Stabilize Bridge numerical CI tests

[emerardd]

Description

Stabilize two TransformerBridge integration assertions that cross the numerical envelope of different GitHub-hosted CPU runners. This is independent of the Qwen gated-query change in #1653; that PR exposed the existing baseline failures through its full-coverage run.

The left-padding neighbour test previously compared a three-row batch against separate one-row forwards. Those different matrix shapes can select different CPU/BLAS kernels and produced a 9.1552734375e-05 maximum absolute difference despite equivalent position handling. The test now compares two equal-shaped batches that differ only in whether the middle row requires derived positions, isolating the behaviour the regression test is intended to protect.

The optimizer parity test used a step-1 post-update mean-logit threshold of 1e-3. Multiple GitHub CPU runners reproducibly report 0.0010319872526451945, so this one empirical threshold is recalibrated to 2e-3; the initial-forward, maximum-logit, loss, parameter-update, and step-10 thresholds remain unchanged.

The exact pair of failures appeared in these unrelated dev-4.x runs:

• https://github.com/TransformerLensOrg/TransformerLens/actions/runs/31398580334
• https://github.com/TransformerLensOrg/TransformerLens/actions/runs/31395356122
• https://github.com/TransformerLensOrg/TransformerLens/actions/runs/31233076384

The current #1653 failure is:

• https://github.com/TransformerLensOrg/TransformerLens/actions/runs/31577290649

The same base SHA also completed successfully in:

• https://github.com/TransformerLensOrg/TransformerLens/actions/runs/31517217634

No production code or dependencies are changed.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Screenshots

Not applicable.

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

The documentation change is limited to the affected test docstring and explanatory comment; no user-facing API documentation changes are needed. The full unit suite was not rerun because this PR changes only two integration tests.

Validation

• Exact two previously failing tests: 2 passed
• Both complete integration test files with CI-style -n 2 --dist loadscope, OMP_NUM_THREADS=1, and MKL_NUM_THREADS=1: 31 passed
• Full-repository pycln, isort, and Black checks pass
• uv run mypy .: Success: no issues found in 431 source files
• git diff --check passes

[jlarson4]

Looks great @emerardd, thanks for cleaning this up!