Module validate 

Historical band scan, validation report, and constrained weight tuning — Units C and D of the own-repo defect-calibration design.

Unit C — the historical band scan + validation report

band_history reuses health_trend’s at-rev machinery (sampled_commits → live_paths_at → ingest_complexity_at_rev + materialize_imports_at_rev → a history_cutoff-scoped run_code_health_scoped) to recompute code-health bands at ≤12 evenly spaced historical revisions — but, unlike health_trend::file_series, with NO top-50 path cap: the validation report needs every scored file, since any of them might be the one a mined defect touched.

validate matches each SzzLink’s defect-introducing commit to the nearest band sample at-or-before its date, tallies the headline band table, and scores HEAD’s structural_risk against the defect-implicated file labels via stats::{auc, precision_at_k}.

Unit D — constrained weight tuning

tune_weights runs a deterministic coordinate-descent search over the eight SMELL_WEIGHTS and only adopts a tuned set when the evidence clears an honesty floor: fewer than 30 linked defect-changes, fewer than 10 implicated files, a tuned validation AUC below random (0.5), or a validation-AUC improvement short of the acceptance margin, each keep the defaults instead of the tuned weights.

Design decision: re-scoring without re-running the SQL pass

Evaluating a candidate weight set’s training/validation AUC needs a structural_risk for every file under THAT candidate — but re-running run_code_health_scoped (the full SQL pass, including the coupling/god-class/clone sub-scans) once per candidate is far too expensive across a coordinate-descent search. Instead:

1. capture_intensities reads the code_health_biomarkers_v1 session temp table’s rows ONCE, immediately after a single HEAD run_code_health_scoped call, in the SAME connection/session (that table is deliberately left queryable after the scan returns — see its own doc comment in code_health.rs), into a HashMap<path, [f64; 8]> keyed in SMELL_WEIGHTS order.
2. Each candidate’s risk is then computed Rust-side by structural_risk_from_intensities: Σ wᵢ·intensityᵢ clamped to 1.0 — mirroring the SQL LEAST(1.0, SUM(intensity * CASE …)) formula exactly (the HEAD, include_clones = true path, which carries no renormalization divisor).

The two are unit-tested for parity (tolerance 1e-9) against a real run’s structural_risk on the biomarker fixture in tests/defect_calibration_test.rs.

Functions

band_history

Uncapped per-sample band maps: for each of ≤12 sampled revs (the same evenly-spaced history sampler health_trend uses, oldest-first, newest always included), the file→band map at that point in history.

capture_intensities

Capture the per-file, per-smell biomarker intensities tune_weights needs to re-score candidate weight sets, without re-running the whole code-health SQL pass for every candidate — see this module’s design- decision doc comment above.

default_weights

The [SMELL_WEIGHTS] defaults converted to owned (name, weight) tuples, in SMELL_WEIGHTS order — the canonical defaults argument for tune_weights and the reference weight vector structural_risk_from_intensities is parity-tested against.

structural_risk_from_intensities

Rust-side mirror of code_health’s SQL formula LEAST(1.0, SUM(intensity * weight)) — specifically the HEAD (include_clones = true) path, which carries no renormalization divisor (STRUCTURAL_SCALE_NO_DRY only applies when DRY is excluded, which never happens at HEAD). weights must be in the same fixed order as intensities’s 8 slots — default_weights’s order (SMELL_WEIGHTS order) — for the parity to hold; see the parity test in tests/defect_calibration_test.rs.

tune_weights

Constrained deterministic coordinate search over the eight SMELL_WEIGHTS (Unit D). intensities are per-file 8-smell intensity vectors in SMELL_WEIGHTS order (see capture_intensities); train/validation are (path, label) splits already partitioned by fix date (60/40, older/newer — a leakage guard against a random split, prepared by the caller); defaults are the weights to fall back to and the search’s starting point (see default_weights).

validate

Band table + AUC + precision@k against defect labels.