強化 PChome rescore 漏掃與錯配防線

TODO_NEXT_STEPS.txt

@@ -4,6 +4,7 @@
 ================================================================================
 
 【已完成】
+   - V10.475 補 PChome rescore 操作與高分錯配防線：`scripts/audit_competitor_match_attempt_rescore.py` 預設不再只掃 `strong_exact_spec_match`，避免漏掉 `focused_exact_*` 等新版 matcher 理由；matcher 新增暖燈 S/M/L 尺寸差、NITORI 香氛噴霧器型號差的 hard veto，並把彩妝色號單邊出現的高分候選送進 `variant_selection_review`，避免 LA MER 氣墊等色號型商品被誤入 accepted queue。測試：`tests/test_marketplace_product_matcher.py`、`tests/test_competitor_match_attempts_persistence.py`、`tests/test_competitor_match_attempt_rescore_audit.py` 通過。
    - V10.474 補 PChome near-threshold matcher / feeder 下一階段：新增 HOOOME 白色經典香氛暖燈與 Gdesign Aroma Lava 2.0 的窄範圍 total-price exact 回收；Recipe Box 可撕式水性兒童指甲油只進 identity_review，不自動寫正式價差；Pavaruni 蠟燭 vs 精油、DASHING DIVA 不同款式仍維持 veto/低信心。known-id refresh 現在會對 hard-veto 舊候選執行 fresh search recovery，missing known-id 若 fresh search 只找到低分候選也會保留 best candidate + diagnostics，而非落成 `refresh_no_result`；正式覆寫保護新增 stronger existing guard，避免較弱新候選以高分覆蓋既有強正式配對。測試：`tests/test_marketplace_product_matcher.py`、`tests/test_competitor_match_attempts_persistence.py`、`tests/test_competitor_match_attempt_rescore_audit.py` 通過。
    - V10.473 補背景 embedding host_health skip：`allow_111_fallback=false` 會讀最近 `host_health_probes`，跳過 runtime unhealthy 的 GCP 節點（預設 20 分鐘，DB fail-open），避免每筆任務都等待已知壞節點 timeout；路由安全不變，不把背景 embedding 落 111。
    - V10.472 補 GCP Ollama failover rootless 診斷：新增 `scripts/ops/diagnose_ollama_gcp_failover.sh`，可一鍵檢查 GCP-A direct、GCP-B direct、111、110:11435、110:11436 與 GCP-B `bge-m3` runtime；目前輸出確認 GCP-A `11434` refused、GCP-B direct/embed OK、110:11435 502、110:11436 OK。110 無免密 sudo、GCP-A 22 refused、GCP-B SSH key denied，因此 primary 修復需 GCP/SSH 或 110 root 權限；應用層維持 GCP-A → GCP-B → 111，不把背景 embedding 落 111。

config.py

@@ -350,7 +350,7 @@ YOUTUBE_API_KEY = os.getenv('YOUTUBE_API_KEY', '')
 # ==========================================
 # 系統版本與路徑
 # ==========================================
-SYSTEM_VERSION = "V10.474"
+SYSTEM_VERSION = "V10.475"
 LOG_FILE_PATH = os.path.join(BASE_DIR, 'logs/system.log')
 public_url = PUBLIC_URL  # 用於模板顯示
 

