v3.1.2

Fixed

• AgenticHub's LLM intent-classification fallback (free-form questions with no matching keyword)
  could route to lineage/dictionary/quality/builder even when that agent isn't configured in
  the hub, and _dispatch treated it as a hard "Agent '<x>' non configure dans ce hub." error
  instead of falling back to GenBI — on a hub exposing only GenBI, a free-form metrics question like
  "l'évolution mensuelle du CA des commandes" could be misrouted and fail outright. Fixed by
  restricting _llm_classify_intent's candidate intents (and its system prompt's per-intent
  descriptions) to the agents actually configured on the hub, and by falling back to GenBI in
  _dispatch only when the intent came from that LLM fallback — an explicit keyword-detected
  request for an unconfigured agent still returns the "non configure" error as before.
• A failed Spark/Spark Connect query (e.g. an AnalysisException) leaked its full message —
  including the JVM stacktrace: marker and thousands of characters of at org.apache.spark.../
  at com.databricks... frames — into AgentResponse.error/QualityResponse.error, and from there
  into the Markdown rendered for Playground/the supervisor. Fixed by sanitizing the message at the
  point each agent builds its error response (new agentic/_errors.py::sanitize_exception_message),
  keeping only the readable prefix (error code, short message) up to the JVM stacktrace: marker (or
  the first stack-frame line when that marker is absent) and capping the result's length; the full
  original message is still available via exc_info=True logging. The Provenance block is unaffected
  — it reads only from evidence, never from error.