From c68de849dabad5eddd3fcd6b7ccd73801d88c1a9 Mon Sep 17 00:00:00 2001
From: iomgaa <iomgaaycz@gmail.com>
Date: Thu, 16 Jul 2026 05:38:05 -0400
Subject: [PATCH] fix: slow-update R2 dual_metric uses slow_candidate kind (not
 final)

---
 app/harness/runner.py                  |  4 +++-
 tests/unit/test_harness_observation.py | 31 ++++++++++++++++++++++++++
 2 files changed, 34 insertions(+), 1 deletion(-)

diff --git a/app/harness/runner.py b/app/harness/runner.py
index f468e12..068601d 100644
--- a/app/harness/runner.py
+++ b/app/harness/runner.py
@@ -1443,11 +1443,13 @@ class Runner:
             r2_skills_version = self._current_version("skills")
             new_prompts_version = self._current_version("prompts")
             eval_r2 = await self._eval_full_val(epoch, pools, run_suffix="_p2")
+            # R2 是可能被 revert 的慢更新候选，用 slow_candidate 口径，不占 final
+            # （epoch 终值唯一由 Phase 2 的 final 承载）
             write_dual_metric(
                 str(self._paths.db_path),
                 run_id=self._config.run_id,
                 epoch=epoch,
-                version_kind="final",
+                version_kind="slow_candidate",
                 skills_version=r2_skills_version,
                 prompts_version=new_prompts_version,
                 pool="val",
diff --git a/tests/unit/test_harness_observation.py b/tests/unit/test_harness_observation.py
index d37cc40..c6b7fd2 100644
--- a/tests/unit/test_harness_observation.py
+++ b/tests/unit/test_harness_observation.py
@@ -71,6 +71,37 @@ def test_write_read_dual_metric(db_path: str, run_id: str) -> None:
     assert row["run_id"] == run_id
 
 
+def test_dual_metric_kinds_distinguishable(db_path: str, run_id: str) -> None:
+    """observation 层接受任意 version_kind：final 与 slow_candidate 可区分共存。"""
+    write_dual_metric(
+        db_path,
+        run_id=run_id,
+        epoch=1,
+        version_kind="final",
+        skills_version="v1",
+        prompts_version="v1",
+        pool="val",
+        hard_acc=0.7,
+        soft_score=None,
+        mixed_score=None,
+    )
+    write_dual_metric(
+        db_path,
+        run_id=run_id,
+        epoch=1,
+        version_kind="slow_candidate",
+        skills_version="v2",
+        prompts_version="v2",
+        pool="val",
+        hard_acc=0.6,
+        soft_score=None,
+        mixed_score=None,
+    )
+    rows = read_dual_metric(db_path, run_id=run_id)
+    kinds = {row["version_kind"] for row in rows if row["epoch"] == 1}
+    assert kinds == {"final", "slow_candidate"}
+
+
 # ---------------------------------------------------------------------------
 # shadow_gate
 # ---------------------------------------------------------------------------