"""육안 8건 wave (2026-08-13, #106) — 심각도 게이트·ab 변주·팩 v18/v10 계약. 결정론 층만 잠근다(LLM 품질 아님): ①critical 만 fix 를 발동시키는 코드 게이트(S39sh4: major 소파 지적이 fix 를 발동시켜 인물이 바뀐 실측이 근거) ②2롤 ab 변주 맵이 백엔드 불문 둘째 롤부터 변주 절을 동반 ③신규 팩 스템의 실재·로드·주입. """ from __future__ import annotations from pathlib import Path from typing import Any, Dict, List from app.modules.pipeline.multiroll_gemini import ( QK_CRITIQUE_PACK_VERSION, STILL_FIX_REJUDGE_HEADER_PACK_VERSION, load_fix_rejudge_header, resolve_judge_pack_version, ) from app.modules.pipeline.multiroll_select import ( build_critique_schema, run_multiroll_select, ) from app.modules.pipeline.still_recipe import ( IDENTITY_ROLE_PROMPT_VERSION, SUPPORT_STILLNESS_PROMPT_VERSION, build_ab_roll_prompt_map, build_identity_ref_role_clause, build_still_prompt, resolve_prompt_version, ) FIX_HEAD = "Edit this photograph. Apply ONLY the corrections listed below:" FIX_TAIL = "PRESERVE EVERYTHING ELSE EXACTLY." class FakeGen: def __init__(self): self.calls: List[Any] = [] def __call__(self, tag, prompt, labeled_refs, out_path: Path) -> Path: self.calls.append((tag, prompt, list(labeled_refs), out_path)) out_path.parent.mkdir(parents=True, exist_ok=True) key = out_path.stem.rsplit("_", 1)[-1] out_path.write_bytes(f"img:{key}".encode()) return out_path def _judge(tag, prompt, labeled_refs, cand_paths, labels): # 라벨 계약=대문자(roll_labels) — B 를 단독 최고점으로 선정. return { "winner": "B", "ranking": list(reversed(list(labels))), "verdicts": [ {"label": lab, "score": (9 if lab == "B" else 5), "verdict_ko": ""} for lab in labels ], } def _run(tmp_path: Path, issues: List[Dict[str, Any]]): fix_gen = FakeGen() def critique_fn(tag, prompt, labeled_refs, image_path): return {"issues": issues} (tmp_path / "ref.png").write_bytes(b"ref") sel, record = run_multiroll_select( tag="t1", prompt="PROMPT", labeled_refs=[("REF", tmp_path / "ref.png")], out_stem=tmp_path / "out" / "s1", gen_fn=FakeGen(), judge_fn=_judge, critique_fn=critique_fn, fix_gen_fn=fix_gen, roll_count=2, critique_enabled=True, fix_head=FIX_HEAD, fix_tail=FIX_TAIL, fix_label="PHOTOGRAPH TO EDIT", record=None, ) return sel, record, fix_gen # ── 심각도 게이트 ────────────────────────────────────────────────────── def test_severity_gate_skips_fix_when_nothing_critical(tmp_path): """S39sh4 재발 방지: major/minor 뿐이면 fix 자체가 없다 — 원본 확정.""" sel, record, fix_gen = _run(tmp_path, [ {"issue_ko": "가구 소실", "fix_en": "Restore the sofa.", "severity": "major"}, {"issue_ko": "바닥 무늬", "fix_en": "Match the floor grain.", "severity": "minor"}, ]) assert fix_gen.calls == [] assert sel.read_bytes() == b"img:b" assert record["fix_skipped"] is True assert record["fix_skip_reason"] == "no_critical_issue" assert record["fix_severity_skipped_count"] == 2 def test_severity_gate_fixes_only_critical(tmp_path): sel, record, fix_gen = _run(tmp_path, [ {"issue_ko": "다른 인물", "fix_en": "Replace with the referenced man.", "severity": "critical"}, {"issue_ko": "가구 소실", "fix_en": "Restore the sofa.", "severity": "major"}, ]) assert len(fix_gen.calls) == 1 fix_prompt = fix_gen.calls[0][1] assert "Replace with the referenced man." in fix_prompt assert "Restore the sofa." not in fix_prompt assert record["fix_severity_skipped_count"] == 1 def test_severity_gate_absent_keeps_legacy_behavior(tmp_path): """severity 필드가 전무한 구 팩 경로(v7 critique·GQ v8)는 기존 그대로.""" sel, record, fix_gen = _run(tmp_path, [ {"issue_ko": "가구 소실", "fix_en": "Restore the sofa."