Closes #372.

Why

state.session.model is set once in initialState() and never mutated. /model <id>, /preset {auto,flash,pro} only route through loop.configure({ model }) — they never feed the agent store. So the next streaming.start / reasoning.start / "thinking · X" card snapshots the launch-time model id, regardless of which model actually answered the turn.

Repro:

1. Start session on flash (or default).
2. /model deepseek-v4-pro.
3. Send a new message.
4. Streaming card pill still reads flash.

What changes

• src/cli/ui/state/events.ts — new typed event:

  const sessionModelChange = z.object({
    type: z.literal("session.model.change"),
    model: z.string().min(1),
  });

• src/cli/ui/state/reducer.ts — handles session.model.change. Returns the same state when the id is unchanged so React.memo'd consumers don't re-render needlessly.

• src/cli/ui/slash/handlers/model.ts — /model and /preset {auto,flash,pro} dispatch the event after loop.configure(...). ctx.dispatch is already plumbed via agentStore.dispatch in App.tsx (existing wire — used by other handlers).

Cards already opened keep their captured model — that snapshot at streaming.start / reasoning.start is intentional so mid-turn auto-escalation doesn't relabel completed reasoning. Fresh cards opened after the switch render the new id.

Tests

tests/ui-reducer.test.ts — 3 new cases:

• session.model.change updates state.session.model and the next streaming card snapshots it
• no-op (referential identity) when id unchanged
• existing card opened on the prior model is NOT relabeled (preserves the existing snapshot guarantee)

Test plan

• npm run verify — 133 files / 2141 passed / 1 skipped (was 2138, +3 new)
• Manual: chat /model deepseek-v4-pro → send a message → next assistant card pill reads pro
• Manual: /preset flash → /preset pro → next card pill matches the active preset
• Manual: pre-existing assistant card from before the switch still shows the model that produced it (no retroactive relabel)