新增 PostgreSQL 去重检索链路与 hard 评估集支持

- 新增 PostgreSQL 导入脚本、评估脚本和 schema 定义，支持基于 exact_hash、pg_trgm 和行级 hash 的三层召回策略
- 评估 CLI 新增 hard profile，覆盖错别字、OCR 错误、整段翻译、medley 片段等更贴近业务边界的场景
- 调整 checker.py 复核阈值与匹配理由文案，优化翻译行相似与仅副歌重复场景的判定逻辑
- 同步更新 README、TEST_WORKFLOW 和单元测试

Co-Authored-By: Claude <noreply@anthropic.com>

@@ -85,15 +85,33 @@ python -m lyric_dedup.cli generate-eval-set \
   --positive-ratio 0.3
 ```
 
-生成器的业务口径：
+默认 `--profile standard` 生成常规生产评估集。也可以生成更贴近业务边界的 hard 集：
+
+```bash
+python -m lyric_dedup.cli generate-eval-set \
+  --profile hard \
+  --library-dir data/library \
+  --lyrics-dir data/generated_eval/hard_incoming \
+  --csv data/generated_eval/eval_hard_5000.csv \
+  --eval-index data/generated_eval/eval_hard_5000.csv.index.pkl \
+  --size 5000 \
+  --positive-ratio 0.3
+```
+
+standard 业务口径：
 
 - 先扫描整个曲库，按有效歌词行数、语言类型、文件来源前缀做分层采样，不再按排序前缀取样。
-- `应去重` 样本只生成全曲歌词的样式变化，例如时间戳、标点、平台噪声、空行、重复副歌次数变化、附加中文翻译。
+- `应去重` 样本只生成全曲歌词的样式变化，例如时间戳、标点、平台噪声、空行、重复副歌次数变化、附加中文翻译、少量错别字/英文拼写错误。
 - `不应去重` 样本以真实 holdout 完整歌词为主，也包含片段歌词、重复副歌碰撞、仅翻译相似、同主题新歌词、短歌词/占位边界样本。
 - 片段歌词即使命中已有歌曲的一部分，也不应该输出 `duplicate`；最多进入 `review`。
 - 生成器会额外写出 `--eval-index`，这个索引排除了 holdout 歌，评估生成 CSV 时应使用它。
 - 同时会生成 `*.manifest.json`，记录 seed、曲库规模、holdout 数、样本类型分布、语言/来源分桶和样本来源覆盖数。
 
+hard 业务口径不故意制造反常输入，主要覆盖上线更容易踩边界的情况：
+
+- `应去重`: 同曲平台版本噪声、较完整歌词缺少一段、整段中文翻译附加、较真实的录入/OCR 错别字、时间戳和平台元信息混合。
+- `不应去重`: 真实 holdout 新歌、从 holdout 中优先挑选和曲库有行重合的近邻新歌、较长但不完整的单曲片段、多曲 medley/串烧式片段、重复副歌碰撞、仅翻译相似、短歌词边界。
+
 先准备一个 CSV，例如 `data/eval/eval.csv`：
 
 ```csv

@@ -108,6 +108,20 @@ python -m lyric_dedup.cli generate-eval-set \
   --positive-ratio 0.3
 ```
 
+如需生成更贴近业务边界的 hard 口径测试集：
+
+```bash
+python -m lyric_dedup.cli generate-eval-set \
+  --profile hard \
+  --library-dir data/library \
+  --lyrics-dir data/generated_eval/hard_incoming \
+  --csv data/generated_eval/eval_hard_5000.csv \
+  --index outputs/indexes/library_lyrics.pkl \
+  --eval-index data/generated_eval/eval_hard_5000.csv.index.pkl \
+  --size 5000 \
+  --positive-ratio 0.3
+```
+
 默认生产评估口径：
 
