FF.FFFFF..FF.s                                                           [100%]
=================================== FAILURES ===================================
__________________ test_the_default_cue_is_the_instructed_one __________________

    def test_the_default_cue_is_the_instructed_one() -> None:
>       assert schema.DEFAULT_CUE == "json_instructed"
               ^^^^^^^^^^^^^^^^^^
E       AttributeError: module 'ggufone.schema' has no attribute 'DEFAULT_CUE'

tests/test_policy_v2.py:125: AttributeError
________________ test_the_default_chat_format_is_the_role_split ________________

    def test_the_default_chat_format_is_the_role_split() -> None:
>       assert schema.DEFAULT_CHAT_FORMAT == schema.ROLE_SPLIT
               ^^^^^^^^^^^^^^^^^^^^^^^^^^
E       AttributeError: module 'ggufone.schema' has no attribute 'DEFAULT_CHAT_FORMAT'

tests/test_policy_v2.py:135: AttributeError
__________ test_the_bench_and_the_cli_ask_for_the_v2_cell_by_default ___________

    def test_the_bench_and_the_cli_ask_for_the_v2_cell_by_default() -> None:
        """The bench's own literals (it never imports `schema`) must be the schema's defaults."""
>       assert harness.DEFAULT_CUE == schema.DEFAULT_CUE
                                      ^^^^^^^^^^^^^^^^^^
E       AttributeError: module 'ggufone.schema' has no attribute 'DEFAULT_CUE'

tests/test_policy_v2.py:151: AttributeError
_______________ test_the_reproduce_tool_defaults_to_the_v2_cell ________________

    def test_the_reproduce_tool_defaults_to_the_v2_cell() -> None:
        """The tool that regenerates every published row: no flags means the product's own policy."""
        import importlib.util
    
        path = ROOT / "tools" / "e2_reproduce.py"
        spec = importlib.util.spec_from_file_location("e2_reproduce_policy_v2", path)
        assert spec and spec.loader
        module = importlib.util.module_from_spec(spec)
        spec.loader.exec_module(module)
        args = module.make_parser().parse_args(["--suite", "quality", "--model", "m.gguf"])
>       assert args.cue == schema.DEFAULT_CUE
                           ^^^^^^^^^^^^^^^^^^
E       AttributeError: module 'ggufone.schema' has no attribute 'DEFAULT_CUE'

tests/test_policy_v2.py:178: AttributeError
_______ test_a_no_flags_request_renders_the_v2_cell_byte_for_byte_plain ________

    def test_a_no_flags_request_renders_the_v2_cell_byte_for_byte_plain() -> None:
        """The card's parity pin, plain assembly: no options == the explicitly flagged v2 cell."""
        default = rendered(parsed())
        flagged = rendered(parsed(**V2))
>       assert default["prefix"] == flagged["prefix"]
E       AssertionError: assert 'You are a de...fter login.\n' == 'You are a de...fter login.\n'
E         
E         Skipping 120 identical leading characters in diff, use -v to show
E         - eply with a single JSON object and nothing else — no explanation, no code fence, no extra keys.
E         + eply with the label alone — no explanation, no punctuation, no extra words.
E           STATE:
E           The billing dashboard is blank for every user after login.

tests/test_policy_v2.py:191: AssertionError
___ test_a_no_flags_request_renders_the_v2_cell_byte_for_byte_chat_template ____

    def test_a_no_flags_request_renders_the_v2_cell_byte_for_byte_chat_template() -> None:
        """The same pin through a family chat template: prefix + tail + opener, and the token ids."""
        default = rendered(parsed(), chat=True)
        flagged = rendered(parsed(**V2), chat=True)
>       assert default["prefix"] == flagged["prefix"]
E       AssertionError: assert '<|im_start|>...|>assistant\n' == '<|im_start|>....<|im_end|>\n'
E         
E         Skipping 139 identical leading characters in diff, use -v to show
E         - eply with a single JSON object and nothing else — no explanation, no code fence, no extra keys.<|im_end|>
E         + eply with the label alone — no explanation, no punctuation, no extra words.<|im_end|>
E           <|im_start|>user
E           STATE:
E           The billing dashboard is blank for every user after login.<|im_end|>
E         + <|im_start|>assistant

tests/test_policy_v2.py:210: AssertionError
__ test_the_no_flags_engine_publishes_the_v2_policy_and_the_value_row_verdict __

