"""C2 owned_object_usage[] echo schema v1 — integrated test (G1-G7). Spec ref: docs/superpowers/specs/2026-05-20-c2-owned-object-usage-echo-v1-design.md Plan ref: docs/superpowers/plans/2026-05-20-c2-owned-object-usage-echo-v1-implementation.md Codex APPROVED_FOR_W3_ENTRY_WITH_NARROW_PLAN_GAP_ABSORB. no live LLM + NO VLM + no regex semantic. 모든 gate 는 순수 helper 함수 + prompt-pack 파일 read 만 사용한다. """ import ast import json import pathlib import pytest from app.core.errors import AppError from app.core.steps._owned_helpers import ( OWNED_SENTINEL_SCHEMA_VERSION, _REQUIRED_SENTINEL_FIELDS, assert_close_framing_absent_echo, assert_owned_sentinel_shape, assert_owned_usage_hash_matches, build_owned_sentinel, compute_owned_usage_hash, has_redraw_violation, validate_owned_object_usage_coverage, ) REPO_ROOT = pathlib.Path(__file__).resolve().parents[3] def _resolve_latest_major_pack(stem: str, major: str) -> pathlib.Path: """prompt-pack stem 의 major 버전 dir 를 glob 으로 resolve (UTC suffix 무관).""" candidates = sorted((REPO_ROOT / "prompts/_base" / stem).glob(f"{major}.*")) assert candidates, f"no prompt-pack found for {stem}/{major}.*" return candidates[-1] SCENE_DETAIL_V29_DIR = _resolve_latest_major_pack("scene_detail", "29") SCENE_DETAIL_V30_DIR = _resolve_latest_major_pack("scene_detail", "30") OWNED_JUDGE_V3_DIR = _resolve_latest_major_pack("scene_detail_owned_judge", "3") OWNED_JUDGE_V4_DIR = _resolve_latest_major_pack("scene_detail_owned_judge", "4") # === HISTORICAL AUDIT TRAIL — v3 owned_judge closed-list residue catalog BEGIN === # 아래 문자열은 v3 system.md 의 closed-list 잔재 (redraw 동사 화이트리스트 / # redraw 예시 / 판정 우선순위 / ambiguous default) 로, judge v4 echo-verify-only # prompt 에서 0 임을 증명하는 G4 gate 의 검사 대상이다. live policy 아님 — # audit / residue scan 이 본 catalog 를 잔재로 오인하지 않도록 marker 로 감싼다. _V3_CLOSED_LIST_RESIDUE = ( "redraw 동사 화이트리스트", "판정 우선순위", "ambiguous mention 기본값", "add a new door on the right wall", "render a glowing portal in the wall", "set up a desk by the window", ) # === HISTORICAL AUDIT TRAIL — v3 owned_judge closed-list residue catalog END === _VLM_FORBIDDEN = {"vlm", "vision_lm", "vision_language", "lvm_judge", "lvm_validator"} def _assert_no_forbidden_model_refs(tree: ast.AST, where: str) -> None: """AST-based import/call leak check (substring grep 보다 엄격).""" for node in ast.walk(tree): if isinstance(node, (ast.Import, ast.ImportFrom)): for alias in node.names: name_lower = (alias.name or "").lower() for kw in _VLM_FORBIDDEN: assert kw not in name_lower, f"VLM import leak in {where}: {alias.name}" if isinstance(node, ast.ImportFrom): mod_lower = (node.module or "").lower() for kw in _VLM_FORBIDDEN: assert kw not in mod_lower, f"VLM module leak in {where}: {node.module}" if isinstance(node, ast.Call) and isinstance(node.func, ast.Name): for kw in _VLM_FORBIDDEN: assert kw not in node.func.id.lower(), f"VLM call leak in {where}: {node.func.id}" def _entry(token: str, kind: str = "anchor", phrase: str = "ref phrase") -> dict: return { "owned_token": token, "usage_kind": kind, "source_phrase": "" if kind == "absent" else phrase, } # ── G1 — Producer schema owned_object_usage[] required + field shape ── def test_g1_producer_schema_owned_object_usage_required(): schema = json.loads((SCENE_DETAIL_V30_DIR / "detail_schema.json").read_text()) item = schema["properties"]["t2i_variations"]["items"] assert "owned_object_usage" in item["required"] oou = item["properties"]["owned_object_usage"] assert oou["type"] == "array" oitem = oou["items"] props = oitem["properties"] assert set(props) == {"owned_token", "usage_kind", "source_phrase"} assert set(oitem["required"]) == {"owned_token", "usage_kind", "source_phrase"} assert oitem["additionalProperties"] is False assert props["owned_token"]["minLength"] == 1 def test_g1_coverage_duplicate_token_raises(): # cardinality 