recon + sqlglot validator + drill_down package; guard ReAct dimension picks against candidate list

tests/test_yaml_datasets.py

@@ -3,6 +3,11 @@
 Bug captured: schema_docs.yaml had values that began with a quote character
 (e.g. `gender: 'M' or 'F'.`), which YAML treats as a quoted scalar — anything
 after the closing quote is a parse error.
+
+The YAML structure is now flat — tables.<name> is a description string,
+columns are keyed `table.column`, and relationships are an explicit list.
+The DDL structure (column types, nullability) lives in extracted_schema.json
+generated by `make recon`.
 """
 
 from pathlib import Path
@@ -20,6 +25,9 @@ def test_schema_docs_parses(dataset):
     assert docs is not None
     assert "tables" in docs
     assert isinstance(docs["tables"], dict) and docs["tables"]
+    # New shape: each table value is a description string, not a dict.
+    for name, val in docs["tables"].items():
+        assert isinstance(val, str), f"tables.{name} should be a string description, got {type(val)}"
 
 
 @pytest.mark.parametrize("dataset", DATASETS)
@@ -31,16 +39,24 @@ def test_metrics_parses(dataset):
 
 
 def test_financial_columns_with_quotes_survive_parse():
-    """Specific regression: values that contain quoted codes (M, F, OWNER,
-    PRIJEM, etc.) must parse without truncation."""
+    """Values containing quoted codes (M, F, OWNER, PRIJEM, ...) must parse
+    without truncation. Columns are flat `table.col` keys now."""
     docs = yaml.safe_load(
         (DATASETS_DIR / "financial" / "schema_docs.yaml").read_text()
     )
-    cols = docs["tables"]["client"]["columns"]
-    assert '"M"' in cols["gender"] and '"F"' in cols["gender"]
+    cols = docs["columns"]
+    assert '"M"' in cols["client.gender"] and '"F"' in cols["client.gender"]
+    assert '"OWNER"' in cols["disp.type"] and '"DISPONENT"' in cols["disp.type"]
+    assert '"PRIJEM"' in cols["trans.type"] and '"VYDAJ"' in cols["trans.type"]
 
-    disp = docs["tables"]["disp"]["columns"]
-    assert '"OWNER"' in disp["type"] and '"DISPONENT"' in disp["type"]
 
-    trans = docs["tables"]["trans"]["columns"]
-    assert '"PRIJEM"' in trans["type"] and '"VYDAJ"' in trans["type"]
+def test_financial_has_relationships():
+    docs = yaml.safe_load(
+        (DATASETS_DIR / "financial" / "schema_docs.yaml").read_text()
+    )
+    rels = docs.get("relationships") or []
+    assert len(rels) >= 5  # we have 8; assert ≥5 to allow some pruning
+    # all entries have from/to
+    for r in rels:
+        assert "from" in r and "to" in r
+        assert "." in r["from"] and "." in r["to"]