Widen center-invariant norms before computing squared deviations - #776

Problem

CenterInvariantRegularizer computes norms and squared deviations in the input dtype. For bfloat16 centers [16, 0] and [17, 0], this gives 64 instead of the reference loss 68.0625. Float16 norms around 300 also overflow when squared, before the final centered loss is computed.

Changes

Promote low-precision centers before computing their norms and retain float32 through reduction. Float32/float64 inputs and custom Lp norms retain their existing behavior. Tests check the squared-norm variance and its analytic gradient against float64 references; the integration test reference now also uses float64.

The component documentation, numerical rationale, and unreleased notes describe the behavior.

Validation

• Python 3.12.14 and PyTorch 2.15.0+cpu.
• TEST_DEVICE=cpu TEST_DTYPES=float32,float64,bfloat16 OMP_NUM_THREADS=1 python -m pytest tests/regularizers/test_center_invariant_regularizer.py -q: 10 passed.
• Ruff lint/format checks and git diff --check passed.

The checksum-verified pinned Code Polishy 0.31.10 commands were attempted. check --git-changes is blocked by an unavailable policy-owned CPython 3.12 interpreter/security providers. test --changed --base 0db540f1b8adb0f65ea91075ee1625e13176758b could not start the declared suites because the repository-specific .venv/bin/uv is absent. The direct focused tests below ran in the available Python/PyTorch environment; these results do not constitute a pass of the complete locked verification gates. Float16 execution is also blocked on this host by RuntimeError: Failed to initialize cpuinfo!; its parameterized cases are retained for supported CI hosts.