fix(lib): isolate build-mode store caches and tolerate non-numeric cells in release filters

[SkyeAv]

Summary

tablassert build-kg --release shipped applied_to_treat edges with number_of_cases < 25 even though the release-mode case-count filter exists. Reproduced against installed 16.6.0: a 1-case edge shipped in a --release KG. Two independent bugs, both fixed here.

Bug 1 — section-parquet cache key ignores build mode

In build_graph_pipeline (stage 3), each section's cached subgraph parquet was named STORE / f"{mkhash(s)}.parquet" where s is only the section config dict; release was injected into Tcode.model_validate({**s, ..., "release": release}) outside the hashed dict. Tcode.collect() then quick-exits on any existing store file (if self.store.is_file(): return self.store), skipping the entire op chain — including the release filters wired in _source_ops. So a parquet cached by a non-release build (or a pre-16.6.0 build, where the filter silently no-op'd per #137) was reused verbatim in --release builds: cache poisoning shipped the 1-case edge.

--head already dodged this class of bug with a distinct .head.parquet name. The fix generalizes that idiom: a small _section_store_path(h, head, release, qc) helper gives each build-mode flag its own suffix (<h>.parquet, <h>.head.parquet, <h>.release.parquet, <h>.qc.parquet), composing when flags combine (<h>.head.release.qc.parquet).

--qc is included in the fix, not deferred: fullmap_audit drops QC rejects from the output LazyFrame (qc.py:138-139), so a --qc build's cached parquet genuinely differs in content from a non-qc one — the same cache-poisoning class, fixed for free by the same helper.

Bug 2 — strict cast crashes on header rows

drop_low_number_of_cases and drop_zero_effect_size filtered with pl.col(col).cast(pl.Float64) (strict). The csv op reads sources with has_header=False, so a TSV's header row flows through as a data row and the strict cast raised polars.exceptions.InvalidOperationError on the non-numeric header cell (reproduced: conversion from str to f64 failed in column 'number_of_cases' ... ["case_count"]).

Both ops now cast with strict=False — the clean_numeric idiom. Non-numeric cells become null and are kept (ne_missing(0.0) for effect size; .ge(threshold).fill_null(True) for case counts, since polars 1.40 has no ge_missing): header rows die later at entity resolution as they did before the filters existed. This tolerates arbitrary non-numeric cells, including the pending downstream TSV rename of the case_count header to number_of_cases. drop_not_significant casts to String and was never affected.

Test coverage

• test_drop_low_number_of_cases_keeps_non_numeric_cells — header-cell row and nulls kept, "10" dropped, "30" kept.
• test_drop_zero_effect_size_keeps_non_numeric_cells — same tolerance.
• test_section_store_path_suffixes_every_build_mode — every flag combination gets a distinct, composing cache name.
• test_tcode_collect_quick_exit_never_serves_another_mode_cached_parquet — a non-release cached parquet never quick-exits a release build; once the release parquet exists, the quick exit serves it.

uv run pytest tests/test_lib.py -k "release or number_of_cases or effect_size or store or not_significant or head_mode" → 32 passed; tests/test_cover_cli.py → 52 passed; ruff check/format and pyright clean.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Team

Run ID: d4438638-3d95-4642-8a18-8831aaa03074

---

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.}