 ```text
@@ -120,7 +134,7 @@ python -m lyric_dedup.cli generate-eval-set \
 业务口径：
 
 ```text
-positive_* = 应去重，全曲歌词样式变化
+positive_* = 应去重，全曲歌词样式变化，包括少量错别字/英文拼写错误扰动
 negative_real_holdout_full_song = 不应去重，完整真实歌词，已从评估索引中排除
 negative_fragment = 不应去重，单曲片段
 negative_shared_chorus = 不应去重，重复副歌碰撞
@@ -129,6 +143,15 @@ negative_same_theme_synthetic = 不应去重，同主题新歌词
 edge_short_or_placeholder = 不应去重，短歌词/占位边界样本
 ```
 
+hard 口径额外强调真实业务边界，而不是故意制造反常难题：
+
+```text
+positive_realistic_variant = 应去重，同曲平台版本噪声、较完整缺段、整段翻译附加、真实录入/OCR 错
+negative_near_neighbor_holdout_full_song = 不应去重，和曲库有较多行重合的真实 holdout 新歌
+negative_long_fragment = 不应去重，较长但不完整的单曲片段
+negative_catalog_mashup = 不应去重，多首真实歌词片段组成的串烧/混剪式输入
+```
+
 生成器会扫描整个曲库并按有效歌词行数、语言类型、文件来源前缀分层采样。它会分出一批 holdout 完整歌词作为真实新歌负样本，并生成一个排除 holdout 的评估索引。每次还会输出：
 
 ```text

@@ -5,7 +5,7 @@ from __future__ import annotations
 import hashlib
 import pickle
 from dataclasses import dataclass
-from enum import StrEnum
+from enum import Enum
 from pathlib import Path
 
 from lyric_dedup.minhash_lsh import MinHashConfig
@@ -16,7 +16,7 @@ from lyric_dedup.normalization import lyric_tokens
 from lyric_dedup.normalization import normalize_lyrics
 
 
-class DuplicateDecision(StrEnum):
+class DuplicateDecision(str, Enum):
     DUPLICATE = "duplicate"
     REVIEW = "review"
     NEW = "new"

@@ -53,6 +53,12 @@ def main() -> None:
     generate.add_argument("--seed", type=int, default=20260602)
     generate.add_argument("--index", default="", help="optional source index path recorded in the manifest")
     generate.add_argument("--eval-index", default="", help="output index built from non-holdout records for this eval set")
+    generate.add_argument(
+        "--profile",
+        choices=("standard", "hard"),
+        default="standard",
+        help="evaluation sample profile: standard production mix or harder business-realistic edge mix",
+    )
 
     args = parser.parse_args()
     if args.command == "build-index":
@@ -80,6 +86,7 @@ def main() -> None:
             seed=args.seed,
             index_path=Path(args.index) if args.index else None,
             eval_index_path=Path(args.eval_index) if args.eval_index else None,
+            profile=args.profile,
         )
         print(json.dumps(summary, ensure_ascii=False))
 

+# Test runner
+pytest>=8.0
+
+# PostgreSQL storage prototype
+psycopg[binary]>=3.2
+
+# Existing MySQL/COS lyric download utilities
+pymysql>=1.1
+cos-python-sdk-v5>=1.9
+tqdm>=4.66

+"""Initialize PostgreSQL schema for lyric dedup storage."""
+
+from __future__ import annotations
+
+import argparse
+import sys
+from pathlib import Path
+
+
+PROJECT_ROOT = Path(__file__).resolve().parents[1]
+SCHEMA_PATH = PROJECT_ROOT / "scripts" / "postgres_schema.sql"
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Initialize PostgreSQL schema for lyric dedup.")
+    parser.add_argument("--dsn", required=True, help="PostgreSQL DSN, e.g. postgresql://user:pass@localhost:5432/lyric_dedup")
+    parser.add_argument("--schema", default=str(SCHEMA_PATH))
+    args = parser.parse_args()
+
+    psycopg = _import_psycopg()
+    schema_sql = Path(args.schema).read_text(encoding="utf-8")
+    with psycopg.connect(args.dsn) as conn:
+        with conn.cursor() as cursor:
+            cursor.execute(schema_sql)
+        conn.commit()
+    print(f"initialized schema from {args.schema}")
+
+
