fix(graph): key checkpoints on graph shape and expose the LLM retry budget

- checkpoint resume keyed only by ticker+date silently continued the old graph
  under a different analyst selection / depth / asset mode; fold a run signature
  into the thread id #1089
- add llm_max_retries + TRADINGAGENTS_LLM_MAX_RETRIES, forwarded to every provider
  when set (int-coerced, rejects negatives/booleans), so a 429 burst can't abort
  a run #1091

tradingagents/graph/trading_graph.py

@@ -44,6 +44,24 @@ from .signal_processing import SignalProcessor
 logger = logging.getLogger(__name__)
 
 
+def _coerce_max_retries(value):
+    """Validate an ``llm_max_retries`` value to a non-negative int.
+
+    Accepts an int or a numeric string (env vars arrive as strings). Rejects
+    booleans and negatives loudly so a misconfiguration fails at startup rather
+    than silently disabling retries.
+    """
+    if isinstance(value, bool):
+        raise ValueError(f"llm_max_retries must be an integer, not a boolean: {value!r}")
+    try:
+        n = int(value)
+    except (TypeError, ValueError) as exc:
+        raise ValueError(f"llm_max_retries must be an integer, got {value!r}") from exc
+    if n < 0:
+        raise ValueError(f"llm_max_retries must be >= 0, got {n}")
+    return n
+
+
 class TradingAgentsGraph:
     """Main class that orchestrates the trading agents framework."""
 
@@ -124,6 +142,9 @@ class TradingAgentsGraph:
         self.ticker = None
         self.log_states_dict = {}  # date to full state dict
 
+        # Graph-shape-affecting run choices, kept for the checkpoint signature.
+        self.selected_analysts = tuple(selected_analysts)
+
         # Set up the graph: keep the workflow for recompilation with a checkpointer.
         self.workflow = self.graph_setup.setup_graph(selected_analysts)
         self.graph = self.workflow.compile()
@@ -156,6 +177,12 @@ class TradingAgentsGraph:
         if temperature is not None and temperature != "":
             kwargs["temperature"] = float(temperature)
 
+        # SDK retry budget is cross-provider. Forward it only when explicitly set
+        # so each provider keeps its own default (usually 2) otherwise (#1091).
+        max_retries = self.config.get("llm_max_retries")
+        if max_retries is not None and max_retries != "":
+            kwargs["max_retries"] = _coerce_max_retries(max_retries)
+
         return kwargs
 
     def _create_tool_nodes(self) -> dict[str, ToolNode]:
@@ -318,6 +345,20 @@ class TradingAgentsGraph:
         identity = resolve_instrument_identity(ticker)
         return build_instrument_context(ticker, asset_type, identity)
 
+    def _run_signature(self, asset_type: str) -> str:
+        """Graph-shape inputs that must invalidate a checkpoint if changed.
+
+        Keyed into the checkpoint thread ID so a resume under a different analyst
+        selection, debate/risk depth, or asset mode starts fresh instead of
+        silently continuing the previous graph (#1089).
+        """
+        return "|".join([
+            "analysts=" + ",".join(self.selected_analysts),
+            f"debate={self.config['max_debate_rounds']}",
+            f"risk={self.config['max_risk_discuss_rounds']}",
+            f"asset={asset_type}",
+        ])
+
     def propagate(self, company_name, trade_date, asset_type: str = "stock"):
         """Run the trading agents graph for a company on a specific date.
 
@@ -342,7 +383,8 @@ class TradingAgentsGraph:
             self.graph = self.workflow.compile(checkpointer=saver)
 
             step = checkpoint_step(
-                self.config["data_cache_dir"], company_name, str(trade_date)
+                self.config["data_cache_dir"], company_name, str(trade_date),
+                self._run_signature(asset_type),
             )
             if step is not None:
                 logger.info(
@@ -389,9 +431,10 @@ class TradingAgentsGraph:
         )
         args = self.propagator.get_graph_args()
 
-        # Inject thread_id so same ticker+date resumes, different date starts fresh.
+        # Inject thread_id so same ticker+date+graph-shape resumes; a different
+        # date or graph shape starts fresh (#1089).
         if self.config.get("checkpoint_enabled"):
-            tid = thread_id(company_name, str(trade_date))
+            tid = thread_id(company_name, str(trade_date), self._run_signature(asset_type))
             args.setdefault("config", {}).setdefault("configurable", {})["thread_id"] = tid
 
         if self.debug:
@@ -432,7 +475,8 @@ class TradingAgentsGraph:
         # Clear checkpoint on successful completion to avoid stale state.
         if self.config.get("checkpoint_enabled"):
             clear_checkpoint(
-                self.config["data_cache_dir"], company_name, str(trade_date)
+                self.config["data_cache_dir"], company_name, str(trade_date),
+                self._run_signature(asset_type),
             )
 
         return final_state, self.process_signal(final_state["final_trade_decision"])