Guard loss() and pct_close() denominators and catch Inf, not just NaN

[MaxGhenis]

Summary

Finding #7 (MED). Both functions divide by target + 1, which crosses zero at target == -1 — producing Inf and silently corrupting the loss. The NaN guard in loss() was specifically torch.isnan, so Inf passed through undetected and calibration diverged without any diagnostic. Any other upstream non-finite value (e.g. an Inf estimate from a broken estimate function) slipped past the same check.

Factored out _safe_denominator(targets) that clamps target + 1 to eps = 1e-12 so the value is unchanged for any well-behaved target but never zero. Switched the guard to torch.isfinite(...).all() so Inf is rejected alongside NaN, with an error message pointing at likely causes.

Test plan

• Add tests/test_metrics_regression.py: target == -1 no longer produces Inf, explicit Inf estimates now raise, value preservation for normal targets, and pct_close gets the same guard.
• All existing tests pass (uv run pytest tests -x -q -> 19 passed).

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
microcalibrate	Ready	Preview, Comment	Apr 17, 2026 0:40am

[MaxGhenis]

LGTM (cannot self-approve; posting as comment).

• _safe_denominator(targets) clamps targets + 1 to eps = 1e-12, preventing divide-by-zero when target == -1. For any well-behaved target (target + 1 >= eps) the numeric result is unchanged — verified by test_loss_preserves_values_for_well_behaved_targets.
• Guard broadened from torch.isnan(rel_error).any() to not torch.isfinite(rel_error).all(), so Inf is now caught (was silently passing through before). Error message is informative.
• pct_close uses the same _safe_denominator, so the metric stays finite on pathological targets.

Scope note (not blocking): finding #7 also flagged the "magnitude asymmetry" of the +1 smoothing for sub-unit targets. This PR deliberately preserves the original +1 semantics and only guards the denominator floor — a narrow, conservative fix. Revisiting the principled scale (e.g. max(|target|, eps)) is a bigger behaviour change best handled separately. Leaving as-is is fine.