fix(eval): parse_anomaly_segments must tolerate non-numeric junk entries

Full-scale eval crashed in the pure-LLM QA-judge path: pure_llm emits
free-form JSON whose 'segments' list can contain entries like
["CPU Usage", 30] (channel name instead of an int). _seg_from_entry called
int() unconditionally and raised ValueError, killing the whole report run
(model predictor had already finished; result lost).

Fix: wrap _seg_from_entry in try/except (ValueError, TypeError) → return
None for any unparseable entry, so good segments alongside junk are still
extracted. Verified: {["CPU Usage",30],[10,20]} → [(10,20)]. + unit test.
Also archive logs/train_full_run.log (stage1 31250 EMA1.26, stage2 9375 EMA0.75).

src/tsmm/eval/parse_answer.py

@@ -79,17 +79,22 @@ def _seg_from_entry(s):
 
     Accepts either ``[start, end]`` lists/tuples or ``{"start": s, "end": e}``
     objects (the schema emitted by ``data/instruct.py``). Returns None if no
-    usable pair is found.
+    usable pair is found, or if the values aren't ints (e.g. a pure-LLM emits
+    ``["CPU Usage", 30]`` — junk that must not crash the eval pipeline).
     """
-    if isinstance(s, dict):
-        if "start" in s and "end" in s:
-            return (int(s["start"]), int(s["end"]))
+    try:
+        if isinstance(s, dict):
+            if "start" in s and "end" in s:
+                return (int(s["start"]), int(s["end"]))
+            return None
+        if isinstance(s, (list, tuple)):
+            if len(s) >= 2:
+                return _normalize_seg(s)
+            if len(s) == 1:
+                v = int(s[0]); return (v, v)
+    except (ValueError, TypeError):
         return None
-    if isinstance(s, (list, tuple)):
-        if len(s) >= 2:
-            return _normalize_seg(s)
-        if len(s) == 1:
-            v = int(s[0]); return (v, v)
+    return None
     return None
 
 

tests/test_parse_answer.py

@@ -73,6 +73,18 @@ class TestParseSegments:
         assert any('"a"' in b and '"b"' in b for b in blobs)
         assert '{"d": 4}' in blobs
 
+    def test_non_numeric_segment_entries_are_skipped(self):
+        # a pure-LLM may emit junk like ["CPU Usage", 30] — must not crash, just skip
+        from tsmm.eval.parse_answer import _seg_from_entry
+        assert _seg_from_entry(["CPU Usage", 30]) is None
+        assert _seg_from_entry({"start": "foo", "end": "bar"}) is None
+        assert _seg_from_entry("not a segment") is None
+        # good entries alongside junk are still extracted
+        text = '{"segments": [["CPU Usage", 30], [10, 20]]}'
+        segs, ok = parse_anomaly_segments(text)
+        assert ok is True
+        assert segs == [(10, 20)]
+
     def test_empty_segments_list(self):
         text = '{"segments": [], "type": "none"}'
         segs, ok = parse_anomaly_segments(text)