feat(search): Phase 2Q Diagnose Phase 1B — scaffold + dispatcher

phase-2q-query-rewrite-diagnose.md v6 plan Phase 1 의 fixture 외 잔여.
Phase 1A 446ba82 위 dispatcher + cache + LLM call + API param + eval flag + 21 unit test.
retrieval 합성 (search_with_rewrite) 은 Phase 2 별 commit.

신규:
- app/services/search/query_rewriter.py — LLM_BACKEND_MAP + _resolve + cache + rewrite()
  · slug-based allowlist (no silent fallback), httpx 직접, Priority.FOREGROUND semaphore
  · sampling 박제 (gemma response_format json_object / qwen prompt rule only — Phase 0 inspect 9)
  · manual TTL cache (query_analyzer 패턴 1:1, sha256[:32] NFKC key, LLM_REWRITE_TIMEOUT_MS=15000)
- tests/test_query_rewriter.py — 21 test PASS (resolve / cache key / parser / cache TTL / constants)

수정:
- app/api/search.py — ?rewrite_backend= query param + 400 unknown / 503 unavailable.
  scaffold = call but discard variants (retrieval path 영향 0). Phase 2 에서 합성.
- tests/search_eval/run_eval.py — --rewrite-backend flag + 4 hot spot wire-up.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

tests/search_eval/run_eval.py

@@ -203,6 +203,7 @@ async def call_search(
     snapshot_doc_id_max: int | None = None,
     snapshot_chunk_id_max: int | None = None,
     reranker_backend: str | None = None,
+    rewrite_backend: str | None = None,
 ) -> tuple[list[int], float]:
     """검색 API 호출 → (doc_ids, latency_ms)."""
     url = f"{base_url.rstrip('/')}/api/search/"
@@ -222,6 +223,8 @@ async def call_search(
         params["snapshot_chunk_id_max"] = snapshot_chunk_id_max
     if reranker_backend is not None:
         params["reranker_backend"] = reranker_backend
+    if rewrite_backend is not None:
+        params["rewrite_backend"] = rewrite_backend
 
     import time
 
@@ -253,6 +256,7 @@ async def evaluate(
     snapshot_doc_id_max: int | None = None,
     snapshot_chunk_id_max: int | None = None,
     reranker_backend: str | None = None,
+    rewrite_backend: str | None = None,
 ) -> list[QueryResult]:
     """전체 쿼리셋 평가."""
     results: list[QueryResult] = []
@@ -266,6 +270,7 @@ async def evaluate(
                     snapshot_doc_id_max=snapshot_doc_id_max,
                     snapshot_chunk_id_max=snapshot_chunk_id_max,
                     reranker_backend=reranker_backend,
+                    rewrite_backend=rewrite_backend,
                 )
                 results.append(
                     QueryResult(
@@ -843,6 +848,7 @@ async def call_search_full(
     snapshot_doc_id_max: int | None = None,
     snapshot_chunk_id_max: int | None = None,
     reranker_backend: str | None = None,
+    rewrite_backend: str | None = None,
 ) -> tuple[list[dict], float]:
     """call_search와 동일 로직. 단 full result dict 리스트 반환."""
     url = f"{base_url.rstrip('/')}/api/search/"
@@ -864,6 +870,8 @@ async def call_search_full(
         params["snapshot_chunk_id_max"] = snapshot_chunk_id_max
     if reranker_backend is not None:
         params["reranker_backend"] = reranker_backend
+    if rewrite_backend is not None:
+        params["rewrite_backend"] = rewrite_backend
 
     import time
 
@@ -1322,6 +1330,12 @@ def main() -> int:
         default=None,
         help="Phase 2B Diagnose reranker dispatcher slug (baseline | cand_gte_ml_base). 미지정 = production.",
     )
+    parser.add_argument(
+        "--rewrite-backend",
+        type=str,
+        default=None,
+        help="Phase 2Q Diagnose query rewrite dispatcher slug (baseline | cand_multi_query_macmini | cand_multi_query_macbook). 미지정 = single-query path. Phase 1B scaffold = variants 박제만, retrieval 합성은 Phase 2.",
+    )
 
     args = parser.parse_args()
 
