All scheduled tasks for a user shared one scheduler thread, so each run
appended its messages to the same LangGraph checkpoint. Over time the
history blew past the model's input window and every task on that
account started failing with context-too-long.
_execute_task now mints a fresh thread_id per run and threads it through
ensure_thread_meta / snapshot / start_run / extract_*. The per-task
scheduler_thread_id column is left in place for backwards compatibility
but no longer drives the LLM call.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
