fix(lib): emit p-value columns in scientific notation again

[SkyeAv]

Restores the controlled scientific-notation output for p-value columns: the KGX-compliance rework (#71) switched format_numeric() to real JSON numbers for p_value / adjusted_p_value, silently retiring the {:.4e} branch for exactly the columns it existed for — edges shipped "p_value":0.0001 instead of the documented "p_value":"1.0000e-03".

What was broken

• Dead notation branch: format_numeric() consulted numeric_slot_kind() first, which reports "float" for p_value / adjusted_p_value (declared Optional[float] on the base Association class in the pinned biolink-model), so the float-kind arm short-circuited with a cast-only path. The {:.4e} formatting could only fire for p-value columns without a numeric slot — which don't exist.
• Result: shortest-repr numbers in whatever notation polars/serde chose (0.0001, 0.05, 6.52e-6) instead of controlled four-significant-figure scientific notation.

Fix

• src/tablassert/lib.py: any p-value-like column (same substring match as numeric_columns) is now formatted {:.4e} before the numeric-slot short-circuit; branch ordering is the guard. Non-p-value numeric columns (effect_size, supporting_study_size) keep the model-typed behavior: a real JSON number once a future biolink-model types the slot (#1770 / #1774), controlled {:.4g} strings until then.
• KGX validity preserved: validate_record() / validate_kgx() validate in Pydantic lax mode, which coerces the numeric string back for the float slot — verified empirically against the pinned model classes before choosing this fix.

Tests

• New guard in tests/test_biolink.py: the validate_kgx matrix gains an edge with "p_value":"1.0000e-03" that must count as fully valid (while "not-a-number" still fails).
• tests/test_lib.py: p-value formatting tests assert scientific-notation strings; the null-stripping test pins a zero p-value surviving strip_nulls as "0.0000e+00".
• tests/test_e2e_smoke.py: the real-pipeline coercion smoke asserts "p_value" == "1.0000e-02" on the emitted edge.

Docs

• docs/tutorial.md: the example edge output is accurate again — including dropping the flat supporting_study_size that never reaches the edge on biolink-model 4.4.3 (it rides the inlined supporting study).
• CHANGELOG.md: Unreleased / Fixed entry documenting the regression and restoration; format_numeric() / numeric_slot_kind() docstrings updated.

Risks / caveats

• Accepted tension with fix(biolink): make KGX output validate against the Biolink Model #71: item 8 of that PR moved p-values to real numbers for type fidelity; this deliberately trades strict-typed emission back for notation control. Pydantic-lax validation (our validate-kgx, the agent self-audit) stays green; a hypothetical strict-mode downstream consumer would coerce the string.

Testing

• uv run ruff check . → clean; uv run ruff format --check . → clean; uv run pyright → 0 errors, 0 warnings.
• uv run pytest -q → 898 passed, 35 skipped (identical count to pristine main).

[coderabbitai]

Important

Review available on request

• 🔍 Trigger review

Reviews should be triggered manually for repositories with fewer than 10 stars. Select Trigger review above or comment @coderabbitai review to review the latest changes. For a full review, comment @coderabbitai full review.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 2ac5e607-dd65-43ed-bac8-59cea394ee72

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}