docs/memory/current_execution_queue_20260524.md

@@ -28,6 +28,7 @@
 - 2026-05-25 14:10 CST 起，`V10.472` 補 rootless GCP Ollama failover 診斷腳本與 DevOps SOP：`scripts/ops/diagnose_ollama_gcp_failover.sh` 會檢查 direct GCP-A/GCP-B/111、110 proxy `11435/11436` 與 GCP-B `bge-m3` runtime。現況輸出：GCP-A direct `/api/version` failed/refused、GCP-B direct OK、111 OK、110:11435 502、110:11436 OK、GCP-B embed OK；110 無免密 sudo，`ssh gcp-a` 22 refused、`ssh gcp-b` publickey denied，因此 primary 修復需 GCP/SSH 或 110 root 權限。
 - 2026-05-25 14:12 CST 起，`V10.473` 進行背景 embedding host_health skip：`allow_111_fallback=false` 的背景 embedding 會讀最近 `host_health_probes`，若 GCP-A/GCP-B runtime 已被標 unhealthy，直接跳過該節點並開 GCP circuit，不等待 30 秒 timeout、不落 111；DB 讀取失敗 fail-open。
 - 2026-05-25 14:45 CST 起，`V10.474` 補 PChome near-threshold matcher / feeder 下一階段：HOOOME 白色經典香氛暖燈、Gdesign Aroma Lava 2.0 進 total-price exact；Recipe Box 可撕式水性兒童指甲油保留 identity_review，不自動寫正式價差；Pavaruni 蠟燭 vs 精油與 DASHING DIVA 不同款式仍不放行。known-id refresh 會對 hard-veto 舊候選跑 fresh search recovery；missing known-id 若 fresh search 只有低分候選，也保留 best candidate + diagnostics，不再只記 `refresh_no_result`；正式覆寫保護新增 stronger existing guard。
+- 2026-05-25 15:20 CST 起，`V10.475` 補 rescore CLI 與高分錯配防線：audit CLI 預設不再只掃 `strong_exact_spec_match`，避免新版 `focused_exact_*` 理由漏掃；matcher 對香氛暖燈 S/M/L 尺寸差、NITORI 香氛噴霧器型號差直接 hard veto，彩妝色號單邊出現時送 `variant_selection_review`，避免高分但不同 variant 的候選被誤推入 accepted queue。
 - 2026-05-25 12:05 CST 狀態：`main` 已部署到 188，正式 `/health` 為 `V10.467`，待推 Gitea。兩段變更已合併驗證：V10.466 rescore duplicate 改看 latest-state，7 筆 SKU 最新 attempt 全為 `rescore_accepted_current`，`competitor_prices` / `competitor_price_history` 目標計數未變；V10.467 focused exact matcher 在容器內回 `exact / total_price / price_alert_exact`。本輪 recreate `momo-app`、`scheduler`、`telegram-bot`；未使用 `--remove-orphans`，未碰 `momo-db`。Smoke 通過：三容器 healthy、PChome rescore queue API HTTP 200、Gemini 24 小時無 provider 紀錄、Ollama env 順序維持 GCP-A → GCP-B → 111、3 分鐘三容器 log 未見 Traceback / ERROR / CRITICAL / IntegrityError。
 
 ## 1. MOMO / PChome 核心比價準確率
@@ -66,6 +67,7 @@
 - 2026-05-25 12:20 CST 起，matcher 新增 `focused_exact_total_price_safe` 窄範圍通道；目前只覆蓋 3W CLINIC 粉底液 2入、花美水凝膠 3支、The Ordinary 咖啡因 EGCG 30ml、KUSSEN 屁屁膏 3入、Bone 擴香禮盒、1990 融燭燈白色款與 CANMAKE 淚袋盤等已確認同款樣本。這讓高信心 `exact/manual_review` 能轉為 `exact/total_price` 供 rescore pilot 入人工覆核；DASHING DIVA、唇彩、香味、色號/款式敏感商品仍不放行。
 - 2026-05-25 12:25 CST production pilot：SKU `6101639`、`10074951`、`7760902`、`TP00074980000005`、`14774766`、`10142589`、`10262470`、`10262471`、`11308520` 已從 `true_low_confidence` materialize 為 `rescore_accepted_current`，全數 `exact/total_price/price_alert_exact` 且理由含 `focused_exact_total_price_safe`。SKU `6101784` 因「即期品」商業條件不同，刻意保留在 `true_low_confidence`，不納入本輪自動入隊。
 - 2026-05-25 14:45 CST 起，matcher 擴充至香氛/精油近門檻安全 cohort：HOOOME 白色經典香氛暖燈與 Gdesign Aroma Lava 2.0 可進 `exact/total_price/price_alert_exact`；Recipe Box 可撕式水性兒童指甲油只進 `identity_review`，因兒童指甲油仍可能藏色款/款式。DASHING DIVA 與 Pavaruni cross-type 負例已補測試，避免跨款式、跨劑型誤配。
+- 2026-05-25 15:20 CST 起，新增三個正式觀察到的高分負例防線：PRAY 守夜人暖燈 L vs S、NITORI 香氛噴霧器 5510 vs YX168、LA MER 氣墊粉霜通用 listing vs `11 Rosy Ivory` 色號。前兩者 hard veto，後者保留高分但不進 accepted queue。
 
