transcription/tests/test_model_contract_guards.py

"""Deterministic guards for the model and persistence contracts.

These cover three checks that `.github/skills/python-code-reviewer/skill.md` requires
on every review but that previously had no automated enforcement:

* **Status vocabulary conformance** - status comparisons and assignments must use the
  `JobStatus` / `JobSourceStatus` / `JobPurpose` enums rather than string literals.
* **Relationship loading contract** - relationships declare ``lazy="raise"`` so read
  paths must eager-load explicitly, and the documented exceptions in `docs/schema.md`
  must match the code exactly.
* **Schema contract fidelity** - the "Field-Accurate Table Contracts" tables in
  `docs/schema.md` must list exactly the fields each SQLModel table declares.

See `docs/requirements.md` (REQ-4-102), `docs/schema.md` ("Relationship Loading
Contract"), and `.github/instructions/services.instructions.md`.
"""

from __future__ import annotations

import ast
import re
from pathlib import Path

from sqlmodel import SQLModel

from transcription.db import models as models_module
from transcription.db.models import JobPurpose
from transcription.db.models import JobSourceStatus
from transcription.db.models import JobStatus

PROJECT_ROOT = Path(__file__).resolve().parents[1]
SOURCE_DIR = PROJECT_ROOT / "src" / "transcription"
MODELS_PATH = SOURCE_DIR / "db" / "models.py"
SCHEMA_DOC = PROJECT_ROOT / "docs" / "schema.md"

STATUS_ENUMS = (JobStatus, JobSourceStatus, JobPurpose)
STATUS_VALUES = frozenset(member.value for enum in STATUS_ENUMS for member in enum)

# Attribute names that carry a status enum. A string literal compared against or
# assigned to one of these is a stringly-typed status, even if it happens to match.
STATUS_ATTRIBUTES = frozenset({"status", "purpose"})

# `JobSource.execution_attempts` loads attempt evidence on demand rather than raising,
# because evidence is fetched deliberately by the services that own it. Documented in
# `docs/schema.md` under "Relationship Loading Contract".
DOCUMENTED_LOADING_EXCEPTIONS = {"execution_attempts": "noload"}


def _python_files() -> list[Path]:
    return sorted(SOURCE_DIR.rglob("*.py"))


def _relative(path: Path) -> str:
    return path.relative_to(PROJECT_ROOT).as_posix()


def _is_status_target(node: ast.expr) -> bool:
    """True for `x.status`, `x.purpose`, and their `.value` unwrappings."""
    if isinstance(node, ast.Attribute):
        if node.attr in STATUS_ATTRIBUTES:
            return True
        if node.attr == "value":
            return _is_status_target(node.value)
    return False


def _string_constant(node: ast.expr) -> str | None:
    return node.value if isinstance(node, ast.Constant) and isinstance(node.value, str) else None


def _status_literal_violations(tree: ast.Module) -> list[tuple[int, str]]:
    found: list[tuple[int, str]] = []
    for node in ast.walk(tree):
        if isinstance(node, ast.Compare):
            operands = [node.left, *node.comparators]
            if not any(_is_status_target(operand) for operand in operands):
                continue
            for operand in operands:
                literal = _string_constant(operand)
                if literal is not None:
                    found.append((node.lineno, literal))
        elif isinstance(node, ast.Call):
            for keyword in node.keywords:
                if keyword.arg not in STATUS_ATTRIBUTES:
                    continue
                literal = _string_constant(keyword.value)
                if literal is not None:
                    found.append((node.lineno, literal))
    return found


def test_status_enums_expose_expected_vocabulary():
    """Guard the guard: the scan below is meaningless if the enums are empty."""
    assert {member.value for member in JobStatus} == {
        "queued",
        "processing",
        "transcribed",
        "partial_success",
        "failed",
    }
    assert {member.value for member in JobSourceStatus} == {
        "pending",
        "transcribed",
        "failed",
        "cancelled",
    }
    assert {member.value for member in JobPurpose} == {"transcription", "retranscription"}


def test_no_stringly_typed_status_comparisons_or_assignments():
    """Status handling must go through the enums, never raw strings.

    `attempt.status.value == "transcribed"` silently survives an enum rename and
    compares a projection of the value rather than the value itself.
    """
    violations: dict[str, list[tuple[int, str]]] = {}
    for path in _python_files():
        tree = ast.parse(path.read_text(encoding="utf-8"))
        found = _status_literal_violations(tree)
        if found:
            violations[_relative(path)] = found
    assert violations == {}


def test_status_string_literals_outside_models_are_accounted_for():
    """Any bare status-valued literal in the package must be a known non-status use.

    This is deliberately narrower than the comparison scan: it catches literals that
    merely *look* like statuses, so a genuine new one cannot slip in unnoticed.
    """
    allowed = {
        # `JobStatus` / `JobSourceStatus` / `JobPurpose` member definitions.
        "src/transcription/db/models.py",
        # `WorkerHealthState` is a separate Literal vocabulary that reuses "failed".
        "src/transcription/worker.py",
        # Distribution name lookups for `transcription`, not the JobPurpose member.
        "src/transcription/config.py",
        "src/transcription/providers/evidence.py",
        # UI placeholder copy where a value is absent, not a status render.
        "src/transcription/ui/pages/jobs_page.py",
    }
    unexpected: dict[str, list[tuple[int, str]]] = {}
    for path in _python_files():
        relative = _relative(path)
        if relative in allowed:
            continue
        tree = ast.parse(path.read_text(encoding="utf-8"))
        found = [
            (node.lineno, node.value)
            for node in ast.walk(tree)
            if isinstance(node, ast.Constant) and isinstance(node.value, str) and node.value in STATUS_VALUES
        ]
        if found:
            unexpected[relative] = found
    assert unexpected == {}


