fix: use savepoints for row-by-row error isolation in Excel import

Prevents a single duplicate fall_id from rolling back the entire
import session. Each row insert now uses db.begin_nested() so
constraint violations are isolated to the offending row.

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

backend/app/services/excel_import.py

@@ -494,12 +494,22 @@ def import_abrechnung_sheet(
                 abrechnung_datum=abrechnung_datum,
                 import_source=f"Abrechnung_DAK.xlsx:{sheet_name}",
             )
-            db.add(case)
-            imported += 1
-
-            # Flush in batches of 100 to catch constraint violations early
-            if imported % 100 == 0:
+            # Use savepoint so a single row failure doesn't break the session
+            nested = db.begin_nested()
+            try:
+                db.add(case)
                 db.flush()
+                nested.commit()
+                imported += 1
+            except Exception as flush_err:
+                nested.rollback()
+                errors.append(f"Row {row_num} ({nachname}): {flush_err}")
+                logger.warning(
+                    "Import error in sheet '%s' row %d: %s",
+                    sheet_name, row_num, flush_err,
+                )
+                skipped += 1
+                continue
 
         except Exception as e:
             nachname_display = _str_or_none(_get(row, col_map, "nachname")) or "?"
@@ -509,10 +519,6 @@ def import_abrechnung_sheet(
                 sheet_name, row_num, e,
             )
 
-    # Final flush
-    if imported > 0:
-        db.flush()
-
     logger.info(
         "Sheet '%s': %d imported, %d skipped, %d errors",
         sheet_name, imported, skipped, len(errors),