do we need to implement this update?
If so, we can do it in another PR, and for now, we only consider to update the test_sliding_window.py.

Quick fix in the local test cases:

We have changed the asserts of the pymatching cases to compare syndrome reproduction ((H @ e) % 2 == syndrome) instead of exact equality, because same-H diff-O columns are identical in H, so the exact column is not unique but every valid correction must still reproduce the syndrome.
single_error_lut cases are kept for strict equality.

We have also added one more test case test_pymatching_parallel_edges_use_observable_faults to make up for direct-PyMatching coverage to make sure that H-only PyMatching rejects parallel edges while the O + merge_strategy path decodes successfully.

[bmhowe23]

/ok to test 0e431bb

[bmhowe23]

Hi @bmhowe23 ,

To cap warnings at 10 messages per function invocation -- DONE in the new push

Cause of the CI failure above here:

1. We updated canonicalize_for_rounds to keep same-H diff-O signatures, which gives parallel edges in matching-graph terms.
2. z_dem_from_memory_circuit calls canonicalize_for_rounds, so dem.detector_error_matrix now contains those parallel-edge columns.
3. The failed test case test_sliding_window_1 builds the full decoder first qec.get_decoder(decoder_name, dem.detector_error_matrix). For pymatching, plain construction uses the default merge_strategy = disallow without observables provided, so it rejects the parallel edges, giving ValueError:.

Potential need to update sliding_window.cpp:

Setting merge_strategy="independent" alone does not fix the test logic because it only stops the construction crash, not the assertion. Without O, PyMatching decodes in column space and the full vs. windowed decoders pick different representative columns, so exact-equality still fails.

We may need further updates in sliding_window.cpp (slice and forward the observables matrix O per window, and decode/commit in observable space) to enable a genuinely observable-aware sliding-window decode — i.e., a faithful full-vs-window equivalence for matching decoders.

Do we need to implement this update? If so, we can do it in another PR, and for now, we only consider to update the test_sliding_window.py.

Quick fix in the local test cases:

We have changed the asserts of the pymatching cases to compare syndrome reproduction ((H @ e) % 2 == syndrome) instead of exact equality, because same-H diff-O columns are identical in H, so the exact column is not unique but every valid correction must still reproduce the syndrome. single_error_lut cases are kept for strict equality.

We have also added one more test case test_pymatching_parallel_edges_use_observable_faults to make up for direct-PyMatching coverage to make sure that H-only PyMatching rejects parallel edges while the O + merge_strategy path decodes successfully.

@kaiqiy-nv I believe the underlying cause of the problem was in the original QEC circuit. In particular, our memory circuits terminate after the stabilizer rounds but do not include the boundary detectors synthesized from the final data qubit readout. This makes the circuits too sensitive to errors in the final stabilizer rounds. The true fix for this is in work, and it will incorporate use of CUDA-Q detectors (freshly added to CUDA-Q) in NVIDIA/cuda-quantum#4602.

Bottom line: I would like to incorporate your updates, but please allow a bit of time to sort out the multiple PRs that it will take to properly resolve this.