Fix gpt-oss harmony leak: drop truncated analysis channel + raise narrator budget

The Narrator's 160-token budget truncated gpt-oss mid-analysis-channel, so
harmony_final (which only stripped when 'assistantfinal' was present) returned
the raw chain-of-thought and it leaked into the public headline and the JSON
parser. Now harmony_final drops an unfinished analysis channel (returns ''),
the narrator gets 512 tokens + a 'Reasoning: low' hint so it reaches its final
channel, and the prompt no longer names a TV show. Reviewer APPROVED.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

tests/test_normalize.py

@@ -33,6 +33,22 @@ def test_harmony_without_marker_is_unchanged():
     assert harmony_final(text) == text
 
 
+def test_harmony_truncated_analysis_is_dropped():
+    """When the budget runs out before `assistantfinal`, the model is still in
+    its analysis channel; raw chain-of-thought must never leak downstream."""
+    text = "analysisWe need to produce a punchy headline, optionally a short qu"
+    assert harmony_final(text) == ""
+
+
+def test_harmony_empty_string_is_empty():
+    assert harmony_final("") == ""
+
+
+def test_harmony_final_channel_is_stripped():
+    """Leading/trailing whitespace around the final-channel answer is removed."""
+    assert harmony_final('analysis ...assistantfinal  {"x":1}  ') == '{"x":1}'
+
+
 def test_identity_returns_text_unchanged():
     text = "anything at all, even {broken json"
     assert identity(text) == text

ttw/narrator.py

@@ -125,7 +125,10 @@ class DummyNarrator:
 class LLMNarrator:
     """gpt-oss-backed Narrator. Reuses the council's gptoss ModalLLM (no add)."""
 
-    def __init__(self, client, max_tokens: int = 160):
+    def __init__(self, client, max_tokens: int = 512):
+        # gpt-oss reasons in an analysis channel before its final answer; a tight
+        # budget truncates before `assistantfinal`, so give it room to land the
+        # final channel (the normalizer drops anything that never gets there).
         self.client = client
         self.max_tokens = max_tokens
 
@@ -142,8 +145,9 @@ class LLMNarrator:
                 {
                     "role": "system",
                     "content": (
+                        "Reasoning: low\n"
                         "You are the Narrator of the Wood Street Journal, a "
-                        "Billions-style financial dramatist. One punchy headline, "
+                        "high-stakes financial dramatist. One punchy headline, "
                         "optionally one short character quip. No spoilers."
                     ),
                 },

ttw/normalize.py

@@ -14,11 +14,16 @@ from typing import Callable
 def harmony_final(text: str) -> str:
     """Return the harmony final channel: text after the first `assistantfinal` marker.
 
-    gpt-oss emits `...analysis...assistantfinal<answer>`. With no marker the text
-    is already bare, so return it unchanged.
+    gpt-oss emits `...analysis...assistantfinal<answer>`. When the generation
+    budget runs out before `assistantfinal` the model is still in its analysis
+    (chain-of-thought) channel; return "" so raw reasoning never leaks downstream
+    (into a headline or into the JSON parser). With no marker and no analysis
+    channel the text is already bare, so return it unchanged.
     """
     if "assistantfinal" in text:
-        return text.split("assistantfinal", 1)[1]
+        return text.split("assistantfinal", 1)[1].strip()
+    if text.lstrip().lower().startswith("analysis"):
+        return ""  # truncated reasoning, no final channel emitted -- drop it
     return text