Prove asset-level embedding upserts against live PostgreSQL

Constraint: The schema already declared asset-level idempotency, but without live evidence future work could mistake it for an unverified design note.
Rejected: Rely on DDL inspection alone | It would not prove duplicate inserts are blocked and upserts reuse the same embedding row.
Confidence: high
Scope-risk: narrow
Directive: Keep asset-level writer implementations aligned with the verified ON CONFLICT (feature_set_id, asset_id) WHERE window_id IS NULL contract.
Tested: /usr/local/miniconda3/bin/python -m py_compile scripts/validate_audio_embedding_asset_upsert_live.py; git diff --check; /usr/local/miniconda3/bin/python scripts/validate_audio_embedding_asset_upsert_live.py --dsn 'postgres://d2:d2pass@127.0.0.1:5432/d2' --schema acr_asset_upsert_test --output data/pgvector_eval/music20/audio_embedding_asset_upsert_live_report.json
Not-tested: No production semantic writer uses the asset-level contract yet; this commit validates the DB contract, not an end-to-end extractor.

+{
+  "schema": "acr_asset_upsert_test",
+  "dsn_redacted": "postgres://d2:***@127.0.0.1:5432/d2",
+  "seed_ids": {
+    "model_id": 1,
+    "feature_set_id": 1,
+    "canonical_song_id": 1,
+    "work_id": 1,
+    "recording_id": 1,
+    "asset_id": 1
+  },
+  "first_insert_embedding_id": 1,
+  "duplicate_insert_guard": {
+    "passed": true,
+    "error_type": "UniqueViolation",
+    "message": "duplicate key value violates unique constraint \"uq_audio_embedding_feature_asset\""
+  },
+  "upsert_embedding_id": 1,
+  "same_embedding_id_reused": true,
+  "counts": {
+    "audio_embedding": 1,
+    "audio_embedding_vector_192": 1
+  },
+  "final_state": {
+    "embedding_id": 1,
+    "asset_id": 1,
+    "window_id": null,
+    "checksum": "checksum-v2",
+    "embedding_uri": "inline://asset-probe-upsert",
+    "metadata_json": {
+      "probe": "asset_level_upsert_v2"
+    },
+    "vector_literal": "[0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2,0.2]"
+  },
+  "passed": true
+}
\ No newline at end of file

 ## 2026-06-04
 
+- 新增 `scripts/validate_audio_embedding_asset_upsert_live.py` 与 `audio_embedding_asset_upsert_live_report.json`，在隔离 schema `acr_asset_upsert_test` 上真实验证 `uq_audio_embedding_feature_asset`：重复普通 insert 会触发 `UniqueViolation`，而 `ON CONFLICT ... DO UPDATE` 会复用同一 `embedding_id`，最终 `audio_embedding/audio_embedding_vector_192` 行数都保持为 `1`。
 - 新增 `scripts/run_phase1_embedding_preflight_matrix_live.py` 与 `phase1_embedding_preflight_matrix_report.json`，对 `mert / muq / ecapa` 四条 semantic jobs 做了统一 live preflight 矩阵验证；结果表明 4 条 job 全都稳定落到 `preflight_failed`，且 blocker 已收敛为 `/workspace/downloads` 未挂载与语义模型 runtime 缺失，而不是单条 job 的偶发异常。
 - 更新 `run_embedding_job.py`，把 semantic lane 从“只有 dry-run”推进到“真实 scope 读取 + vector table 校验 + runtime 依赖校验 + 缺音频校验 + PostgreSQL failed 落账”的 preflight write contract；当前 live `mert` job 会把 `unreadable_audio_assets` 与 `model_runtime_unavailable` 同时写入 `feature_extraction_job.metadata_json`，不再只停留在纸面设计。
 - 给 `audio_embedding` 补上 `UNIQUE(feature_set_id, window_id) WHERE window_id IS NOT NULL` 与 `UNIQUE(feature_set_id, asset_id) WHERE window_id IS NULL AND asset_id IS NOT NULL` 两条幂等唯一键，为后续真实 `MERT / MuQ / ECAPA` upsert 落库固定主键策略。

@@ -343,6 +343,17 @@ MERT 5s/2.5s job (`extraction_job_id=2`) 在 `acr_test` 上已经真实验证：
 
