Validation fixes: pandas-3 band check (#128), relative-path annotation (#129), and the builder layer in CI

[mmcky]

Both regressions were filed by the #127 validation run and are reproduced and fixed here, together with the CI gap that let the first one sit unseen for a week.

The two fixes

#128. builders/business_cycle.py checked value bands with values.stack().between(lo, hi).all(). pandas 3's stack() keeps NaN, so the structural nulls the manifest places (every economy's YR1960, GBR's and FRA's leading years, an unpublished newest year) failed between() and all three World Bank tables were rejected under the lectures' pandas. The line dates from #114; #126 kept it. Fix: .dropna() before the band check, with a comment naming the trap, and a note in _template.py so the next builder does not inherit it.

#129. scripts/validate_datasets.py resolved REPO but used the CLI's paths as given, so a relative manifest path that failed crashed on relative_to(REPO) instead of printing the annotation. Fix: resolve the arguments once. A relative-path run now prints ::error file=lectures/lingcod_msy_recovery.csv.yml::F_over_Fmsy: 1 nulls, manifest says exactly 2.

Closing the gap

The workflow's own comment said it was "the standing check that it stays green on both" pandas majors. It ran one major and one layer. Now:

• each dynamic builder exposes check_committed(): its own validate() on the committed bytes, no network (business_cycle.py, business_cycle_fred.py, and the template);
• validate_datasets.py --builders runs them, once per builder even when several manifests name the same one;
• validate-datasets.yml is a matrix over pandas 2.3.3 and 3.x and runs both layers in each.

Verified locally: with the pre-fix band line, --builders under pandas 3 fails with ::error file=builders/business_cycle.py::ValidationError: gdp_growth_annual.csv: value out of band [-50, 50]; with the fix, both layers pass under 2.3.3 and 3.0.5 (44 of 44 manifests; 4 committed snapshots through their builders).

Closes #128. Closes #129.

🤖 Generated with Claude Code

[Copilot]

🟡 Changes recommended

The new --builders validation path has a robustness bug that can crash without emitting annotations for malformed manifests, and the CI pandas-3 install is unpinned in a way that can cause non-reproducible failures.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR fixes two validator regressions discovered during the #127 independent validation run and strengthens CI so builder-specific validation runs across both pandas majors (shared schema validator + each dynamic builder’s own committed-bytes checks).

Changes:

• Fix builders/business_cycle.py band validation under pandas 3 by dropping NaNs before the between() check.
• Fix scripts/validate_datasets.py handling of relative CLI manifest paths and add --builders to run each committed dynamic builder’s check_committed() validation.
• Update CI (validate-datasets.yml) to run both validation layers under a pandas 2.3.3 and pandas 3.x matrix; document the new layer in builder docs.

File summaries

File	Description
scripts/validate_datasets.py	Adds --builders builder-layer validation and fixes relative-path argument resolution.
builders/business_cycle.py	Fixes pandas 3 stack() NaN behavior in band checks; adds check_committed().
builders/business_cycle_fred.py	Adds check_committed() for builder-layer validation.
builders/_template.py	Documents the pandas-3 stack()/NaN pitfall and adds a check_committed() template.
.github/workflows/validate-datasets.yml	Runs validation as a pandas-version matrix and executes both shared and builder layers.
builders/README.md	Documents that CI runs check_committed() for dynamic builders too.
AGENTS.md	Documents the check_committed() requirement and how CI uses it.

Review details

• Files reviewed: 7/7 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.