"""outdoor_frame_mode — 레인2 샷별 구도 판정 결정론 테스트 (E2E6 ③). 판정 완성도=E2E/육안 — 여기는 검증(인용·parity·low 반려)·재시도· fail-closed·팩 중립성·등록만 잠근다. """ from typing import Any, Dict, List import pytest from app.core.errors import AppError from app.modules.pipeline.outdoor_frame_mode import ( FRAME_MODES, build_frame_mode_schema, resolve_prompt_version, run_outdoor_frame_mode, validate_frame_modes, ) SCENE_TEXTS = {4: "옥탑방 전경. 철문이 반쯤 열려 있다.", 17: "고양이가 담을 넘는다."} HEADINGS = {4: "S#4. 옥상 (낮)", 17: "S#17. 옥상 (밤)"} SHOTS = [ {"scene_index": 4, "shot_index": 1, "description": "부감 전경", "camera_direction": "높은 부감"}, {"scene_index": 17, "shot_index": 1, "description": "담 위 고양이", "camera_direction": ""}, ] TAGS = ["S4sh1", "S17sh1"] def _item(tag, mode="structure_dominant", conf="high", si=4, quote="철문이 반쯤 열려 있다."): return {"shot": tag, "frame_mode": mode, "confidence": conf, "evidence": {"scene_index": si, "quote_ko": quote}, "rationale_ko": "근거"} def test_validate_parity_and_evidence(): ok = [_item("S4sh1"), _item("S17sh1", mode="layout_dominant", si=17, quote="담을 넘는다")] assert validate_frame_modes(ok, TAGS, SCENE_TEXTS) == [] # 누락 v = validate_frame_modes([_item("S4sh1")], TAGS, SCENE_TEXTS) assert any("판정 누락" in x for x in v) # 중복 + 대상 밖 v = validate_frame_modes( [_item("S4sh1"), _item("S4sh1"), _item("S99sh1")], TAGS, SCENE_TEXTS) assert any("중복" in x for x in v) assert any("대상 밖" in x for x in v) # 인용 원문 부재 v = validate_frame_modes( [_item("S4sh1", quote="원문에 없는 문장")], ["S4sh1"], SCENE_TEXTS) assert any("원문에 없음" in x for x in v) # low confidence = 위반 (mixed 임의 라우팅 금지) v = validate_frame_modes([_item("S4sh1", conf="low")], ["S4sh1"], SCENE_TEXTS) assert any("confidence=low" in x for x in v) def test_run_retry_then_success_and_fail_closed(): calls: List[str] = [] def flaky(step_tag, system, user, schema, **kw): calls.append(user) if len(calls) == 1: return {"items": [_item("S4sh1")]} # S17sh1 누락 → 재시도 return {"items": [ _item("S4sh1"), _item("S17sh1", mode="layout_dominant", si=17, quote="담을 넘는다"), ]} out = run_outdoor_frame_mode( shots=SHOTS, scene_texts=SCENE_TEXTS, scene_headings=HEADINGS, call_structured_fn=flaky, ) assert set(out["shots"]) == set(TAGS) assert out["shots"]["S4sh1"]["frame_mode"] == "structure_dominant" assert len(calls) == 2 # 재시도 프롬프트에 위반 피드백 포함 + 씬 원문 전문 유지(자르지 않음) assert "계약 위반" in calls[1] for si, txt in SCENE_TEXTS.items(): assert txt in calls[1] def always_bad(step_tag, system, user, schema, **kw): return {"items": []} with pytest.raises(AppError): run_outdoor_frame_mode( shots=SHOTS, scene_texts=SCENE_TEXTS, scene_headings=HEADINGS, call_structured_fn=always_bad, ) def test_validate_cross_scene_evidence_rejected(): """Codex f7c81336 B2: 다른 씬의 실제 인용도 반려 — 해당 샷의 씬만.""" it = _item("S4sh1", si=17, quote="고양이가 담을 넘는다.") v = validate_frame_modes([it], ["S4sh1"], SCENE_TEXTS) assert any("해당 샷의 씬" in x for x in v) def test_validate_frame_cp_entries_status_and_contract(): """Codex f7c81336 H3: persisted CP 소비 재검증 — status·enum·parity.""" from app.modules.pipeline.outdoor_frame_mode import ( validate_frame_cp_entries, ) ok_entry = { "status": "ok", "frame_mode": "structure_dominant", "confidence": "high", "evidence": {"scene_index": 4, "quote_ko": "철문이 반쯤 열려 있다."