@@ -1375,21 +1389,21 @@ def main() -> int:
         if args.base_url:
             print(f"\n>>> evaluating: {args.base_url}")
             results = asyncio.run(
-                evaluate(queries, args.base_url, args.token, "single", mode=args.mode, fusion=args.fusion, rerank=args.rerank, analyze=args.analyze, embedding_backend=args.embedding_backend, snapshot_doc_id_max=args.snapshot_doc_id_max, snapshot_chunk_id_max=args.snapshot_chunk_id_max, reranker_backend=args.reranker_backend)
+                evaluate(queries, args.base_url, args.token, "single", mode=args.mode, fusion=args.fusion, rerank=args.rerank, analyze=args.analyze, embedding_backend=args.embedding_backend, snapshot_doc_id_max=args.snapshot_doc_id_max, snapshot_chunk_id_max=args.snapshot_chunk_id_max, reranker_backend=args.reranker_backend, rewrite_backend=args.rewrite_backend)
             )
             print_summary("single", results, eval_version=args.eval_version)
             all_results.extend(results)
         else:
             print(f"\n>>> baseline: {args.baseline_url}")
             baseline_results = asyncio.run(
-                evaluate(queries, args.baseline_url, args.token, "baseline", mode=args.mode, fusion=args.fusion, rerank=args.rerank, analyze=args.analyze, embedding_backend=args.embedding_backend, snapshot_doc_id_max=args.snapshot_doc_id_max, snapshot_chunk_id_max=args.snapshot_chunk_id_max, reranker_backend=args.reranker_backend)
+                evaluate(queries, args.baseline_url, args.token, "baseline", mode=args.mode, fusion=args.fusion, rerank=args.rerank, analyze=args.analyze, embedding_backend=args.embedding_backend, snapshot_doc_id_max=args.snapshot_doc_id_max, snapshot_chunk_id_max=args.snapshot_chunk_id_max, reranker_backend=args.reranker_backend, rewrite_backend=args.rewrite_backend)
             )
             baseline_summary = print_summary("baseline", baseline_results, eval_version=args.eval_version)
 
             print(f"\n>>> candidate: {args.candidate_url}")
             candidate_results = asyncio.run(
                 evaluate(
-                    queries, args.candidate_url, args.token, "candidate", mode=args.mode, fusion=args.fusion, rerank=args.rerank, analyze=args.analyze, embedding_backend=args.embedding_backend, snapshot_doc_id_max=args.snapshot_doc_id_max, snapshot_chunk_id_max=args.snapshot_chunk_id_max, reranker_backend=args.reranker_backend
+                    queries, args.candidate_url, args.token, "candidate", mode=args.mode, fusion=args.fusion, rerank=args.rerank, analyze=args.analyze, embedding_backend=args.embedding_backend, snapshot_doc_id_max=args.snapshot_doc_id_max, snapshot_chunk_id_max=args.snapshot_chunk_id_max, reranker_backend=args.reranker_backend, rewrite_backend=args.rewrite_backend
                 )
             )
             candidate_summary = print_summary("candidate", candidate_results, eval_version=args.eval_version)