일치하나 owned_token 중복 → AppError. dup = [_entry("door"), _entry("door")] with pytest.raises(AppError) as exc: validate_owned_object_usage_coverage(dup, ["door", "window"]) assert exc.value.code == "step.contract_violation" assert "duplicate" in exc.value.message.lower() def test_g1_coverage_missing_token_raises(): # owned 에 door 있는데 echo 가 비어 cardinality 부족 → AppError. with pytest.raises(AppError) as exc: validate_owned_object_usage_coverage([], ["door"]) assert exc.value.code == "step.contract_violation" def test_g1_coverage_extra_token_raises(): # echo 에 owned 밖 token 추가 → cardinality 초과 → AppError. extra = [_entry("door"), _entry("window")] with pytest.raises(AppError) as exc: validate_owned_object_usage_coverage(extra, ["door"]) assert exc.value.code == "step.contract_violation" def test_g1_coverage_empty_owned_ok(): # owned 가 비면 echo 도 [] — required field 존재, no raise. validate_owned_object_usage_coverage([], []) # ── G2 — usage_kind enum strict + source_phrase non-empty unless absent ── def test_g2_usage_kind_enum_strict(): bad = [{"owned_token": "door", "usage_kind": "future_kind", "source_phrase": "x"}] with pytest.raises(AppError) as exc: validate_owned_object_usage_coverage(bad, ["door"]) assert exc.value.code == "step.contract_violation" def test_g2_source_phrase_required_unless_absent(): # anchor 인데 source_phrase 비면 위반. with pytest.raises(AppError): validate_owned_object_usage_coverage( [{"owned_token": "door", "usage_kind": "anchor", "source_phrase": ""}], ["door"], ) # absent 면 빈 source_phrase 허용. validate_owned_object_usage_coverage( [{"owned_token": "door", "usage_kind": "absent", "source_phrase": ""}], ["door"], ) def test_g2_schema_usage_kind_enum(): schema = json.loads((SCENE_DETAIL_V30_DIR / "detail_schema.json").read_text()) oitem = ( schema["properties"]["t2i_variations"]["items"]["properties"] ["owned_object_usage"]["items"] ) assert set(oitem["properties"]["usage_kind"]["enum"]) == {"redraw", "anchor", "absent"} # ── G3 — sentinel v2 + owned_usage_hash tamper mismatch → AppError ── def test_g3_sentinel_v2_schema_version_and_fields(): assert OWNED_SENTINEL_SCHEMA_VERSION == 2 assert "owned_usage_hash" in _REQUIRED_SENTINEL_FIELDS assert len(_REQUIRED_SENTINEL_FIELDS) == 7 def test_g3_build_sentinel_has_owned_usage_hash(): sentinel = build_owned_sentinel( owned=[], camera_direction="", t2i_prompt="x", is_close_framing=False, violations=[], owned_object_usage=[], ) assert "owned_usage_hash" in sentinel assert_owned_sentinel_shape(sentinel, where="g3") def test_g3_owned_usage_hash_tamper_mismatch_raises(): usage = [_entry("door", "anchor", "near the door")] sentinel = build_owned_sentinel( owned=["door"], camera_direction="medium", t2i_prompt="a room near the door", is_close_framing=False, violations=[], owned_object_usage=usage, ) # unmodified echo → match (no raise). assert_owned_usage_hash_matches(sentinel, usage) # tampered echo → owned_usage_hash mismatch → AppError. tampered = [_entry("door", "redraw", "draw a new door")] with pytest.raises(AppError) as exc: assert_owned_usage_hash_matches(sentinel, tampered) assert exc.value.code == "step.contract_violation" def test_g3_owned_usage_hash_deterministic_order_independent(): a = [_entry("door"), _entry("window")] b = [_entry("window"), _entry("door")] assert compute_owned_usage_hash(a) == compute_owned_usage_hash(b) # ── G4 — judge v4 no closed-list residue + echo-verify prompt-pack ── def test_g4_judge_v4_no_closed_list_residue(): sys_md = (OWNED_JUDGE_V4_DIR / "system.md").read_text() for residue in _V3_CLOSED_LIST_RESIDUE: assert residue not in sys_md, f"judge v4 closed-list residue leak: {residue!r}" def test_g4_judge_v4_is_echo_verify(): sys_md = (OWNED_JUDGE_V4_DIR / "system.md").read_text() assert sys_md.strip(), "judge v4 