 ## 3. 12 Agent 決策信封整合
 

scripts/audit_competitor_match_attempt_rescore.py

@@ -59,7 +59,14 @@ def main(argv: list[str] | None = None) -> int:
     parser.add_argument("--input", help="JSONL file path, or '-' for stdin. If omitted, query DATABASE_PATH.")
     parser.add_argument("--source", default="pchome")
     parser.add_argument("--status", action="append", dest="statuses", help="Attempt status to include; repeatable.")
-    parser.add_argument("--reason-filter", default="strong_exact_spec_match")
+    parser.add_argument(
+        "--reason-filter",
+        default=None,
+        help=(
+            "Only include rows whose diagnostic_codes contain this text. "
+            "Default is unfiltered so newer focused matcher reasons are not missed."
+        ),
+    )
     parser.add_argument("--limit", type=int, default=100)
     parser.add_argument("--sample-limit", type=int, default=20)
     parser.add_argument("--min-score", type=float, default=MIN_MATCH_SCORE)

services/marketplace_product_matcher.py

@@ -1970,6 +1970,12 @@ def score_marketplace_match(
     aroma_scent_variant_conflict = _has_aroma_scent_variant_conflict(left, right)
     if aroma_scent_variant_conflict:
         reasons.append("aroma_scent_variant_conflict")
+    wax_lamp_size_letter_conflict = _has_wax_lamp_size_letter_conflict(left, right)
+    if wax_lamp_size_letter_conflict:
+        reasons.append("size_letter_variant_conflict")
+    nitori_diffuser_model_conflict = _has_nitori_diffuser_model_conflict(left, right)
+    if nitori_diffuser_model_conflict:
+        reasons.append("nitori_diffuser_model_conflict")
     variant_selection_review = _has_named_variant_selection_review(left, right, shared_anchor)
     if variant_selection_review:
         reasons.append("variant_selection_review")
@@ -2021,6 +2027,10 @@ def score_marketplace_match(
         hard_veto = True
     if aroma_scent_variant_conflict:
         hard_veto = True
+    if wax_lamp_size_letter_conflict:
+        hard_veto = True
+    if nitori_diffuser_model_conflict:
+        hard_veto = True
 
     focused_exact_line_reason = _has_focused_low_score_exact_identity_line(left, right)
     if focused_exact_line_reason in FOCUSED_IDENTITY_REVIEW_ONLY_REASONS:
@@ -3007,6 +3017,59 @@ def _has_aroma_scent_variant_conflict(left: ProductIdentity, right: ProductIdent
     return False
 
 
+def _standalone_size_letter_tokens(identity: ProductIdentity) -> set[str]:
+    text = identity.searchable_name
+    return {
+        match.group(1).lower()
+        for match in re.finditer(r"(?<![a-z0-9])([sml])(?![a-z0-9])", text, re.I)
+    }
+
+
+def _has_wax_lamp_size_letter_conflict(left: ProductIdentity, right: ProductIdentity) -> bool:
+    pair_text = f"{left.searchable_name} {right.searchable_name}"
+    if not any(term in pair_text for term in ("香氛蠟燭暖燈", "蠟燭暖燈", "融蠟燈")):
+        return False
+    left_sizes = _standalone_size_letter_tokens(left)
+    right_sizes = _standalone_size_letter_tokens(right)
+    return bool(left_sizes and right_sizes and not (left_sizes & right_sizes))
+
+
+def _has_nitori_diffuser_model_conflict(left: ProductIdentity, right: ProductIdentity) -> bool:
+    if not ({"nitori", "宜得利家居"} & (left.brand_tokens & right.brand_tokens)):
+        return False
+    if "香氛噴霧器" not in left.searchable_name or "香氛噴霧器" not in right.searchable_name:
+        return False
+    left_models = _extract_model_tokens(left.searchable_name) | set(
+        re.findall(r"(?<![a-z0-9])\d{3,5}(?![a-z0-9])", left.searchable_name)
+    )
+    right_models = _extract_model_tokens(right.searchable_name) | set(
+        re.findall(r"(?<![a-z0-9])\d{3,5}(?![a-z0-9])", right.searchable_name)
+    )
+    return bool(left_models and right_models and not (left_models & right_models))
+
+
+def _makeup_shade_tokens(identity: ProductIdentity) -> set[str]:
+    text = identity.searchable_name
+    tokens = set(_explicit_variant_option_tokens(identity))
+    shade_pattern = (
+        r"(?<![a-z0-9])(?:#|no\.?|色號|號色)?\s*(\d{1,3})\s+"
+        r"(rosy ivory|ivory|beige|sand|fair|light|medium|porcelain|rose)(?![a-z0-9])"
+    )
+    for match in re.finditer(shade_pattern, text, re.I):
+        tokens.add(match.group(1).lower())
+        tokens.add(match.group(2).lower().replace(" ", "_"))
+    return tokens
+
+
+def _has_makeup_shade_selection_gap(left: ProductIdentity, right: ProductIdentity) -> bool:
+    pair_text = f"{left.searchable_name} {right.searchable_name}"
+    if not any(term in pair_text for term in ("氣墊粉霜", "粉底", "粉霜", "蜜粉", "唇釉", "唇膏")):
+        return False
+    left_shades = _makeup_shade_tokens(left)
+    right_shades = _makeup_shade_tokens(right)
+    return bool(left_shades) != bool(right_shades)
+
+
 def _has_taicend_baby_spray_equivalence(left: ProductIdentity, right: ProductIdentity) -> bool:
     brand_tokens = {"taicend", "泰陞"}
     return (
@@ -3734,6 +3797,8 @@ def _has_named_variant_selection_review(
     right: ProductIdentity,
     shared_anchor: str,
 ) -> bool:
+    if _has_makeup_shade_selection_gap(left, right):
+        return True
     left_options = _explicit_variant_option_tokens(left)
     right_options = _explicit_variant_option_tokens(right)
     if bool(left_options) != bool(right_options):