tests/test_query_rewriter.py

@@ -0,0 +1,212 @@
+"""Phase 2Q Diagnose Phase 1B — query_rewriter scaffold + dispatcher 단위 테스트.
+
+가드레일 (plan v6 §5 + §7 Phase 1):
+  1. `_resolve_rewrite_backend` — slug resolve, unknown ValueError, baseline → None
+  2. `_cache_key` — deterministic + NFKC normalize + backend slug 분리
+  3. `_extract_variants` — valid shape / wrong count / type mismatch / empty / non-list
+  4. cache set/get/TTL (LRU evict 시뮬레이션)
+  5. `allowed_slugs` — LLM_BACKEND_MAP keys 1:1
+"""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+import os
+import sys
+import time
+
+import pytest
+
+# logs/llm_gate.log 가 root 소유 (운영 fastapi daemon write) → pytest 가 hyungi user 로
+# import 시 PermissionError. 본 test 한정 FileHandler safe-wrap (다른 test 영향 0).
+_orig_file_handler = logging.FileHandler
+
+def _safe_file_handler(filename, *args, **kwargs):  # type: ignore
+    try:
+        return _orig_file_handler(filename, *args, **kwargs)
+    except PermissionError:
+        return logging.NullHandler()
+
+logging.FileHandler = _safe_file_handler  # type: ignore[assignment]
+
+# tests/ → 프로젝트 루트 → app/
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "app"))
+
+from services.search import query_rewriter
+from services.search.query_rewriter import (
+    EXPECTED_N_VARIANTS,
+    LLM_BACKEND_MAP,
+    PROMPT_VERSION,
+    _cache_key,
+    _extract_variants,
+    _resolve_rewrite_backend,
+    allowed_slugs,
+)
+
+
+# ─── 1. _resolve_rewrite_backend ──────────────────────────
+
+
+def test_resolve_baseline_returns_none():
+    assert _resolve_rewrite_backend(None) is None
+    assert _resolve_rewrite_backend("baseline") is None
+
+
+def test_resolve_known_slugs():
+    cfg = _resolve_rewrite_backend("cand_multi_query_macmini")
+    assert cfg is not None
+    assert "endpoint" in cfg and "model" in cfg and "sampling" in cfg
+    assert cfg["model"] == "gemma-4-26b-a4b-it-8bit"
+
+    cfg = _resolve_rewrite_backend("cand_multi_query_macbook")
+    assert cfg is not None
+    assert cfg["model"] == "mlx-community/Qwen3.6-27B-8bit"
+    # qwen sampling 에 response_format 없음 (Phase 0 inspect 9 박제)
+    assert "response_format" not in cfg["sampling"]
+
+
+def test_resolve_unknown_slug_raises():
+    with pytest.raises(ValueError, match="unknown_rewrite_backend"):
+        _resolve_rewrite_backend("cand_bogus")
+    with pytest.raises(ValueError):
+        _resolve_rewrite_backend("cand_multi_query_other")
+
+
+def test_allowed_slugs_matches_map():
+    assert allowed_slugs() == list(LLM_BACKEND_MAP.keys())
+    assert "baseline" in allowed_slugs()
+    assert "cand_multi_query_macmini" in allowed_slugs()
+    assert "cand_multi_query_macbook" in allowed_slugs()
+
+
+# ─── 2. _cache_key ────────────────────────────────────────
+
+
+def test_cache_key_deterministic():
+    k1 = _cache_key("산업안전보건법 제6장", "cand_multi_query_macmini")
+    k2 = _cache_key("산업안전보건법 제6장", "cand_multi_query_macmini")
+    assert k1 == k2
+    assert len(k1) == 32  # sha256[:32]
+
+
+def test_cache_key_nfkc_normalize_and_strip_lower():
+    # whitespace + uppercase → 동일 key
+    base = _cache_key("ASME Section VIII", "cand_multi_query_macmini")
+    assert _cache_key("  asme section viii  ", "cand_multi_query_macmini") == base
+    assert _cache_key("ASME SECTION VIII", "cand_multi_query_macmini") == base
+
+
+def test_cache_key_differs_by_backend_slug():
+    k_a = _cache_key("query", "cand_multi_query_macmini")
+    k_b = _cache_key("query", "cand_multi_query_macbook")
+    assert k_a != k_b
+
+
+def test_cache_key_includes_prompt_version():
