fix: Kelly per-verdict caps, more noise words, higher token limit, Pydantic warning

- Cap Kelly per verdict: STRONG BUY 25%, BUY 15%, WATCH 5% (was flat 25%)
- Cap raw Kelly fraction at 0.5 to prevent extreme values from small avg_loss
- Add THING, TXTW, MRC, HERE and 20+ more common words to noise filter
- Increase structured output max_tokens from 8192 to 16384
- Fix Pydantic warning filter to match by module instead of message string

Made-with: Cursor

src/agent.py

@@ -375,7 +375,7 @@ def analyst_node(state):
 
         structured_llm = get_structured_llm().with_structured_output(InvestmentVerdict)
         with warnings.catch_warnings():
-            warnings.filterwarnings("ignore", message="Pydantic serializer warnings")
+            warnings.filterwarnings("ignore", category=UserWarning, module="pydantic")
             result = structured_llm.invoke(prompt)
 
         stats = get_kelly_stats()

src/core/ticker_utils.py

@@ -42,6 +42,10 @@ NOISE_WORDS = frozenset({
     "SPAC", "NBER", "OPEC", "MSCI", "EMEA", "APAC", "OECD", "FIFO",
     "FINRA", "SIPC", "FDIC", "LISA", "ISA", "ATM", "AMA", "FDA",
     "PHNX", "IPG", "GAAP", "IFRS", "FASB", "IASB", "PCAOB",
+    "THING", "TXTW", "MRC", "HERE", "ELSE", "SURE", "WORK",
+    "SAFE", "IDEA", "PLAN", "RULE", "STEP", "PLAY", "OPEN",
+    "PART", "NOTE", "LINE", "READ", "FILL", "SIZE", "WIDE",
+    "SIGN", "RISE", "LEAD", "PUSH", "PULL", "DROP", "JUMP",
 })
 
 _MAX_TICKER_LEN = 8  # longest valid ticker with suffix: e.g. CHE.UN.TO

src/llm.py

@@ -48,7 +48,7 @@ def get_llm() -> ChatOpenAI:
     return _llm_instance
 
 
-def get_structured_llm(max_tokens: int = 8192) -> ChatOpenAI:
+def get_structured_llm(max_tokens: int = 16384) -> ChatOpenAI:
     """Return an LLM instance configured for structured output.
 
     Uses a capped ``max_tokens`` to prevent reasoning models from

src/models/kelly.py

@@ -34,6 +34,12 @@ _VERDICT_SCALE = {
 }
 _POS_FLOOR = 1.0
 _POS_CAP = 25.0
+_VERDICT_CAPS = {
+    "STRONG BUY": 25.0,
+    "BUY": 15.0,
+    "WATCH": 5.0,
+}
+_MAX_KELLY_FRACTION = 0.5  # cap raw Kelly to avoid extreme values
 
 
 @dataclass
@@ -160,7 +166,7 @@ def get_kelly_stats() -> KellyStats:
     else:
         kelly = 0.0
 
-    kelly = max(kelly, 0.0)
+    kelly = max(min(kelly, _MAX_KELLY_FRACTION), 0.0)
 
     result = KellyStats(
         total_trades=total,
@@ -201,5 +207,6 @@ def calculate_position_size(stats: KellyStats, verdict: str) -> float:
         return 0.0
 
     raw = stats.half_kelly * scale * 100
+    cap = _VERDICT_CAPS.get(verdict, _POS_CAP)
 
-    return round(max(_POS_FLOOR, min(raw, _POS_CAP)), 1)
+    return round(max(_POS_FLOOR, min(raw, cap)), 1)

src/whale_hunter.py

@@ -313,7 +313,7 @@ def analyst_node(state):
 
         structured_llm = get_structured_llm().with_structured_output(InvestmentVerdict)
         with warnings.catch_warnings():
-            warnings.filterwarnings("ignore", message="Pydantic serializer warnings")
+            warnings.filterwarnings("ignore", category=UserWarning, module="pydantic")
             result = structured_llm.invoke(prompt)
 
         stats = get_kelly_stats()