}, "rationale_ko": "근거", } assert validate_frame_cp_entries( {"S4sh1": ok_entry}, ["S4sh1"], SCENE_TEXTS) == [] # status!=ok bad = dict(ok_entry, status="failed") v = validate_frame_cp_entries({"S4sh1": bad}, ["S4sh1"], SCENE_TEXTS) assert any("ok 아님" in x for x in v) # enum 밖 frame_mode = layout 하강 아님, 위반 weird = dict(ok_entry, frame_mode="SAMPLE_mode") v = validate_frame_cp_entries({"S4sh1": weird}, ["S4sh1"], SCENE_TEXTS) assert any("frame_mode 값 위반" in x for x in v) # 대상 샷 누락 = parity 위반 v = validate_frame_cp_entries({}, ["S4sh1"], SCENE_TEXTS) assert any("판정 누락" in x for x in v) def test_step_config_hash_sensitive_to_physical_model(monkeypatch): """Codex f7c81336 H4: alias 아닌 물리 모델 스탬프 — 교체=드리프트.""" from app.core.config import settings from app.core.steps.outdoor_frame_mode_step import OutdoorFrameModeStep step = object.__new__(OutdoorFrameModeStep) step.project_config = {} monkeypatch.setattr(settings, "openai_model", "SAMPLE-model-a") h1 = OutdoorFrameModeStep._config_hash(step) monkeypatch.setattr(settings, "openai_model", "SAMPLE-model-b") h2 = OutdoorFrameModeStep._config_hash(step) assert h1 != h2 def test_decide_direct_seed_location_ref_mode_contract(): """Codex f7c81336 B1: prev=classify SOT 고정 — asset 결손=fail-closed.""" from app.modules.pipeline.still_recipe import ( decide_direct_seed_location_ref_mode, ) assert decide_direct_seed_location_ref_mode( prev_tag=None, prev_sel_exists=False, bg_only=False) == "seed_only" assert decide_direct_seed_location_ref_mode( prev_tag="S1sh1", prev_sel_exists=True, bg_only=False) == "prev_only" # bgonly 는 classify 가 prev 무효화 — seed_only assert decide_direct_seed_location_ref_mode( prev_tag="S1sh1", prev_sel_exists=False, bg_only=True) == "seed_only" with pytest.raises(ValueError): decide_direct_seed_location_ref_mode( prev_tag="S1sh1", prev_sel_exists=False, bg_only=False) def test_schema_enum_locked(): schema = build_frame_mode_schema() enum = (schema["properties"]["items"]["items"]["properties"] ["frame_mode"]["enum"]) assert tuple(enum) == FRAME_MODES def test_pack_scenario_neutral(): from pathlib import Path repo = Path(__file__).resolve().parents[3] ver = resolve_prompt_version("1") text = (repo / "prompts" / "_base" / "outdoor_frame_mode" / ver / "judge_system.md").read_text(encoding="utf-8") for word in ("옥탑", "정류장", "rooftop", "bus stop", "항구"): assert word not in text assert "structure_dominant" in text and "layout_dominant" in text def test_step_registered_and_gated(): from app.core.step_manifest import get_manifest_dict from app.core.steps import STEP_CLASSES assert "outdoor_frame_mode" in STEP_CLASSES meta = get_manifest_dict("outdoor_frame_mode") assert meta["applicability"] == "if_outdoor_frame_mode" assert meta["allow_partial_downstream"] is False assert meta["order"] == 21.78 assert "outdoor_lane_plan" in meta["depends_on"] def test_unknown_version_raises(): with pytest.raises(ValueError): resolve_prompt_version("99")