Module defect_validation 

defect-validation analysis — surfaces a defect-calibration artifact’s own evidence as flat, human-scannable (metric, value) rows.

Unlike every other registered analysis, this one reads NOTHING from the fact store: all of its numbers were already computed at mining time and frozen into a defects.calib.json artifact (see crate::defect_calibration). It resolves that artifact from opts.defect_calibration, applies the SAME repo-identity guard the weight application uses (check_repo_identity, honouring --allow-foreign-calibration), and flattens its ValidationMetrics + TuningDecision + MiningStats into rows.

Presentation follows the project’s honesty framing: the band table shows where mined defects landed relative to code-health bands, but this is an association, not causation — a red file that a defect-introducing commit touched is evidence the score ranks it high, not proof the score caused the defect. Every count carries its n; every absent Option metric renders as an explicit n/a (<why>) rather than being silently dropped. The tuning-decision AUCs are always surfaced — including the case where tuning was applied yet the validation AUCs sit below 0.5, so a reader can judge the outcome for themselves.

Without a configured artifact the analysis returns zero rows and prints a one-line hint on stderr pointing at codelore calibrate-defects — an honest absence, not an error.

Structs

DefectValidationRow

One flattened evidence row from a defect-calibration artifact: (metric, value). The value is a string so counts, ratios, honest n/a (<why>) placeholders, and textual verdicts (e.g. the weights source) share one row shape — mirroring ArchitectureMetricRow.

Functions

run_defect_validation

Run the defect-validation analysis.