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/default_config.py

@@ -18,6 +18,7 @@ _ENV_OVERRIDES = {
     "TRADINGAGENTS_CHECKPOINT_ENABLED":   "checkpoint_enabled",
     "TRADINGAGENTS_BENCHMARK_TICKER":     "benchmark_ticker",
     "TRADINGAGENTS_TEMPERATURE":          "temperature",
+    "TRADINGAGENTS_LLM_MAX_RETRIES":      "llm_max_retries",
     # Provider-specific reasoning/thinking knobs (None = each provider's own
     # default). Settable here for non-interactive runs; the CLI also offers an
     # interactive choice, which is skipped when the matching var is set.
@@ -95,6 +96,10 @@ DEFAULT_CONFIG = _apply_env_overrides({
     # variation on models that honor it; reasoning models largely ignore it
     # and no setting makes LLM output bit-identical across runs (see README).
     "temperature": None,
+    # SDK retry budget forwarded to every provider chat client. None leaves each
+    # provider/SDK at its own default (usually 2). Raise it to ride out bursty
+    # 429 throttling on rate-limited deployments instead of aborting a run (#1091).
+    "llm_max_retries": None,
     # Checkpoint/resume: when True, LangGraph saves state after each node
     # so a crashed run can resume from the last successful step.
     "checkpoint_enabled": False,