+def _import_psycopg():
+    try:
+        import psycopg
+
+        return psycopg
+    except ModuleNotFoundError:
+        print(
+            "Missing dependency: psycopg. Install it with:\n"
+            "  python -m pip install 'psycopg[binary]'",
+            file=sys.stderr,
+        )
+        raise SystemExit(1)
+
+
+if __name__ == "__main__":
+    main()

+create extension if not exists pg_trgm;
+
+create table if not exists lyrics (
+  id bigserial primary key,
+  record_id text not null unique,
+  source_path text not null,
+  title text,
+  artist text,
+  raw_text text not null,
+  normalized_text text not null,
+  primary_text text not null,
+  translation_text text,
+  exact_hash text not null,
+  split_confidence text,
+  split_reason text,
+  line_count integer not null,
+  created_at timestamptz not null default now(),
+  updated_at timestamptz not null default now(),
+  deleted_at timestamptz
+);
+
+create index if not exists lyrics_exact_hash_idx
+on lyrics (exact_hash)
+where deleted_at is null;
+
+create index if not exists lyrics_primary_text_trgm_idx
+on lyrics using gin (primary_text gin_trgm_ops);
+
+create table if not exists lyric_lines (
+  lyric_id bigint not null references lyrics(id) on delete cascade,
+  role text not null,
+  line_no integer not null,
+  normalized_line text not null,
+  line_hash text not null,
+  primary key (lyric_id, role, line_no)
+);
+
+create index if not exists lyric_lines_hash_idx
+on lyric_lines (line_hash);
+
+create index if not exists lyric_lines_lyric_id_idx
+on lyric_lines (lyric_id);

@@ -316,6 +316,40 @@ def test_generated_eval_set_uses_stratified_production_mix(tmp_path) -> None:
     assert all(row["expected"] == "不应去重" for row in rows if row["sample_type"].startswith("negative_"))
 
 
+def test_generated_hard_eval_set_uses_business_realistic_edge_mix(tmp_path) -> None:
+    library = tmp_path / "library"
+    incoming = tmp_path / "generated" / "incoming"
+    eval_csv = tmp_path / "generated" / "eval_hard.csv"
+    library.mkdir()
+    for idx in range(24):
+        prefix = "AY" if idx % 3 == 0 else "WHHY"
+        lyric = BASE_LYRIC.replace("我爱你", f"我想你{idx}").replace("城市", f"城市{idx}")
+        if idx % 4 == 0:
+            lyric += "\nI miss you tonight\nUnder the moonlight\nNever let me go\n"
+        (library / f"{idx}_{prefix}{idx:06d}.txt").write_text(lyric, encoding="utf-8")
+
+    generate_eval_set(
+        library_dir=library,
+        output_dir=incoming,
+        csv_path=eval_csv,
+        size=40,
+        positive_ratio=0.3,
+        profile="hard",
+    )
+
+    rows = list(csv.DictReader(eval_csv.open(encoding="utf-8")))
+    manifest = json.loads((tmp_path / "generated" / "eval_hard.csv.manifest.json").read_text(encoding="utf-8"))
+    sample_types = {row["sample_type"] for row in rows}
+
+    assert len(rows) == 40
+    assert manifest["profile"] == "hard"
+    assert "positive_realistic_variant" in manifest["plan"]
+    assert "negative_near_neighbor_holdout_full_song" in manifest["plan"]
+    assert "negative_long_fragment" in sample_types
+    assert "negative_catalog_mashup" in sample_types
+    assert any(row["sample_type"].startswith("positive_") for row in rows)
+
+
 def test_foreign_original_with_added_chinese_translation_is_duplicate() -> None:
     checker = DuplicateChecker()
     checker.add_record(