Lifecycle: distinguish model steps from user turns - #183

Review: approved ✅

Small (+31/-17) but it fixes a real semantic conflation: there was no way for a client to tell "the model finished one round-trip" from "the user's turn is over". Anything driving UI state off the old signal was guessing.

What I checked

• turn_complete is emitted exactly once per runAgent terminal state, with all five exits (abort, provider error, end-turn, approval abort, max-turns) funnelled through one completion path. That single-exit discipline is what makes the once-only guarantee actually hold rather than being a comment — worth preserving in future edits to the loop.
• Carrying the terminal stop reason plus the optional final assistant message on the event means clients don't have to reconstruct why a turn ended from event archaeology.
• CLI renderers stay intentionally silent for both events. Correct — these are bookkeeping for programmatic clients, and the terminal already renders completion its own way.

This is the event contract the app-server (#188) and every thin client downstream depend on, so landing it as its own reviewable slice ahead of them is the right ordering.

Validation: CI green; full suite green locally at the stack tip.