tests/test_competitor_match_attempt_rescore_audit.py

@@ -1,4 +1,6 @@
 import json
+import importlib.util
+from pathlib import Path
 
 from sqlalchemy import create_engine, text
 
@@ -28,6 +30,32 @@ def _create_match_attempts_table(conn):
     """))
 
 
+def _load_rescore_cli_module():
+    module_path = Path(__file__).resolve().parents[1] / "scripts" / "audit_competitor_match_attempt_rescore.py"
+    spec = importlib.util.spec_from_file_location("audit_competitor_match_attempt_rescore_cli", module_path)
+    module = importlib.util.module_from_spec(spec)
+    assert spec and spec.loader
+    spec.loader.exec_module(module)
+    return module
+
+
+def test_match_attempt_rescore_cli_defaults_to_unfiltered_reasons(monkeypatch, capsys):
+    cli = _load_rescore_cli_module()
+    captured = {}
+
+    def fake_build_match_attempt_rescore_audit(_engine, **kwargs):
+        captured.update(kwargs)
+        return {"scanned": 0}
+
+    monkeypatch.setattr(cli, "create_engine", lambda _database_path: object())
+    monkeypatch.setattr(cli, "build_match_attempt_rescore_audit", fake_build_match_attempt_rescore_audit)
+
+    assert cli.main(["--limit", "1"]) == 0
+
+    capsys.readouterr()
+    assert captured["reason_filter"] is None
+
+
 def test_match_attempt_rescore_audit_classifies_current_gate_pass_and_veto():
     from services.competitor_match_attempt_rescore_audit import summarize_match_attempt_rescore
 

tests/test_marketplace_product_matcher.py

@@ -2187,6 +2187,41 @@ def test_marketplace_matcher_keeps_aroma_and_nail_variant_gaps_blocked():
     assert "variant_descriptor_conflict" in dashing_cross_style.reasons
 
 
+def test_marketplace_matcher_blocks_high_score_wax_lamp_and_device_variant_gaps():
+    from services.marketplace_product_matcher import score_marketplace_match
+
+    pray_size_gap = score_marketplace_match(
+        "【韓國PRAY】守夜人金屬香氛蠟燭暖燈-復古金(L/專櫃公司貨)",
+        "【韓國EPOCHSIA x Pray】守夜人金屬香氛蠟燭暖燈(S)-復古金",
+    )
+    nitori_model_gap = score_marketplace_match(
+        "【NITORI 宜得利家居】香氛噴霧器 5510(香氛)",
+        "【NITORI 宜得利家居】香氛噴霧器 YX168 WH",
+    )
+
+    assert pray_size_gap.hard_veto is True
+    assert pray_size_gap.comparison_mode == "not_comparable"
+    assert "size_letter_variant_conflict" in pray_size_gap.reasons
+    assert nitori_model_gap.hard_veto is True
+    assert nitori_model_gap.comparison_mode == "not_comparable"
+    assert "nitori_diffuser_model_conflict" in nitori_model_gap.reasons
+
+
+def test_marketplace_matcher_sends_single_sided_makeup_shade_to_review():
+    from services.marketplace_product_matcher import score_marketplace_match
+
+    diagnostics = score_marketplace_match(
+        "【LA MER 海洋拉娜】奇蹟煥采氣墊粉霜 24g（12g x 2）專櫃公司貨",
+        "《LA MER 海洋拉娜》奇蹟煥采氣墊粉霜 24g（12g x 2）-11 Rosy Ivory",
+    )
+
+    assert diagnostics.hard_veto is False
+    assert diagnostics.score >= 0.76
+    assert diagnostics.price_basis == "manual_review"
+    assert diagnostics.alert_tier == "identity_review"
+    assert "variant_selection_review" in diagnostics.reasons
+
+
 def test_marketplace_matcher_promotes_eaoron_classic_tone_up_cream_exact_line():
     from services.marketplace_product_matcher import score_marketplace_match