system.md empty" # echo-verify-only prompt 은 owned_object_usage 를 cross-check 대상으로 명시. assert "owned_object_usage" in sys_md def test_g4_v4_user_template_placeholders(): tmpl = (OWNED_JUDGE_V4_DIR / "user_template.md").read_text() assert "{owned_object_usage}" in tmpl for ph in ("{t2i_prompt}", "{owned_list_block}", "{camera_direction}"): assert ph in tmpl, f"judge v4 user_template missing placeholder {ph}" # ── G5 — verdict enum + has_redraw_violation preserve ── def test_g5_v4_schema_byte_identical_v3(): v3 = (OWNED_JUDGE_V3_DIR / "schema.json").read_bytes() v4 = (OWNED_JUDGE_V4_DIR / "schema.json").read_bytes() assert v3 == v4, "v4 schema.json must be byte-identical to v3 (verdict enum preserve)" def test_g5_v4_schema_verdict_enum(): sch = json.loads((OWNED_JUDGE_V4_DIR / "schema.json").read_text()) verdict = sch["properties"]["violations"]["items"]["properties"]["verdict"] assert set(verdict["enum"]) == {"redraw_violation", "anchor_reference"} def test_g5_has_redraw_violation_preserve(): assert has_redraw_violation([]) is False redraw = [{"owned_object": "door", "violating_phrase": "p", "reason": "r", "verdict": "redraw_violation"}] anchor = [{"owned_object": "door", "violating_phrase": "p", "reason": "r", "verdict": "anchor_reference"}] assert has_redraw_violation(redraw) is True assert has_redraw_violation(anchor) is False # ── G6 — boundary preserve (t2i_prompt body + C1 + Area C) ── def test_g6_t2i_prompt_body_shape_unchanged_v29_to_v30(): v29 = json.loads((SCENE_DETAIL_V29_DIR / "detail_schema.json").read_text()) v30 = json.loads((SCENE_DETAIL_V30_DIR / "detail_schema.json").read_text()) i29 = v29["properties"]["t2i_variations"]["items"]["properties"] i30 = v30["properties"]["t2i_variations"]["items"]["properties"] assert i29["t2i_prompt"] == i30["t2i_prompt"], "t2i_prompt body shape drift" # v30 = v29 + owned_object_usage 만 추가 (sidecar evidence, 다른 field 변경 0). assert set(i30) - set(i29) == {"owned_object_usage"} assert set(i29) - set(i30) == set() def test_g6_c1_perception_mode_boundary_preserve(): from app.core.perception_mode import PERCEPTION_MODES assert "mirror" in PERCEPTION_MODES assert len(PERCEPTION_MODES) == 8 # v30 = v29 copy — C1 perception_mode cross-reference 보존. sys_md = (SCENE_DETAIL_V30_DIR / "system.md").read_text() assert "REPRODUCTION_PERCEPTION_MODES" in sys_md or "perception_mode.py" in sys_md def test_g6_area_c_boundary_preserve(): # C2 는 render_prompt_card 를 건드리지 않음 — Area C directionality anchor 보존. src = (REPO_ROOT / "backend/app/core/steps/render_prompt_card.py").read_text() assert "_ALLOWED_DIRECTIONALITY_CLASSES" in src # ── G7 — NO VLM (AST) + no silent fallback + close-framing absent echo ── def test_g7_no_vlm_owned_modules(): for mod_path in ( "backend/app/core/steps/_owned_judge.py", "backend/app/core/steps/_owned_helpers.py", ): tree = ast.parse((REPO_ROOT / mod_path).read_text()) _assert_no_forbidden_model_refs(tree, mod_path) def test_g7_no_vlm_test_module(): _assert_no_forbidden_model_refs( ast.parse(pathlib.Path(__file__).read_text()), __file__ ) def test_g7_close_framing_absent_echo_ok(): ok = [_entry("door", "absent"), _entry("window", "absent")] assert_close_framing_absent_echo(ok, where="g7") # no raise def test_g7_close_framing_non_absent_raises(): bad = [_entry("door", "anchor", "near the door"), _entry("window", "absent")] with pytest.raises(AppError) as exc: assert_close_framing_absent_echo(bad, where="g7") assert exc.value.code == "step.contract_violation" def test_g7_coverage_mismatch_raises_no_silent_fallback(): # owned_object_usage missing / non-list / coverage mismatch → AppError fail-fast. with pytest.raises(AppError): validate_owned_object_usage_coverage("not a list", ["door"]) with pytest.raises(AppError): validate_owned_object_usage_coverage([_entry("window")], ["door"])