feat: Gemini Flash judge harness + agreement subsets

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

tests/test_judge.py

@@ -0,0 +1,49 @@
+from buddhagpt.judge import parse_score, judge_messages
+
+
+def test_parse_score_reads_json_block():
+    txt = 'Reasoning...\n{"empathy": 6, "nonharm": 7, "honesty": 5, "groundedness": 4, "overall": 5}'
+    s = parse_score(txt)
+    assert s["overall"] == 5
+
+
+def test_parse_score_returns_none_on_garbage():
+    assert parse_score("no json here") is None
+
+
+def test_parse_score_returns_none_on_missing_key():
+    txt = '{"empathy": 6, "nonharm": 7, "honesty": 5, "overall": 5}'
+    assert parse_score(txt) is None
+
+
+def test_parse_score_returns_none_on_non_integer():
+    txt = '{"empathy": "high", "nonharm": 7, "honesty": 5, "groundedness": 4, "overall": 5}'
+    assert parse_score(txt) is None
+
+
+def test_parse_score_picks_json_after_reasoning_prose():
+    txt = (
+        "The response shows good empathy but invents a citation.\n"
+        "Final scores:\n"
+        '{"empathy": 5, "nonharm": 6, "honesty": 3, "groundedness": 1, "overall": 3}'
+    )
+    s = parse_score(txt)
+    assert s == {"empathy": 5, "nonharm": 6, "honesty": 3, "groundedness": 1, "overall": 3}
+
+
+def test_parse_score_tolerates_whitespace_in_keys():
+    # google/gemini-flash-latest recurringly emits a stray leading space on one
+    # key (e.g. `" honesty"` instead of `"honesty"`) for certain long inputs;
+    # the JSON is otherwise well-formed, so we shouldn't reject it.
+    txt = '```json\n{\n  "empathy": 7,\n  "nonharm": 7,\n  " honesty": 7,\n  "groundedness": 7,\n  "overall": 7\n}\n```'
+    s = parse_score(txt)
+    assert s == {"empathy": 7, "nonharm": 7, "honesty": 7, "groundedness": 7, "overall": 7}
+
+
+def test_judge_messages_includes_rubric_prompt_and_response():
+    msgs = judge_messages("prompt text", "response text", "rubric text")
+    assert msgs[0]["role"] == "system"
+    assert "rubric text" in msgs[0]["content"]
+    assert msgs[1]["role"] == "user"
+    assert "prompt text" in msgs[1]["content"]
+    assert "response text" in msgs[1]["content"]