test: add Gate 3 numerical oracle contracts

tests/test_cli.py

@@ -8,7 +8,10 @@ downloading real models or running any pipeline.  They use
 
 from __future__ import annotations
 
+import json
 from io import StringIO
+from types import SimpleNamespace
+from unittest.mock import MagicMock
 from unittest.mock import patch
 
 import pytest
@@ -155,6 +158,203 @@ class TestCLIDispatch:
             assert args_passed.contribute is True
             assert args_passed.contribute_notes == "Testing contribution system"
 
+    @pytest.mark.parametrize("flag", ["--prompt-pairs-file", "--prompt-pair-file"])
+    @pytest.mark.parametrize("command", ["obliterate", "abliterate"])
+    def test_prompt_pairs_file_flag_is_available_on_obliterate_and_alias(
+        self,
+        command,
+        flag,
+        tmp_path,
+    ):
+        """Explicit prompt-pair files are parsed for both commands and flag spellings."""
+        path = tmp_path / "pairs.json"
+        path.write_text("{}", encoding="utf-8")
+        with patch("obliteratus.cli._cmd_abliterate") as mock_cmd:
+            main([command, "fake/model", flag, str(path)])
+            args_passed = mock_cmd.call_args[0][0]
+            assert args_passed.prompt_pairs_file == str(path)
+
+    @pytest.mark.parametrize("flag", ["--prompt-pairs-file", "--prompt-pair-file"])
+    @pytest.mark.parametrize("command", ["obliterate", "abliterate"])
+    def test_prompt_pairs_file_is_mutually_exclusive_with_residue_files(
+        self,
+        command,
+        flag,
+        tmp_path,
+    ):
+        """Explicit prompt-pair files and mined residue construction cannot be mixed."""
+        path = tmp_path / "pairs.json"
+        path.write_text("{}", encoding="utf-8")
+        stderr_text = _capture_exit(
+            [
+                command,
+                "fake/model",
+                flag,
+                str(path),
+                "--residue-file",
+                "audit.json",
+            ],
+            expect_code=2,
+        )
+        assert "not allowed with argument" in stderr_text.lower()
+
+    @pytest.mark.parametrize(
+        ("option", "value"),
+        [
+            ("--dataset", "custom"),
+            ("--residue-weight", "7"),
+            ("--residue-max", "3"),
+        ],
+    )
+    @pytest.mark.parametrize("flag", ["--prompt-pairs-file", "--prompt-pair-file"])
+    @pytest.mark.parametrize("command", ["obliterate", "abliterate"])
+    def test_prompt_pairs_file_rejects_residue_only_options(
+        self,
+        command,
+        flag,
+        option,
+        value,
+        tmp_path,
+    ):
+        """Explicit prompt-pair files cannot be mixed with residue-only options."""
+        path = tmp_path / "pairs.json"
+        path.write_text("{}", encoding="utf-8")
+        stderr_text = _capture_exit(
+            [command, "fake/model", flag, str(path), option, value],
+            expect_code=2,
+        )
+        assert option in stderr_text
+        assert "can only be used with --residue-file" in stderr_text
+
+
+def test_cmd_abliterate_wires_prompt_pairs_file_into_pipeline(tmp_path):
+    """Loaded prompt-pair files are passed directly to AbliterationPipeline."""
+    path = tmp_path / "pairs.json"
+    path.write_text(
+        json.dumps(
+            {
+                "harmful": [f"harm {index}" for index in range(5)],
+                "harmless": [f"safe {index}" for index in range(5)],
+            }
+        ),
+        encoding="utf-8",
+    )
+    result_path = tmp_path / "result"
+    result_path.mkdir()
+    pipeline = MagicMock()
+    pipeline.run.return_value = str(result_path)
+
+    class FakeLive:
+        def __init__(self, *_args, **_kwargs):
+            self.update = MagicMock()
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *_args):
+            return False
+
+    args = SimpleNamespace(
+        model="org/model",
+        output_dir=str(tmp_path / "out"),
+        device="cpu",
+        dtype="float32",
+        method="basic",
+        n_directions=1,
+        direction_method=None,
+        regularization=None,
+        refinement_passes=1,
+        min_layer_fraction=None,
+        max_layer_fraction=None,
+        harmless_pc_count=None,
+        shield_concept_count=None,
+        shield_ridge=None,
+        shield_residualize=None,
+        shield_layer_penalty=None,
+        projection_target=None,
+        projection_row_fraction=None,
+        quantization=None,
+        gpu_memory_utilization=None,
+        large_model=False,
+        verify_sample_size=1,
+        refusal_max_tokens=1,
+        residue_file=[],
+        dataset="builtin",
+        residue_weight=5,
+        residue_max=None,
+        prompt_pairs_file=str(path),
+        contribute=False,
+        contribute_notes="",
+    )
+    with (
+        patch("rich.live.Live", FakeLive),
+        patch("obliteratus.abliterate.AbliterationPipeline", return_value=pipeline) as factory,
+        patch("obliteratus.telemetry.maybe_send_pipeline_report"),
+    ):
+        from obliteratus import cli
+
+        cli._cmd_abliterate(args)
+
+    assert factory.call_args.kwargs["harmful_prompts"] == [f"harm {index}" for index in range(5)]
+    assert factory.call_args.kwargs["harmless_prompts"] == [f"safe {index}" for index in range(5)]
+
+    args.prompt_pairs_file = None
+    pipeline.reset_mock()
+    pipeline.run.return_value = str(result_path)
+    with (
+        patch("rich.live.Live", FakeLive),
+        patch("obliteratus.abliterate.AbliterationPipeline", return_value=pipeline) as factory,
+        patch("obliteratus.telemetry.maybe_send_pipeline_report"),
+    ):
+        from obliteratus import cli
+
+        cli._cmd_abliterate(args)
+
+    assert "harmful_prompts" not in factory.call_args.kwargs
+    assert "harmless_prompts" not in factory.call_args.kwargs
+
+
+def test_cmd_abliterate_reports_invalid_prompt_pairs_file(tmp_path):
+    from obliteratus import cli
+
+    args = SimpleNamespace(
+        model="org/model",
+        output_dir=str(tmp_path / "out"),
+        device="cpu",
+        dtype="float32",
+        method="basic",
+        n_directions=1,
+        direction_method=None,
+        regularization=None,
+        refinement_passes=1,
+        min_layer_fraction=None,
+        max_layer_fraction=None,
+        harmless_pc_count=None,
+        shield_concept_count=None,
+        shield_ridge=None,
+        shield_residualize=None,
+        shield_layer_penalty=None,
+        projection_target=None,
+        projection_row_fraction=None,
+        quantization=None,
+        gpu_memory_utilization=None,
+        large_model=False,
+        verify_sample_size=1,
+        refusal_max_tokens=1,
+        residue_file=[],
+        dataset="builtin",
+        residue_weight=5,
+        residue_max=None,
+        prompt_pairs_file=str(tmp_path / "missing.json"),
+        contribute=False,
+        contribute_notes="",
+    )
+
+    with pytest.raises(SystemExit) as exc:
+        cli._cmd_abliterate(args)
+
+    assert exc.value.code == 2
+
 
 class _EncodingOnlyStdout:
     """Minimal stream stand-in for encoding-selection tests."""