A skipped live check must not exit 0

README.md

@@ -121,9 +121,9 @@ connected through SSH local forwarding.
    return + facade, keyword-parameter rename, payload-key rename,
    env-var rename) are
    offline by default.
-   `--live` skips with exit 0 when Ollama is down; that skip is not a
-   quality pass. `--require-live` writes `eval-runs/live-status.json`
-   `{"skipped": true}` and exits 2 so CI can fail closed:
+   `--live` and `--require-live` write `eval-runs/live-status.json`
+   `{"skipped": true}` and exit 2 when Ollama is down. A skipped live
+   check is not success:
 
    ```bash
    .venv/bin/python scripts/prove_multifile_refactor.py

docs/evaluation-protocol.md

@@ -201,9 +201,9 @@ PYTHONPATH=src .venv/bin/python -m unittest discover -s tests -v
 GitHub Actions (`.github/workflows/tests.yml`) runs the same no-GPU path on every push.
 
 Live protocol (workstation with Ollama). If the server is down, `--live`
-prints `SKIP live` and exits 0. `--require-live` writes
-`eval-runs/live-status.json` `{"skipped": true}` and exits 2. That skip
-is not a model-quality pass:
+and `--require-live` print `SKIP live`, write
+`eval-runs/live-status.json` `{"skipped": true}`, and exit 2. A skipped
+live check is not success:
 
 ```bash
 .venv/bin/python scripts/run_eval.py --live --model fast

scripts/prove_multifile_refactor.py

@@ -1,10 +1,10 @@
 #!/usr/bin/env python3
 """Score the harder multi-file refactor corpus.
 
-Default path is offline fixtures (CI / no GPU). ``--live`` calls real
-``local_refactor`` through stdio MCP and skips with exit 0 when Ollama
-is down. ``--require-live`` is the same path but exits 2 on skip and
-writes ``eval-runs/live-status.json``. A skip is not a model-quality pass.
+Default path is offline fixtures (CI / no GPU). ``--live`` and
+``--require-live`` call real ``local_refactor`` through stdio MCP.
+When Ollama is down they write ``eval-runs/live-status.json`` and exit 2.
+A skipped live check is not success.
 """
 
 from __future__ import annotations

scripts/prove_refactor_acceptance.py

@@ -3,7 +3,7 @@
 
 Single-file whitespace extract only. For the harder multi-file corpus,
 use ``scripts/prove_multifile_refactor.py`` (offline fixtures by default;
-``--live`` skips when Ollama is down; ``--require-live`` exits 2).
+``--live`` and ``--require-live`` exit 2 when Ollama is down).
 """
 
 from __future__ import annotations

scripts/run_eval.py

@@ -91,9 +91,7 @@ def _run_fixtures(case_id: str | None, suite: str) -> int:
     return failed
 
 
-async def _run_live(
-    case_id: str | None, model: str, suite: str, *, require_live: bool
-) -> int:
+async def _run_live(case_id: str | None, model: str, suite: str) -> int:
     from local_coding_slm.ollama_client import OllamaSettings, is_reachable
     from local_coding_slm.server import _load_dotenv
 
@@ -103,9 +101,9 @@ async def _run_live(
         _write_live_status(skipped=True)
         print(
             f"SKIP live: Ollama unreachable at {settings.host_label()} "
-            "(offline fixtures still pass; this is not a model-quality fail)"
+            "(skipped live check is not success)"
         )
-        return 2 if require_live else 0
+        return 2
     _write_live_status(skipped=False)
 
     from mcp import ClientSession, StdioServerParameters
@@ -160,7 +158,7 @@ def main() -> None:
     parser.add_argument(
         "--require-live",
         action="store_true",
-        help="Same as --live, but exit 2 when Ollama is down instead of skip-0.",
+        help="Same as --live. A skipped live check exits 2.",
     )
     parser.add_argument("--model", choices=("fast", "strong"), default="fast")
     parser.add_argument("--case", dest="case_id", default=None)
@@ -178,7 +176,6 @@ def main() -> None:
                     args.case_id,
                     args.model,
                     args.suite,
-                    require_live=args.require_live,
                 )
             )
         )

spec.md

@@ -683,7 +683,7 @@ OLLAMA_BASE_URL
 │   ├── run_mcp.sh                   ← project MCP entry (loads .env)
 │   ├── prove_acceptance.py
 │   ├── prove_refactor_acceptance.py
-│   ├── prove_multifile_refactor.py  ← harder multi-file suite; --live skip-0, --require-live skip-2
+│   ├── prove_multifile_refactor.py  ← harder multi-file suite; skipped live exits 2
 │   ├── run_eval.py                  ← fixture corpus; --live on the workstation
 │   ├── run_harness.py               ← timed MCP campaign (stub or live)
 │   └── check_deployment_safety.py   ← defensive bind / tag / git checks

tests/test_eval_harder.py

@@ -93,7 +93,7 @@ def test_live_flag_skips_when_ollama_is_down(self) -> None:
             text=True,
             check=False,
         )
-        self.assertEqual(proc.returncode, 0, proc.stderr)
+        self.assertEqual(proc.returncode, 2, proc.stderr)
         self.assertIn("SKIP live", proc.stdout)
 
     def test_require_live_exits_2_when_ollama_is_down(self) -> None:

tests/test_fitness_skipped_live.py

@@ -0,0 +1,44 @@
+"""Fitness function: the harness must not treat a skipped live check as success."""
+
+from __future__ import annotations
+
+import json
+import os
+import subprocess
+import sys
+import unittest
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+class SkippedLiveFitnessTests(unittest.TestCase):
+    def test_skipped_live_check_is_not_success(self) -> None:
+        env = os.environ.copy()
+        env["PYTHONPATH"] = str(ROOT / "src") + os.pathsep + env.get("PYTHONPATH", "")
+        env["OLLAMA_BASE_URL"] = "http://127.0.0.1:1"
+        status_path = ROOT / "eval-runs" / "live-status.json"
+        if status_path.exists():
+            status_path.unlink()
+        proc = subprocess.run(
+            [
+                sys.executable,
+                str(ROOT / "scripts" / "run_eval.py"),
+                "--live",
+                "--case",
+                "whitespace_extract",
+            ],
+            cwd=str(ROOT),
+            env=env,
+            capture_output=True,
+            text=True,
+            check=False,
+        )
+        self.assertNotEqual(proc.returncode, 0, proc.stdout + proc.stderr)
+        self.assertIn("SKIP live", proc.stdout)
+        payload = json.loads(status_path.read_text(encoding="utf-8"))
+        self.assertEqual(payload, {"skipped": True})
+
+
+if __name__ == "__main__":
+    unittest.main()