 而不需要先查再写。
 
+当前这两条唯一键里，asset-level 路径也已经有 live 证据：
+
+- `scripts/validate_audio_embedding_asset_upsert_live.py`
+- `audio_embedding_asset_upsert_live_report.json`
+
+已验证：
+
+- 重复 `INSERT` 会被 `uq_audio_embedding_feature_asset` 拒绝
+- `ON CONFLICT ... DO UPDATE` 会复用同一个 `embedding_id`
+- `audio_embedding` / `audio_embedding_vector_192` 行数都保持为 `1`
+
 ### 下一步替换点
 
 当 runtime 与音频挂载到位后，只需要把 guarded failure path 替换成真实 inference：

@@ -774,3 +774,40 @@ cd /workspace/acr-engine
 - 当前真正阻塞 Phase-1 encoder-only 落地的是：
   1. `/workspace/downloads` 音频挂载
   2. 模型 runtime 依赖安装
+
+
+## 新增：asset-level embedding upsert live 验证
+
+为了把 `uq_audio_embedding_feature_asset` 从“DDL 声明”推进到“真实证据”，本轮新增：
+
+- `acr-engine/scripts/validate_audio_embedding_asset_upsert_live.py`
+- `acr-engine/data/pgvector_eval/music20/audio_embedding_asset_upsert_live_report.json`
+
+### 验证动作
+
+脚本会在隔离 schema `acr_asset_upsert_test` 中：
+
+1. 落最小主数据图：`song -> work -> recording -> asset`
+2. 插入第一条 `window_id IS NULL` 的 asset-level embedding
+3. 再做一次普通重复 `INSERT`
+4. 预期被 `uq_audio_embedding_feature_asset` 拒绝
+5. 再做一次 `ON CONFLICT ... DO UPDATE`
+6. 验证最终仍只有 `1` 条 `audio_embedding` 与 `1` 条 `audio_embedding_vector_192`
+
+### 当前结果
+
+| 项 | 结果 |
+|---|---|
+| 首次 `embedding_id` | `1` |
+| 重复普通 `INSERT` | `UniqueViolation` |
+| 唯一键名 | `uq_audio_embedding_feature_asset` |
+| upsert 后 `embedding_id` | `1` |
+| `same_embedding_id_reused` | `true` |
+| `audio_embedding` 行数 | `1` |
+| `audio_embedding_vector_192` 行数 | `1` |
+| 最终 `checksum` | `checksum-v2` |
+
+结论：
+
+- asset-level 唯一键不是“纸面存在”，而是已经在 live PostgreSQL 上真实生效
+- 后续如果补 asset-level semantic writer，可以直接沿用同一个 `ON CONFLICT (feature_set_id, asset_id) ...` 合同

@@ -192,6 +192,7 @@ sed -n '1,320p' acr-engine/sql/acr_pg_schema_v2.sql
 - semantic lane 也已完成 live failure contract：`run_embedding_job.py` 现在会同时暴露 `unreadable_audio_assets` 与 `model_runtime_unavailable`，而不是把失败伪装成 completed
 - `audio_embedding` 已补上 window / asset 双路唯一键，后续真实 encoder 只需替换 inference adapter 即可复用同一 upsert 合同
 - `scripts/run_phase1_embedding_preflight_matrix_live.py` 已跑通，4 条 semantic jobs（mert/muq/ecapa）在 `acr_test` 上都被稳定标记为 `preflight_failed`；当前共性 blocker 已收敛为 `/workspace/downloads` 缺失 + 语义模型 runtime 缺失
+- `scripts/validate_audio_embedding_asset_upsert_live.py` 已在隔离 schema `acr_asset_upsert_test` 上验证 `uq_audio_embedding_feature_asset`：重复 insert 会被唯一键拒绝，upsert 会复用同一 `embedding_id`，说明 asset-level 幂等键也已有真实证据
 - `phase1_hot_reference_v1` 在 `acr_test` 里已经真实补齐 `20` 个 reference members，因此 worker dry-run 当前看到的 scope 已是 `20 recordings / 20 assets / 20 windows`
 - worker contract 现在已有基础前置状态保护；重复执行同一 chromaprint dry-run job 会被 `expected_status=pending` 明确拒绝，证据见 `phase1_worker_double_claim_guard_report.json`
 - exact lane 的 `run_chromaprint_job.py` 已具备非 dry-run 写入路径；当前在 `acr_test` 的 live 结果是因为 `/workspace/downloads/...` 缺失而明确 `failed`，不是继续假装 `completed`