Extracts tool calls from the LLM response.

If the response contains tool calls (non-empty list), populates ctx.tool_calls for downstream stages.

If the response contains no tool calls and non-blank content, sets decision to {:done, response} — the LLM has produced a final text response and no further iteration is needed.

If the response contains no tool calls AND blank content (nil or whitespace-only), sets decision to {:blank_stop, response} — a degenerate completion that must not be accepted as final. Some OpenAI-compatible backends (observed on Ollama gpt-oss) occasionally return HTTP success with finish_reason: "stop", empty content, and no tool calls while a tool-driven task is still incomplete. The outer loop (LlmCore.Agent.Loop) owns the bounded recovery policy for this decision: it retries the turn and terminates with a typed error when blanks repeat.

When ctx.terminal_tool is set and a matching call is present, the pipeline marks the response as done and stores the matching call and raw arguments on the context. The call is not validated, dispatched, or injected into the next turn. A terminal-tool completion with blank text content is still :done — the payload travels in the tool arguments.

Analogous to a context merge stage: takes raw input and normalizes it into the pipeline's working format.