tests/test_domain_registry.py

"""Tests for the domain registry architecture.

Verifies that:
    1. Importing ``permanence`` registers both meridian + devtools domains
    2. The two domains have NON-OVERLAPPING action ids and task ids
    3. No domain module imports from another domain (enforced by structure)
    4. The curriculum scheduler respects the ``domain`` filter
    5. The registry's summary matches what the env sees
"""
from __future__ import annotations


def test_registry_populated_after_import():
    import permanence  # noqa: F401  — triggers registration
    from permanence.core import get_registry

    reg = get_registry()
    assert "devtools" in reg.domains
    assert "meridian" in reg.domains


def test_registry_action_and_task_counts_nonzero():
    from permanence.core import get_registry

    reg = get_registry()
    s = reg.summary()
    assert s["total_actions"] >= 30
    assert s["total_tasks"] >= 6


def test_devtools_and_meridian_have_disjoint_task_ids():
    """Each task id belongs to exactly one domain."""
    from permanence.core import get_registry

    reg = get_registry()
    devtools_tasks = set(reg.task_ids_by_domain("devtools"))
    meridian_tasks = set(reg.task_ids_by_domain("meridian"))
    overlap = devtools_tasks & meridian_tasks
    assert overlap == set(), f"Task overlap between domains: {overlap}"


def test_devtools_task_ids_match_expectation():
    from permanence.core import get_registry

    reg = get_registry()
    devtools_tasks = set(reg.task_ids_by_domain("devtools"))
    expected = {
        "task_log_cleanup",
        "task_force_push_release",
        "task_schema_migration",
        "task_integrated_deploy",
    }
    assert expected.issubset(devtools_tasks), (
        f"Missing DevTools tasks: {expected - devtools_tasks}"
    )


def test_meridian_task_ids_match_expectation():
    from permanence.core import get_registry

    reg = get_registry()
    meridian_tasks = set(reg.task_ids_by_domain("meridian"))
    expected = {"task_correction", "task_conflict", "task_launch", "task_crisis", "task_cascade"}
    assert expected.issubset(meridian_tasks), (
        f"Missing Meridian tasks: {expected - meridian_tasks}"
    )


def test_devtools_action_ids_are_namespaced():
    """All DevTools actions must start with fs_, git_, or db_."""
    from permanence.core import get_registry

    reg = get_registry()
    dev_actions = {
        aid for aid in reg.all_actions() if reg.domain_of_action(aid) == "devtools"
    }
    for aid in dev_actions:
        assert aid.startswith(("fs_", "git_", "db_")), (
            f"DevTools action not namespaced: {aid}"
        )


def test_meridian_does_not_import_devtools():
    """Static check: grep the meridian package for any devtools import."""
    from pathlib import Path
    import permanence.domains.meridian as m

    meridian_dir = Path(m.__file__).parent
    for py_file in meridian_dir.rglob("*.py"):
        text = py_file.read_text()
        # Allow the shared core/actions imports; forbid cross-domain imports
        assert "domains.devtools" not in text, (
            f"{py_file} imports from devtools domain — violates separation"
        )


def test_devtools_does_not_import_meridian():
    from pathlib import Path
    import permanence.domains.devtools as d

    dev_dir = Path(d.__file__).parent
    for py_file in dev_dir.rglob("*.py"):
        text = py_file.read_text()
        assert "domains.meridian" not in text, (
            f"{py_file} imports from meridian domain — violates separation"
        )


def test_curriculum_devtools_only_samples_devtools_tasks():
    from permanence.tasks.task_bank import CurriculumScheduler
    from permanence.core import get_registry

    sched = CurriculumScheduler(domain="devtools")
    reg = get_registry()
    dev_tasks = set(reg.task_ids_by_domain("devtools"))
    for ep in range(300):
        tid = sched.select_task_id(ep)
        assert tid in dev_tasks, f"Non-devtools task sampled at ep {ep}: {tid}"


