Mutation in generator; category H (stop) scoring and judge; category K variants and tool schema variant; evalset plan

- mutation.py: negation mutants, skip WHILE/DO blocks (endless loop blocked RFC ~8 min)
- generator: accept only after mutation check; survivors go back as repair feedback
- runner/judge.py: stop tasks scored 100/30/0; gap by keywords, else judge model
- proxy: tool schema variant generic_v0 (draft); generator make_k_variant
- evalset.py: 88 slots for step F (A-I, H, K)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

harness/proxy.py

@@ -12,6 +12,24 @@ MODEL_TOOLS = {
 }
 WRITE_TOOLS = {"sap_create_object", "sap_push_source", "sap_push_element",
                "sap_push_message", "sap_activate"}
+# Tool schema variants (category K; small variation in training). Draft until abap-mcp-arayuz.md exists.
+# The model sees the variant names and argument names; the proxy translates them to EPOD.
+VARIANTS = {
+    "generic_v0": {
+        "tools": {"search_object": "sap_search_object", "read_source": "sap_pull_source",
+                  "object_structure": "sap_object_structure", "where_used": "sap_usage_references",
+                  "create_object": "sap_create_object", "write_source": "sap_push_source",
+                  "activate": "sap_activate", "syntax_check": "sap_syntax_check",
+                  "run_unit_tests": "sap_run_unit_test", "run_atc": "sap_atc_run",
+                  "object_members": "sap_object_members", "element_info": "sap_element_info",
+                  "inactive_objects": "sap_inactive_objects", "short_dumps": "sap_short_dumps",
+                  "sql_query": "sap_sql_query", "write_element": "sap_push_element",
+                  "write_message": "sap_push_message", "check_object": "sap_check_object",
+                  "pretty_print": "sap_pretty_print", "run_class": "sap_run_class"},
+        "args": {"objectName": "name", "objectType": "type", "functionGroup": "function_group",
+                 "packageName": "package", "includeType": "include", "className": "class_name"},
+    },
+}
 RUN_PREFIX = re.compile(r"^Z\d[0-9A-Z]{5,6}_", re.I)
 
 
@@ -20,8 +38,9 @@ class BudgetExceeded(Exception):
 
 
 class ToolProxy:
-    def __init__(self, mcp, prefix, budget, log_path):
+    def __init__(self, mcp, prefix, budget, log_path, tool_schema=None):
         self.mcp = mcp
+        self.variant = VARIANTS.get(tool_schema) if tool_schema else None
         self.prefix = prefix.upper()
         self.max_calls = budget.get("max_tool_calls", 60)
         self.max_activations = budget.get("max_activations", 15)
@@ -32,7 +51,31 @@ class ToolProxy:
         self.log = open(log_path, "a")
 
     def schemas(self):
-        return [t for t in self.mcp.list_tools() if t["name"] in MODEL_TOOLS]
+        tools = [t for t in self.mcp.list_tools() if t["name"] in MODEL_TOOLS]
+        if not self.variant:
+            return tools
+        back = {v: k for k, v in self.variant["tools"].items()}
+        amap = self.variant["args"]
+        out = []
+        for t in tools:
+            if t["name"] not in back:
+                continue
+            sch = json.loads(json.dumps(t.get("inputSchema", {})))
+            sch["properties"] = {amap.get(k, k): v for k, v in sch.get("properties", {}).items()}
+            if "required" in sch:
+                sch["required"] = [amap.get(k, k) for k in sch["required"]]
+            desc = t.get("description", "")
+            for epod, alias in list(back.items()) + list(amap.items()):
+                desc = re.sub(rf"\b{re.escape(epod)}\b", alias, desc)
+            out.append(dict(t, name=back[t["name"]], description=desc, inputSchema=sch))
+        return out
+
+    def _translate(self, tool, args):
+        """Variant name and arguments to EPOD. EPOD names stay valid (the oracle uses them)."""
+        if not self.variant or tool not in self.variant["tools"]:
+            return tool, args
+        back = {v: k for k, v in self.variant["args"].items()}
+        return self.variant["tools"][tool], {back.get(k, k): v for k, v in (args or {}).items()}
 
     def _foreign(self, name):
         n = (name or "").upper()
@@ -52,6 +95,10 @@ class ToolProxy:
 
     def call(self, tool, args):
         entry = {"t": time.time(), "tool": tool, "args": args}
+        shown = tool
+        tool, args = self._translate(tool, args)
+        if shown != tool:
+            entry.update(tool=tool, args=args, variant_tool=shown)
         if tool not in MODEL_TOOLS:
             result = (True, f"Tool {tool} is not available.")
         elif self.calls >= self.max_calls: