fix(noise): honor GeminiNoiseModelABC scaling_factor in squin lowering

[weinbe58]

Summary

Fixes #785.

generate_simple_noise_model / generate_logical_noise_model read the raw *_px fields off the GeminiNoiseModelABC (e.g. noise_model.mover_px), which ignore the model's scaling_factor. The scaled values only live behind the getter properties (mover_pauli_rates, two_qubit_pauli, …). As a result, in the squin/stim path every value of scaling_factor produced identical noise — it worked at the cirq level but was silently a no-op downstream (see the discussion in QuEraComputing/bloqade-circuit#836).

Changes

• Single-qubit Pauli rates now come from the scaled getters: mover_pauli_rates, sitter_pauli_rates, local_pauli_rates, global_pauli_rates, cz_unpaired_pauli_rates.
• Correlated CZ probabilities come from two_qubit_pauli.error_probabilities (scaled), with .get(k, 0.0) so a rate that scales to zero (and gets dropped from the dict) no longer risks a KeyError.
• Loss probabilities are left raw on purpose — upstream scaling_factor only scales Pauli rates, not loss. Per-category loss scaling is the genuinely-new axis tracked upstream in bloqade-circuit#836. Using the getters here also future-proofs lanes against that upstream work.
• New python/tests/test_noise_scaling.py walks the generated squin kernel IR and asserts the baked-in constants reflect scaling_factor across all noise kernels, that loss stays raw, that scaling_factor=0 zeroes rates without KeyError, and (regression) that distinct scales now produce distinct constants.

Verification

• test_noise_scaling.py + test_noise_init_generator.py: 37 passed
• rewrite/move2squin/test_noise.py + test_device.py: 27 passed, 3 skipped
• black / isort / ruff / pyright: clean

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 91.66667% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
python/bloqade/lanes/noise_model.py	91.66%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[Copilot]

Pull request overview

Fixes a squin/stim-lowering bug where GeminiNoiseModelABC.scaling_factor was effectively ignored by generate_simple_noise_model / generate_logical_noise_model, by switching from raw *_px fields to the scaled getter APIs for Pauli noise rates.

Changes:

• Read single-qubit Pauli rates via *_pauli_rates getters so scaling_factor is honored in generated squin kernels.
• Read correlated CZ Pauli probabilities from two_qubit_pauli.error_probabilities and default missing keys to 0.0 to avoid KeyError at scaling_factor=0.
• Add a new test suite that inspects squin kernel IR constants to regression-test scaling behavior and ensure loss probabilities remain unscaled.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
python/bloqade/lanes/noise_model.py	Switches noise parameter sourcing from raw fields to scaled getter APIs during squin kernel construction.
python/tests/test_noise_scaling.py	Adds regression tests that walk squin kernel IR constants to verify scaling behavior and loss non-scaling.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

☂️ Code Coverage

current status: ✅

Overall Coverage

Statements	Covered	Coverage	Threshold	Status
11684	10038	86%	0%	🟢

New Files

No new covered files...

Modified Files

File	Coverage	Status
python/bloqade/lanes/noise_model.py	59%	🟢
TOTAL	59%	🟢

updated for commit: ea9f8d7 by action🐍

[jasonhan3]

In general, for local_loss_prob, global_loss_prob, cz_gate_loss_prob, cz_unpaired_loss_prob, move_loss_prob, sit_loss_prob, I don't know why you don't call the corresponding getters defined in MoveNoiseModelABC ? It's mentioned in the PR description that you don't but it's unclear to me why, as the issue here: QuEraComputing/bloqade-circuit#836 has a loss_error_scale parameter.

[jasonhan3]

shouldn't this be using cz_paired_errors on MoveNoiseModelABC if we want to use the properties instead of the direct field?

[weinbe58]

That doesn't include the scale the API's are all messed up upstream

[jasonhan3]

Ah ok sounds good! So we'll probably update the downstream API's when the upstream ones are updated?

I guess, do we anticipate that the @Property API's here: https://github.com/QuEraComputing/bloqade-circuit/blob/06116fd450aaeefebb4769bc42290f60ece02671/src/bloqade/qasm2/dialects/noise/model.py would have the scale factor? If so we can proactively change it; if not then we can just leave it alone until upstream API's are changed

[jasonhan3]

same comment; shouldn't this be using cz_paired_errors on MoveNoiseModelABC if we want to use the properties instead of the direct field?

[weinbe58]

Note on scope: this PR only scales the Pauli rates, not the loss probabilities. That's a deliberate limitation of the current upstream API — GeminiNoiseModelABC.scaling_factor (and its getter properties) only apply the scale factor to the Pauli rates; there is no getter that returns loss probabilities with the scale factor applied. So lanes has no way to obtain scaled loss values without re-implementing the scaling itself, which would drift from the upstream source of truth.

Per-category loss scaling (loss_error_scale) is the genuinely-new axis tracked upstream in QuEraComputing/bloqade-circuit#836. Once that lands and exposes a scaled-loss getter, we can switch the loss reads here over to it.

[jasonhan3]

LGTM!