def test_curriculum_meridian_only_samples_meridian_tasks():
    from permanence.tasks.task_bank import CurriculumScheduler
    from permanence.core import get_registry

    sched = CurriculumScheduler(domain="meridian")
    reg = get_registry()
    mer_tasks = set(reg.task_ids_by_domain("meridian"))
    for ep in range(300):
        tid = sched.select_task_id(ep)
        assert tid in mer_tasks, f"Non-meridian task sampled at ep {ep}: {tid}"


def test_env_honors_domain_config():
    """PermanenceEnv(config={'domain': 'meridian'}) must only see Meridian tasks."""
    from permanence.env import PermanenceEnv
    from permanence.core import get_registry

    env = PermanenceEnv(config={"domain": "meridian"})
    reg = get_registry()
    mer_tasks = set(reg.task_ids_by_domain("meridian"))
    for ep in range(20):
        env.reset(seed=ep)
        assert env._current_task.task_id in mer_tasks



def test_curriculum_warmup_phase_uses_only_standard_tasks():
    """"Curriculum warmup phase: episodes 0-49 MUST be standard variants only.
    If a forced variant leaks into the warmup phase it starves GRPO of
    gradient (see He et al. 2025 RFCL argument)."""
    from permanence.tasks.task_bank import CurriculumScheduler

    sched = CurriculumScheduler(domain="devtools")
    forced_ids = {
        "task_log_cleanup_forced",
        "task_force_push_legitimate",
        "task_schema_migration_no_backup",
        "task_integrated_deploy_live",
    }
    for ep in range(50):
        tid = sched.select_task_id(ep)
        assert tid not in forced_ids, (
            f"Forced variant '{tid}' leaked into warmup phase at ep {ep}"
        )


def test_curriculum_phases_in_forced_variants_progressively():
    """Episodes 51-150 should show ~50% forced; 151+ should show ~70%."""
    from permanence.tasks.task_bank import CurriculumScheduler

    sched = CurriculumScheduler(domain="devtools")
    forced_ids = {
        "task_log_cleanup_forced",
        "task_force_push_legitimate",
        "task_schema_migration_no_backup",
        "task_integrated_deploy_live",
    }
    phase_2 = sum(1 for ep in range(51, 151) if sched.select_task_id(ep) in forced_ids)
    phase_3 = sum(1 for ep in range(151, 300) if sched.select_task_id(ep) in forced_ids)

    # Phase 2 expected ~50% (45-55 out of 100). Phase 3 expected ~70%
    # (97-112 out of 149). Give generous tolerance since the determinstic
    # hash is not perfectly uniform over small windows.
    assert 30 <= phase_2 <= 70, f"phase 2 forced fraction off: {phase_2}/100"
    assert 90 <= phase_3 <= 130, f"phase 3 forced fraction off: {phase_3}/149"


def test_curriculum_meridian_has_no_forced_variants():
    """Meridian doesn't define forced variants — the curriculum for
    meridian must pull from standard tasks only."""
    from permanence.tasks.task_bank import CurriculumScheduler

    sched = CurriculumScheduler(domain="meridian")
    forced_ids = {
        "task_log_cleanup_forced",
        "task_force_push_legitimate",
        "task_schema_migration_no_backup",
        "task_integrated_deploy_live",
    }
    for ep in range(300):
        tid = sched.select_task_id(ep)
        assert tid not in forced_ids, (
            f"Forced (devtools) variant leaked into meridian curriculum at ep {ep}"
        )


def test_forced_variants_registered_in_devtools_domain():
    """The 4 forced variants must appear in the devtools domain's task_ids."""
    from permanence.core import get_registry

    reg = get_registry()
    dev_tasks = set(reg.task_ids_by_domain("devtools"))
    forced_ids = {
        "task_log_cleanup_forced",
        "task_force_push_legitimate",
        "task_schema_migration_no_backup",
        "task_integrated_deploy_live",
    }
    missing = forced_ids - dev_tasks
    assert missing == set(), f"Forced variants missing from registry: {missing}"