+    # PROMPT_VERSION 변경 시 cache 분리 — 직접 test 어렵지만 raw 구성 확인
+    assert PROMPT_VERSION == "v1"
+    k = _cache_key("query", "cand_multi_query_macmini")
+    assert len(k) == 32
+
+
+# ─── 3. _extract_variants ─────────────────────────────────
+
+
+def test_extract_variants_valid_shape():
+    raw = '{"variants": ["원본", "한국어 변형", "english"]}'
+    out = _extract_variants(raw, expected_n=3)
+    assert out == ["원본", "한국어 변형", "english"]
+
+
+def test_extract_variants_strips_whitespace():
+    raw = '{"variants": ["  원본  ", "한국어\\n", " english "]}'
+    out = _extract_variants(raw, expected_n=3)
+    assert out == ["원본", "한국어", "english"]
+
+
+def test_extract_variants_wrong_count_returns_none():
+    raw = '{"variants": ["only_one"]}'
+    assert _extract_variants(raw, expected_n=3) is None
+    raw = '{"variants": ["a", "b", "c", "d"]}'
+    assert _extract_variants(raw, expected_n=3) is None
+
+
+def test_extract_variants_missing_key_returns_none():
+    raw = '{"queries": ["a", "b", "c"]}'
+    assert _extract_variants(raw, expected_n=3) is None
+
+
+def test_extract_variants_non_list_returns_none():
+    raw = '{"variants": "single string"}'
+    assert _extract_variants(raw, expected_n=3) is None
+
+
+def test_extract_variants_empty_string_returns_none():
+    raw = '{"variants": ["a", "", "c"]}'
+    assert _extract_variants(raw, expected_n=3) is None
+
+
+def test_extract_variants_non_string_element_returns_none():
+    raw = '{"variants": ["a", 123, "c"]}'
+    assert _extract_variants(raw, expected_n=3) is None
+
+
+def test_extract_variants_invalid_json_returns_none():
+    raw = "not json at all"
+    assert _extract_variants(raw, expected_n=3) is None
+
+
+def test_extract_variants_markdown_fence_fallback():
+    # parse_json_response 가 ```json fenced 블록 내부 추출 — production parser 재사용 검증
+    raw = '```json\n{"variants": ["a", "b", "c"]}\n```'
+    out = _extract_variants(raw, expected_n=3)
+    assert out == ["a", "b", "c"]
+
+
+# ─── 4. cache set / get ───────────────────────────────────
+
+
+@pytest.mark.asyncio
+async def test_cache_set_get_roundtrip():
+    # 격리: 전역 _CACHE 초기화 (다른 테스트와 격리)
+    query_rewriter._CACHE.clear()
+    key = _cache_key("__test_unique_key__", "cand_multi_query_macmini")
+    assert await query_rewriter._get_cached(key) is None
+    await query_rewriter._set_cached(key, ["v0", "v1", "v2"])
+    out = await query_rewriter._get_cached(key)
+    assert out == ["v0", "v1", "v2"]
+
+
+@pytest.mark.asyncio
+async def test_cache_ttl_expiry():
+    query_rewriter._CACHE.clear()
+    key = "ttl_test_key"
+    # manual entry with past expire_at
+    query_rewriter._CACHE[key] = (time.time() - 1.0, ["a", "b", "c"])
+    assert await query_rewriter._get_cached(key) is None
+    # lazy delete verify
+    assert key not in query_rewriter._CACHE
+
+
+@pytest.mark.asyncio
+async def test_cache_returns_copy_not_reference():
+    """_get_cached 반환 list 를 외부에서 수정해도 internal cache 안전."""
+    query_rewriter._CACHE.clear()
+    key = "copy_test_key"
+    await query_rewriter._set_cached(key, ["a", "b", "c"])
+    out = await query_rewriter._get_cached(key)
+    out.append("mutated")
+    out2 = await query_rewriter._get_cached(key)
+    assert out2 == ["a", "b", "c"]
+
+
+# ─── 5. constants ─────────────────────────────────────────
+
+
+def test_constants_match_plan_v6():
+    assert PROMPT_VERSION == "v1"
+    assert EXPECTED_N_VARIANTS == 3
+    assert query_rewriter.LLM_REWRITE_TIMEOUT_MS == 15000
+    assert query_rewriter.CACHE_TTL == 86400
+    assert query_rewriter.CACHE_MAXSIZE == 1000