Fix db inventory: capture guaranteed floor before complex queries

Run estimated_row_count first and save to `guaranteed` variable before
attempting table_statistics JOIN or pg_class — preventing SQLAlchemy
connection-poisoning from an early query failure from blocking the
name-only fallback.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

visualization/debug_page.py

@@ -102,13 +102,37 @@ _LADDER_TB2P_FIELDS = [
 def _query_db_inventory(conn) -> pd.DataFrame:
     """
     List all tables in the public schema with estimated row counts.
-    Uses CockroachDB auto-stats (table_statistics join) with pg_class fallback,
-    and a final-resort name-only fallback so the table always renders.
+    Runs the safe estimated_row_count query first to guarantee a renderable
+    floor result, then attempts richer sources (table_statistics, pg_class).
     Returns DataFrame with columns [table_name, row_count], sorted by table_name.
     """
     from sqlalchemy import text as _t
 
-    # Primary: join crdb_internal.table_statistics for latest collected row counts
+    # Step 1 — always capture a guaranteed floor first.
+    # estimated_row_count is the proven-safe query (worked in 8f09f14).
+    # Runs before any complex query that might corrupt the connection object.
+    guaranteed = pd.DataFrame(columns=["table_name", "row_count"])
+    try:
+        df = pd.read_sql(
+            _t("""
+               SELECT name AS table_name,
+                      estimated_row_count AS row_count
+               FROM crdb_internal.tables
+               WHERE schema_name = 'public'
+                 AND table_type = 'table'
+               ORDER BY name
+            """),
+            conn,
+        )
+        if not df.empty:
+            guaranteed = df
+            if df["row_count"].notna().any():
+                return df   # actual counts available — done
+    except Exception:
+        pass
+
+    # Step 2 — try table_statistics JOIN for richer counts.
+    # If this corrupts conn, guaranteed is already saved above.
     try:
         df = pd.read_sql(
             _t("""
@@ -134,7 +158,7 @@ def _query_db_inventory(conn) -> pd.DataFrame:
     except Exception:
         pass
 
-    # Fallback: pg_class reltuples (approximate, also fast)
+    # Step 3 — pg_class reltuples.
     try:
         df = pd.read_sql(
             _t("""
@@ -154,25 +178,9 @@ def _query_db_inventory(conn) -> pd.DataFrame:
     except Exception:
         pass
 
-    # Final resort: table names only, row_count will be NULL — better than empty
-    try:
-        df = pd.read_sql(
-            _t("""
-               SELECT name AS table_name,
-                      estimated_row_count AS row_count
-               FROM crdb_internal.tables
-               WHERE schema_name = 'public'
-                 AND table_type = 'table'
-               ORDER BY name
-            """),
-            conn,
-        )
-        if not df.empty:
-            return df
-    except Exception:
-        pass
-
-    return pd.DataFrame(columns=["table_name", "row_count"])
+    # Always return at least table names (guaranteed may be empty only if DB is
+    # fully unreachable, in which case the "check DB connection" warning is correct).
+    return guaranteed
 
 
 def _get_table_columns(conn, table_name: str) -> set: