fix(summarization): tag summary LLM calls nostream to stop phantom stream messages (#2503)

[18062706139fcz]

Problem

Fixes #2503. The SummarizationMiddleware runs its summary LLM call inside a before_model hook. Without a nostream tag, the summary tokens are captured by LangGraph's messages-tuple stream callback and broadcast to the frontend as a phantom AI message.

Fix

• Build a dedicated summary model copy tagged with "nostream" so the LangGraph streaming handler skips it (TAG_NOSTREAM).
• The tag is merged on top of any existing tags (e.g. middleware:summarize) instead of overwriting them, so RunJournal / tracing attribution is preserved (RunnableBinding.with_config does a shallow merge).
• Override _create_summary / _acreate_summary to invoke the tagged model directly, rather than temporarily swapping the shared self.model. The previous swap approach would leak the RunnableBinding across concurrent runs and break parent logic that inspects the raw model (profile / _get_ls_params).

Tests

Added regression tests covering:

• nostream tagging on the summary model
• concurrent-run isolation (shared self.model is never mutated mid-await)
• raw model preservation for parent profile inspection
• existing-tag merge (middleware:summarize + nostream, no duplicates)

All tests pass: pytest tests/test_summarization_middleware.py → 28 passed.