    def test_the_no_flags_engine_publishes_the_v2_policy_and_the_value_row_verdict() -> None:
        """The engine surface: the response names the v2 policy and reads the instructed value row."""
        request = parsed()
        session = JsonSession(n_vocab=512)
        plan = decide.plan_context(request, session, template=resolution(), resolve=False)
        view = prompt.build_question(request.questions[0], readout=request.options.readout,
                                    cue=request.options.cue, chat_format=request.options.chat_format,
                                    role=plan.role, index=0, contract=request.options.json_contract)
        value_position = len(plan.prefix_tokens) + len(session.tokenize(view.suffix)) - 1
        label = session.tokenize("billing")[0]
        session.row_fn = (lambda ctx, position=value_position, label=label:
                          biased_row(session.n_vocab, {label: TOP})
                          if ctx.position == position
                          else biased_row(session.n_vocab, {label: TOP}))
        result = decide.DecisionEngine(session).decide(request, plan=plan)
>       assert result.engine["cue"] == schema.DEFAULT_CUE
                                       ^^^^^^^^^^^^^^^^^^
E       AttributeError: module 'ggufone.schema' has no attribute 'DEFAULT_CUE'

tests/test_policy_v2.py:238: AttributeError
____ test_the_reproduce_line_names_the_pre_v2_cell_and_stays_silent_for_v2 _____

    def test_the_reproduce_line_names_the_pre_v2_cell_and_stays_silent_for_v2() -> None:
        """A row that is *not* the default must name its policy; the default row must not (E3e rule)."""
        default = harness.reproduce_command(harness.BenchConfig(suite="quality", model_path="/m.gguf",
                                                                backend="vulkan", runs=1, threads=4,
                                                                items=60))
        assert "--cue" not in default and "--chat-format" not in default
        pre_v2 = harness.reproduce_command(
            harness.BenchConfig(suite="quality", model_path="/m.gguf", backend="vulkan", runs=1,
                                threads=4, items=60, cue="shipped", chat_format="answer_sheet"))
>       assert "--cue shipped" in pre_v2
E       AssertionError: assert '--cue shipped' in 'uv run ggufone bench --suite quality --model /m.gguf --backend vulkan --runs 1 --threads 4 --items 60 --json'

tests/test_policy_v2.py:289: AssertionError
_________ test_the_report_marks_the_pre_v2_policy_and_not_the_default __________

    def test_the_report_marks_the_pre_v2_policy_and_not_the_default() -> None:
        report = {"suite": "quality", "generated_at": "2026-09-20T00:00:00Z",
                  "config": {"backend": "vulkan", "runs": 1, "threads": 4,
                             "cue": "shipped", "chat_format": "answer_sheet"},
                  "model": {"path": "/m.gguf"}, "overall": {}, "per_type": {},
                  "commands": {"reproduce": "uv run ggufone bench --suite quality"}}
        text = harness.render_report(report)
>       assert "- prompt policy: cue=shipped · chat_format=answer_sheet" in text
E       AssertionError: assert '- prompt policy: cue=shipped · chat_format=answer_sheet' in '### quality — m.gguf\n\n- generated: 2026-09-20T00:00:00Z\n- host: None · cpus None · cgroup quota n/a\n- config: bac...ement**\n\n| type | n | correct | agreement | 95% CI |\n|---|---|---|---|---|\n| overall | None | None | — | — – — |\n'

tests/test_policy_v2.py:304: AssertionError
=========================== short test summary info ============================
FAILED tests/test_policy_v2.py::test_the_default_cue_is_the_instructed_one - ...
FAILED tests/test_policy_v2.py::test_the_default_chat_format_is_the_role_split
FAILED tests/test_policy_v2.py::test_the_bench_and_the_cli_ask_for_the_v2_cell_by_default
FAILED tests/test_policy_v2.py::test_the_reproduce_tool_defaults_to_the_v2_cell
FAILED tests/test_policy_v2.py::test_a_no_flags_request_renders_the_v2_cell_byte_for_byte_plain
FAILED tests/test_policy_v2.py::test_a_no_flags_request_renders_the_v2_cell_byte_for_byte_chat_template
FAILED tests/test_policy_v2.py::test_the_no_flags_engine_publishes_the_v2_policy_and_the_value_row_verdict
FAILED tests/test_policy_v2.py::test_the_reproduce_line_names_the_pre_v2_cell_and_stays_silent_for_v2
FAILED tests/test_policy_v2.py::test_the_report_marks_the_pre_v2_policy_and_not_the_default
9 failed, 4 passed, 1 skipped in 0.17s
