feat(sense): T-G1+T-G2 Embedder 与观察埋点

- T-G1 embedder.py：embed() 调 llama-server /v1/embeddings -> 对称 per-vector
  int8 量化（scale=127/max|v|，1024 维余弦扰动 ~1e-4 << 1e-2 验收）；
  超时/连接/非200/空形状 -> EmbedderDown（D-G4 fail-closed）；
  /v1/embeddings OpenAI 兼容透传（双路由组：/sense 前缀 + /v1 无前缀，
  不可用 503）；build_sense_routers 返回列表
- T-G2 observer.py：Observation + 批量缓冲 100ms 刷盘（asyncio.Queue +
  to_thread，D-P10）；三消费方共用；升级阶梯回写 outcome/executed_tier；
  D-G6 无 query 原文列（int8 BLOB + 特征 JSON）
- 测试 +11（量化误差/范围/fail-closed×3/透传/单例/BLOB 持久化/批量刷盘/
  回写/无原文列），全量 381->386->381? 校正：386 passed
- 待真机项：llama-server embedder 端点冒烟（需配置 embedder.base_url）

tests/test_sense_observer.py

@@ -0,0 +1,88 @@
+"""观察埋点测试（T-G2）：批量缓冲/刷盘/三消费方记录/升级回写。"""
+import asyncio
+import time
+
+import pytest
+
+from gateway.sense.observer import Observation, Observer, get_observer, reset_observer
+from gateway.sense.store import SenseStore
+
+
+def test_observer_batch_flush(tmp_path):
+    """log 入队 -> 100ms 内批量落盘；多条一次刷。"""
+    async def scenario():
+        store = SenseStore.init_db(tmp_path / "s.sqlite3")
+        reset_observer()
+        obs = Observer(store, flush_interval=0.05)
+        await obs.start()
+        for i in range(5):
+            obs.log(Observation(
+                request_id=f"r{i}", consumer="pipeline",
+                decided_tier="T2", executed_tier="T2",
+                probs={"t1": 0.2, "t2": 0.6, "t3": 0.2},
+                policy_version="v0-rule", features={"turns": 1}))
+        await asyncio.sleep(0.2)
+        await obs.stop()
+        return store
+
+    store = asyncio.run(scenario())
+    with store._lock, store._connect() as conn:
+        n = conn.execute("SELECT COUNT(*) AS c FROM tier_observations").fetchone()["c"]
+    assert n == 5
+
+
+def test_observer_embedding_blob(tmp_path):
+    """int8 BLOB 原样持久化（D-G6）。"""
+    async def scenario():
+        store = SenseStore.init_db(tmp_path / "s.sqlite3")
+        reset_observer()
+        obs = Observer(store, flush_interval=0.05)
+        await obs.start()
+        obs.log(Observation(request_id="rb", consumer="proxy",
+                            decided_tier="T1", executed_tier="T1",
+                            embedding=bytes([1, 127, 200]), policy_version="v0"))
+        await asyncio.sleep(0.2)
+        await obs.stop()
+        return store
+
+    store = asyncio.run(scenario())
+    with store._lock, store._connect() as conn:
+        row = conn.execute(
+            "SELECT embedding FROM tier_observations WHERE request_id='rb'").fetchone()
+    assert bytes(row["embedding"]) == bytes([1, 127, 200])
+
+
+def test_get_observer_singleton(tmp_path):
+    """同循环内 get_observer 返回单例；reset 后重建。"""
+    async def scenario():
+        store = SenseStore.init_db(tmp_path / "s.sqlite3")
+        reset_observer()
+        o1 = get_observer(store)
+        o2 = get_observer(store)
+        return o1 is o2
+    assert asyncio.run(scenario()) is True
+
+
+def test_upgrade_outcome_writeback(tmp_path):
+    """升级阶梯回写：outcome=escalated + executed_tier 更新。"""
+    store = SenseStore.init_db(tmp_path / "s.sqlite3")
+    store.insert_observation({
+        "request_id": "up1", "consumer": "pipeline", "decided_tier": "T1",
+        "executed_tier": "T1", "probs": "{}", "policy_version": "v0",
+        "features": "{}"})
+    assert store.update_outcome("up1", "escalated", executed_tier="T2") is True
+    with store._lock, store._connect() as conn:
+        row = conn.execute(
+            "SELECT outcome, executed_tier FROM tier_observations"
+            " WHERE request_id='up1'").fetchone()
+    assert row["outcome"] == "escalated" and row["executed_tier"] == "T2"
+
+
+def test_no_query_text_persisted(tmp_path):
+    """D-G6：观察表不含 query 原文（列结构断言）。"""
+    store = SenseStore.init_db(tmp_path / "s.sqlite3")
+    with store._lock, store._connect() as conn:
+        cols = [r["name"] for r in conn.execute(
+            "PRAGMA table_info(tier_observations)").fetchall()]
+    assert "query" not in cols and "query_text" not in cols
+    assert "embedding" in cols and "features" in cols