feat(llm): add a configurable output-token cap

- some model/gateway combinations emit unbounded reasoning/output and hang or
  trip an idle timeout (e.g. some deepseek-v4-flash deployments)
- add an opt-in max_tokens config knob + TRADINGAGENTS_MAX_TOKENS, forwarded to
  every provider when set (Gemini takes it as max_output_tokens); int-coerced,
  rejects non-positive/boolean values #1204

.env.example

@@ -61,6 +61,10 @@ NVIDIA_API_KEY=
 # own default (usually 2). Raise it to ride out bursty 429 rate-limit throttling
 # on rate-limited deployments (e.g. Azure OpenAI) instead of aborting the run.
 #TRADINGAGENTS_LLM_MAX_RETRIES=6
+# Cap on output tokens forwarded to every provider (Gemini's max_output_tokens
+# too). Unset leaves each provider at its default. Set it to bound a model that
+# emits unbounded reasoning/output and hangs or trips a gateway idle timeout.
+#TRADINGAGENTS_MAX_TOKENS=8192
 # Provider-specific reasoning/thinking depth (optional; unset = provider
 # default). Setting one also skips the matching interactive prompt.
 #TRADINGAGENTS_OPENAI_REASONING_EFFORT=medium