Pick the right LangGraph stream_mode for showing tokens vs node updates

**LangGraph** exposes streaming through a `stream_mode` argument on the graph's `.stream` (or `.astream`) method. Three modes carry the load in production. - **messages** streams LLM tokens as they arrive from inside any node. This is what you wire to a UI for typewriter-style output. It also includes tool-call deltas, which is how you show 'the model is calling tool X' before the tool finishes. - **updates** emits the partial state delta each node returned, keyed by node name. Use it when the UI shows 'planner finished, now running tool'. Lightweight, per node, no token noise. - **values** emits the FULL state object after each node runs. Heavier than updates because it ships the entire state every time, useful for debugging or for clients that rebuild UI from scratch on each event. The other options are wrong. There is no `tokens` mode and no `nodes` mode (those names sound right but do not exist). `stream=True` is just the verb for streaming, not a mode picker. And `values` does not carry tokens, it carries state. Picking the wrong mode either floods your UI with deltas or starves it of token-level updates.

**Three real modes, three granularities.** `messages` is per-token (LLM and tool-call deltas streamed from inside any node). `updates` is per-node delta (only what that node added to state). `values` is per-node full state. The right pick depends on the UI. A typewriter-style chat surface needs `messages` because it has to show tokens as they arrive. A status bar that shows 'planner ran, retrieval ran, synthesis running' needs `updates` because it only needs to know when each node finished and what it added. A debugging dashboard or a client that does not maintain its own copy of state can use `values` to get the full snapshot per step, at the cost of bandwidth. The distractors are pure noise. There is no tokens mode. The LLM-token stream is named messages, after what an LLM call actually emits. There is no nodes mode. Node-level events are updates or values. `stream=True` is the verb that turns streaming on; it does not choose a mode. And values does not carry tokens. Naming the three real modes and what each is for is the whole signal this question is testing.