}, ]) assert len(fix_gen.calls) == 1 assert "Restore the sofa." in fix_gen.calls[0][1] assert "fix_severity_skipped_count" not in record # ── critique 스키마 ──────────────────────────────────────────────────── def test_critique_schema_default_has_no_severity(): item = build_critique_schema()["properties"]["issues"]["items"] assert "severity" not in item["properties"] assert item["required"] == ["issue_ko", "fix_en"] def test_critique_schema_with_severity_requires_it(): item = build_critique_schema( with_severity=True)["properties"]["issues"]["items"] assert item["properties"]["severity"]["enum"] == [ "critical", "major", "minor"] assert "severity" in item["required"] # Codex R1 BLOCK-1: 관찰 인덱스 왕복도 스키마 강제 assert item["properties"]["observation_index"] == {"type": "integer"} assert "observation_index" in item["required"] def test_roll_count_validator_locks_two(): """Codex R1 BLOCK-2: 낡은 env(3)가 과금 1.5배+b·c 동일 변주로 새는 창 — 설정 층에서 2 외 전부 fail-fast.""" import pytest from app.core.config import Settings assert Settings._still_recipe_roll_count_range(2) == 2 for bad in (1, 3, 5): with pytest.raises(ValueError): Settings._still_recipe_roll_count_range(bad) # ── 2롤 ab 변주 (전 백엔드) ──────────────────────────────────────────── def test_ab_roll_prompt_map_varies_second_roll(): m = build_ab_roll_prompt_map("BASE PROMPT", ["a", "b"]) assert m["a"] == "BASE PROMPT" assert m["b"].startswith("BASE PROMPT\n\n") assert "COMPOSITION VARIATION" in m["b"] def test_ab_roll_prompt_map_single_label_is_base(): assert build_ab_roll_prompt_map("BASE", ["a"]) == {"a": "BASE"} # ── 팩 v18/v10 스템 실재·로드 ───────────────────────────────────────── def test_still_recipe_v18_resolves_and_loads(): assert resolve_prompt_version("18") == "18.202608131121" assert IDENTITY_ROLE_PROMPT_VERSION == "18" assert SUPPORT_STILLNESS_PROMPT_VERSION == "18" clause = build_identity_ref_role_clause(True) assert "CHARACTER REFERENCE ROLE" in clause assert "covered or hidden" in clause # 얼굴 가림 우선 조항 (S65sh10) assert build_identity_ref_role_clause(False) == "" def test_judge_v10_resolves_and_loads(): assert resolve_judge_pack_version("10") == "10.202608131121" assert QK_CRITIQUE_PACK_VERSION == "10" assert STILL_FIX_REJUDGE_HEADER_PACK_VERSION == "10" header = load_fix_rejudge_header(STILL_FIX_REJUDGE_HEADER_PACK_VERSION) assert "IDENTITY IS NOT TRADABLE" in header # S39sh4 인물 변형 축 def test_roll_prompts_fold_into_fingerprint(tmp_path): """Codex R2: b 변주 절은 roll_prompts 에 실린다 — 공용 지문이 라벨별 전문 해시를 직접 접어, 변주 스템 개정이 그 절을 실은 샷만 정확히 stale 시킨다. roll_prompts 미사용 소비자(배경 플레이트류)는 키 부재 = 기존 지문 보존.""" def one(sub: str, roll_prompts): (tmp_path / sub).mkdir() (tmp_path / sub / "ref.png").write_bytes(b"ref") kw = dict( tag="t1", prompt="PROMPT", labeled_refs=[("REF", tmp_path / sub / "ref.png")], out_stem=tmp_path / sub / "out" / "s1", gen_fn=FakeGen(), judge_fn=_judge, critique_fn=None, fix_gen_fn=None, roll_count=2, critique_enabled=False, fix_head=FIX_HEAD, fix_tail=FIX_TAIL, fix_label="PHOTOGRAPH TO EDIT", record=None, ) if roll_prompts is not None: kw["roll_prompts"] = roll_prompts _, record = run_multiroll_select(**kw) return record["input_fingerprint"] fp_v1 = one("v1", {"A": "PROMPT", "B": "PROMPT\n\nVARIATION 1"}) fp_v1_again = one("v1b", {"A": "PROMPT", "B": "PROMPT\n\nVARIATION 1"}) fp_v2 = one("v2", {"A": "PROMPT", "B": "PROMPT\n\nVARIATION 2"}) fp_none = one("v0", None) assert fp_v1 == fp_v1_again # 결정론 assert fp_v1 != fp_v2 # b 변주 개정 → stale assert fp_none not in (fp_v1, fp_v2) # 키 부재 지문은 별개(구 소비자) def test_recipe_stem_content_hash_exact_stem(): """Codex R2: 스템 단위 해시 — 같은 팩의 동거 스템과 무관해야 한다.""" import hashlib from app.modules.pipeline.still_recipe import ( recipe_stem_content_hash, resolve_prompt_version, ) got = recipe_stem_content_hash("18", "support_stillness_clause") p = (Path(__file__).resolve().parents[3] / "prompts" / "_base" / "still_recipe" / resolve_prompt_version("18") / "support_stillness_clause.md") assert got == hashlib.sha256(p.read_bytes()).hexdigest()[:16] import pytest with pytest.raises(ValueError): recipe_stem_content_hash("18", "no_such_stem") def test_support_stillness_clause_injected_for_people_only(): kw = dict( shot_desc="남자가 문을 연다", place_text="허름한 방.", time_of_day_en="night", world_anchor="", bg_only=False, prev_used=False, prev_usage_en="", pose_clauses=[], movement_en="", figures_en="", carried_en="", char_names=["남자 (male, 40s, stocky)"], ) people = build_still_prompt(**kw) assert "BODY & SUPPORT" in people bg = build_still_prompt(**{**kw, "bg_only": True, "char_names": []}) assert "BODY & SUPPORT" not in bg