test(preflight): pin #152's schema-type-aware label diversity in the parity harness

internal/push/parity_golden_test.go

@@ -23,15 +23,16 @@ import (
 // silent.
 
 type parityCase struct {
-	Name string `json:"name"`
-	Category string `json:"category"`
-	CSV string `json:"csv"`
-	LabelColumn string `json:"label_column"`
-	Extension string `json:"extension"`
-	TargetSize []int `json:"target_size"`
-	CLIVerdict string `json:"cli_verdict"`
-	IngestorVerdict string `json:"ingestor_verdict"`
-	Note string `json:"note"`
+	Name string `json:"name"`
+	Category string `json:"category"`
+	CSV string `json:"csv"`
+	LabelColumn string `json:"label_column"`
+	Extension string `json:"extension"`
+	TargetSize []int `json:"target_size"`
+	Schema map[string]string `json:"schema"`
+	CLIVerdict string `json:"cli_verdict"`
+	IngestorVerdict string `json:"ingestor_verdict"`
+	Note string `json:"note"`
 }
 
 func TestValidatorParity(t *testing.T) {
@@ -85,9 +86,12 @@ func runGoPreflight(t *testing.T, c parityCase) string {
  TargetSize: c.TargetSize,
 	}
 	if IsTabular(c.Category) {
- // Mirror runDataIngest: the schema is inferred from the CSV before
- // the preflight runs (the schema-columns preview needs it).
- if sch, _, _, err := InferSchema(layout.LabelsCSV); err == nil {
+ // Mirror runDataIngest: an explicit schema (the --schema flow) wins,
+ // else it is inferred from the CSV — the golden generator derives the
+ // schema the same way, so dtype-sensitive verdicts stay comparable.
+ if len(c.Schema) > 0 {
+ spec.Schema = c.Schema
+ } else if sch, _, _, err := InferSchema(layout.LabelsCSV); err == nil {
  spec.Schema = sch
  }
 	}

internal/push/testdata/parity/cases.json

@@ -281,6 +281,32 @@
  "cli_verdict": "accept",
  "ingestor_verdict": "accept",
  "note": "pins the text-family dispatch"
+ },
+ {
+ "name": "tabular-varchar-numeric-labels",
+ "category": "tabular_classification",
+ "csv": "data.csv",
+ "label_column": "label",
+ "schema": {
+ "age": "INT",
+ "label": "VARCHAR(255)"
+ },
+ "cli_verdict": "accept",
+ "ingestor_verdict": "accept",
+ "note": "labels '1' vs '1.0' under a VARCHAR label: the ingestor pins dtype=str (no numeric collapse) so they are 2 classes \u2014 pins #152's schema-type-aware collapse (the earlier blanket collapse falsely rejected this)"
+ },
+ {
+ "name": "tabular-float-numeric-labels",
+ "category": "tabular_classification",
+ "csv": "data.csv",
+ "label_column": "label",
+ "schema": {
+ "age": "INT",
+ "label": "FLOAT"
+ },
+ "cli_verdict": "reject",
+ "ingestor_verdict": "reject",
+ "note": "same values under a FLOAT label: numeric read collapses 1/1.0 into one class \u2014 both sides reject (the counterpart pin)"
  }
  ]
 }

internal/push/testdata/parity/cases/tabular-float-numeric-labels/data.csv

@@ -0,0 +1,3 @@
+age,label
+30,1
+40,1.0

internal/push/testdata/parity/cases/tabular-varchar-numeric-labels/data.csv

@@ -0,0 +1,3 @@
+age,label
+30,1
+40,1.0

internal/push/testdata/parity/goldens.json

@@ -90,6 +90,12 @@
  ],
  "verdict": "reject"
  },
+ "tabular-float-numeric-labels": {
+ "errors": [
+ "LabelDiversityValidator: Classification category requires at least 2 distinct label values in column 'label' (after whitespace stripping); this dataset has 1 distinct value(s): [np.float64(1.0)]. Raw value counts: {1.0: 2}. If this is intentional (e.g. you have a continuous target), pick a regression-family category like tabular_regression or time_series_forecasting instead."
+ ],
+ "verdict": "reject"
+ },
  "tabular-header-only": {
  "errors": [
  "IngestableRecordsValidator: No data rows found in CSV 'data.csv': the file has a header but no data rows (0 ingestable records). Add at least one data row and re-ingest.",
@@ -119,6 +125,10 @@
  "errors": [],
  "verdict": "accept"
  },
+ "tabular-varchar-numeric-labels": {
+ "errors": [],
+ "verdict": "accept"
+ },
  "text-clf-ok": {
  "errors": [],
  "verdict": "accept"

scripts/gen-validator-goldens.py

@@ -72,10 +72,15 @@ def run_case(case):
  if case.get("target_size"):
  options["target_size"] = case["target_size"]
  if case["category"].startswith(("tabular", "time_")):
- try:
- schema = infer_schema(csv_path)
- except Exception:
- schema = {}
+ # An explicit per-case schema (mirroring --schema) wins; else
+ # infer — BOTH sides of the harness use the same source so
+ # dtype-sensitive cases (label diversity) stay comparable.
+ schema = case.get("schema")
+ if not schema:
+ try:
+ schema = infer_schema(csv_path)
+ except Exception:
+ schema = {}
  options["schema"] = schema
  options["full_schema"] = schema