def _relationship_loading_strategies() -> dict[str, dict[str, str | None]]:
    """Map each model attribute defined via `Relationship(...)` to its lazy strategy."""
    tree = ast.parse(MODELS_PATH.read_text(encoding="utf-8"))
    strategies: dict[str, dict[str, str | None]] = {}
    for class_node in tree.body:
        if not isinstance(class_node, ast.ClassDef):
            continue
        for statement in class_node.body:
            if not isinstance(statement, ast.AnnAssign) or statement.value is None:
                continue
            call = statement.value
            if not isinstance(call, ast.Call) or getattr(call.func, "id", None) != "Relationship":
                continue
            attribute = statement.target.id if isinstance(statement.target, ast.Name) else "<unknown>"
            lazy: str | None = None
            for keyword in call.keywords:
                if keyword.arg != "sa_relationship_kwargs" or not isinstance(keyword.value, ast.Dict):
                    continue
                for key, value in zip(keyword.value.keys, keyword.value.values, strict=True):
                    if isinstance(key, ast.Constant) and key.value == "lazy" and isinstance(value, ast.Constant):
                        lazy = value.value
            strategies.setdefault(class_node.name, {})[attribute] = lazy
    return strategies


def test_relationships_are_discovered():
    """Guard the guard: the loading rules below are meaningless if nothing is scanned."""
    strategies = _relationship_loading_strategies()
    assert {"Document", "Job", "JobSource", "Source"} <= set(strategies)
    assert sum(len(attributes) for attributes in strategies.values()) >= 25


def test_relationships_declare_lazy_raise_except_documented_cases():
    """REQ-4-102: relationships raise on implicit load so read shape stays explicit."""
    violations: dict[str, str | None] = {}
    for model_name, attributes in _relationship_loading_strategies().items():
        for attribute, lazy in attributes.items():
            expected = DOCUMENTED_LOADING_EXCEPTIONS.get(attribute, "raise")
            if lazy != expected:
                violations[f"{model_name}.{attribute}"] = lazy
    assert violations == {}


def test_documented_loading_exceptions_match_schema_doc():
    """The exception list is only trustworthy while `docs/schema.md` agrees with it."""
    schema_text = SCHEMA_DOC.read_text(encoding="utf-8")
    contract = schema_text.split("## Relationship Loading Contract", 1)[1]
    for attribute, strategy in DOCUMENTED_LOADING_EXCEPTIONS.items():
        assert attribute in contract, f"{attribute} is exempted in code but not documented"
        assert f'`lazy="{strategy}"`' in contract


def _documented_table_fields() -> dict[str, list[str]]:
    schema_text = SCHEMA_DOC.read_text(encoding="utf-8")
    section = schema_text.split("## Field-Accurate Table Contracts", 1)[1]
    documented: dict[str, list[str]] = {}
    for block in re.split(r"\n### ", section)[1:]:
        heading = block.splitlines()[0].strip().strip("`")
        documented[heading] = re.findall(r"^\| `([^`]+)` \|", block, re.MULTILINE)
    return documented


def _table_models() -> dict[str, type[SQLModel]]:
    return {
        name: attribute
        for name, attribute in vars(models_module).items()
        if isinstance(attribute, type)
        and issubclass(attribute, SQLModel)
        and attribute is not SQLModel
        and getattr(attribute, "__table__", None) is not None
    }


def test_schema_doc_documents_every_table_model():
    """Every persisted table needs a field contract, and vice versa."""
    documented = set(_documented_table_fields())
    actual = set(_table_models())
    assert actual, "no table models discovered"
    assert documented - actual == set(), "schema.md documents tables that no longer exist"
    assert actual - documented == set(), "schema.md is missing tables that exist in models.py"


def test_schema_doc_field_contracts_match_models():
    """Check 6: `docs/schema.md` stays field-accurate with `db/models.py`."""
    documented = _documented_table_fields()
    drift: dict[str, dict[str, list[str]]] = {}
    for name, model in _table_models().items():
        expected = set(model.model_fields)
        listed = set(documented.get(name, []))
        if expected != listed:
            drift[name] = {
                "undocumented_fields": sorted(expected - listed),
                "stale_doc_entries": sorted(listed - expected),
            }
    assert drift == {}


def test_schema_doc_lists_fields_in_declaration_order():
    """Ordering drift is how a doc silently stops being reviewable against the model."""
    documented = _documented_table_fields()
    misordered = {
        name: {"documented": documented[name], "declared": list(model.model_fields)}
        for name, model in _table_models().items()
        if documented.get(name, []) != list(model.model_fields)
    }
    assert misordered == {}


def test_schema_doc_enumerations_match_status_enums():
    """The "Authoritative Enumerations" section must list the real members."""
    schema_text = SCHEMA_DOC.read_text(encoding="utf-8")
    section = schema_text.split("## Authoritative Enumerations", 1)[1].split("\n## ", 1)[0]
    drift: dict[str, dict[str, list[str]]] = {}
    for enum in STATUS_ENUMS:
        block = re.split(rf"\n### {enum.__name__}\n", section)
        assert len(block) == 2, f"{enum.__name__} has no section in docs/schema.md"
        listed = re.findall(r"^- `([^`]+)`", block[1].split("\n### ", 1)[0], re.MULTILINE)
        expected = [member.value for member in enum]
        if listed != expected:
            drift[enum.__name__] = {"documented": listed, "declared": expected}
    assert drift == {}