fix: reactive coverage (halt notices, "why is X up") no longer scores as a catalyst

qwen3:14b rated "halted on circuit breaker to the upside, stock now up 172%" at
conviction 0.8 / sentiment +1.0, a bigger rank boost than most real news. So
names that had already run were ranked up.
- Prompt: halt notices, movers roundups and articles that only report a price move
  are not a new catalyst (conviction <= 0.1). If one names a new announcement,
  classify it with conviction <= 0.3.
- A deterministic cap that does not depend on the model: halt type or halt wording
  caps conviction at 0.1; after-the-fact wording ("why is X stock up", "X shares
  surge", "stocks moving") caps it at 0.3. Dilution flags are never touched.
On the 4,046-headline news history the pattern catches 100% of halt notices and 82%
of roundups, 0% of analyst items and 0.8% of earnings items. Against qwen3:14b on a
fixed 20-headline set: reactive mean boost 0.318 -> 0.055 (max 0.80 -> 0.21); real
catalysts 0.353 -> 0.362; offerings 3/3 still flagged dilutive.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B3PXXn2tHBzdv6rYtGPz86

research/ingestion/news_enrichment.py

@@ -41,6 +41,20 @@ CATALYST_TYPES = {
     "none",
 }
 
+# Coverage that only REPORTS a move or a halt is not a new catalyst, but the model rated "halted to the upside,
+# stock up 172%" at conviction 0.8 / sentiment +1.0 (a bigger boost than most real news), which ranks up names
+# that already ran. The prompt says so; this deterministic cap does not depend on the model listening.
+HALT_RE = re.compile(r"\b(halted|trading halt|halt status|circuit breaker|luld|resum(?:e|es|ed|ption) (?:of )?trad)",
+                     re.I)
+REACTIVE_RE = re.compile(
+    r"\b(why|what'?s going on|what'?s driving|here'?s why)\b.{0,80}\b(stock|shares)\b"
+    r"|\b(stock|shares)\b.{0,40}\b(surg|soar|jump|spik|rall|climb|gain|ris|trend|tumbl|plung|slid|sink|drop|fall)\w*"
+    r"|\b(stock|shares)\b.{0,40}\btrading (?:higher|lower)"
+    r"|stocks? moving|\bmovers\b|(?:top|biggest) (?:gainers|losers)|unusual (?:options|volume)|52-week (?:high|low)",
+    re.I)
+HALT_MAX_CONVICTION = 0.1
+REACTIVE_MAX_CONVICTION = 0.3
+
 # Cheap stdlib floor for the dilution veto so it still works when Ollama is down.
 # The LLM CONFIRMS dilution; this regex is the safety net, not the primary signal.
 # "offering" is matched ONLY as a securities offering — a qualifier in front of it
@@ -129,7 +143,12 @@ def build_prompt(headline: str, snippet: str = "", tickers: str = "") -> str:
         'everything else, including an ETF/fund "offering exposure", a product or '
         "service offering, an IPO of a different company, a buyback, or a reverse "
         "split\n"
-        '  "rationale": one short sentence (max ~200 chars)\n\n'
+        '  "rationale": one short sentence (max ~200 chars)\n'
+        "A trading halt/resumption notice, a 'stocks moving'/movers roundup, or an article that only reports or "
+        "explains a price move ('Why is X stock up', 'X shares surge') is NOT a new catalyst: use catalyst_type "
+        '"none" (or "halt_resumption" for a halt) with conviction at most 0.1. If such an article names a NEW '
+        "company announcement as the cause, classify that announcement but with conviction at most 0.3, because "
+        "the move has already happened.\n\n"
         f"<news>\n{news}\n</news>"
     )
 
@@ -141,7 +160,7 @@ def _clamp(value, lo: float, hi: float, default: float) -> float:
         return default
 
 
-def _parse(raw: str, fallback_text: str = "") -> CatalystAnalysis | None:
+def _parse(raw: str, fallback_text: str = "", headline: str = "") -> CatalystAnalysis | None:
     """Tolerant parse of the model's JSON string. None on unrecoverable failure."""
     try:
         data = json.loads(raw)
@@ -162,10 +181,16 @@ def _parse(raw: str, fallback_text: str = "") -> CatalystAnalysis | None:
     if not is_dilutive and looks_dilutive(fallback_text):
         is_dilutive = True
     rationale = str(data.get("rationale", "") or "")[:200]
+    conviction = _clamp(data.get("conviction", 0.0), 0.0, 1.0, 0.0)
+    # cap reactive coverage (never touches is_dilutive: an offering is still an offering)
+    if ctype == "halt_resumption" or HALT_RE.search(headline or ""):
+        conviction = min(conviction, HALT_MAX_CONVICTION)
+    elif REACTIVE_RE.search(headline or ""):
+        conviction = min(conviction, REACTIVE_MAX_CONVICTION)
     return CatalystAnalysis(
         catalyst_type=ctype,
         sentiment=_clamp(data.get("sentiment", 0.0), -1.0, 1.0, 0.0),
-        conviction=_clamp(data.get("conviction", 0.0), 0.0, 1.0, 0.0),
+        conviction=conviction,
         is_dilutive=is_dilutive,
         rationale=rationale,
     )
@@ -213,7 +238,7 @@ def classify_headline(
         return None
     if not response:
         return None
-    return _parse(response, fallback_text=f"{headline} {snippet}")
+    return _parse(response, fallback_text=f"{headline} {snippet}", headline=headline)
 
 
 # ---------------------------------------------------------------------------