fix(test): stabilize flaky LER tests in CI

[ivanbasov]

Summary

• Increase CI sample count for test_ler_improves_with_bd_noise_model from 2,000 to 20,000 so the boundary-detector LER comparison has enough statistical power
• Widen d=13 LER improvement tolerance from 1e-4 to 2e-4 in test_inference_d13_noise25p_ler_quality, raising the guard from ~2.5 sigma to ~5 sigma

Root cause

test_ler_improves_with_bd_noise_model (test_boundary_detectors.py):
With only 2,000 CI samples at p=0.002 (d=5), the test produced ~5 logical errors. Discrete error counts frequently coincided, causing assertLess to fail spuriously:

AssertionError: np.float64(0.0025) not less than np.float64(0.0025)

Observed in PR #2 CI run. Increasing to 20,000 samples yields ~50 errors — enough for reliable separation while adding negligible CI time (Stim sampling is fast).

test_inference_d13_noise25p_ler_quality (test_inference_public_model.py):
With d=13 at 262k shots and LER ~2e-4, the combined standard error for the before/after comparison is ~4e-5. The previous tolerance of 1e-4 (~2.5 sigma) gave a ~1.2% per-run flake probability:

AssertionError: 0.000320 not less than or equal to 0.000314 : Z: LER after (0.000320) > baseline (0.000214) + 0.0001

Observed in PR #6 CI run. Widening to 2e-4 (~5 sigma) effectively eliminates flakes.

Changed files

File	Change
code/tests/test_boundary_detectors.py	CI sample count 2k → 20k for test_ler_improves_with_bd_noise_model
code/tests/test_inference_public_model.py	LER_IMPROVEMENT_TOLERANCE 1e-4 → 2e-4

Test plan

• test_ler_improves_with_bd_noise_model keeps original assertLess — increased samples provide the statistical power
• test_inference_d13_noise25p_ler_quality tolerance widened with ~5-sigma headroom
• CI passes consistently