feat: mid-session model switching per thread, with TUI + desktop UI and e2e tests

[shadaj]

Implements ClientMessage::SwitchModel (previously a stub) so the model can be
changed mid-session. The switch is scoped to a single thread — no propagation
to children — and never disturbs an in-flight completion.

Daemon core

• SessionManager::switch_model: validates the model against the catalog
  (immediate Error reply on unknown model/thread), persists the selection on
  the thread via set_thread_model, then forwards it to a live worker through
  the agent loop; if nothing is running, the persisted selection is resolved
  at the next worker boot. Returns a DaemonMessage::ModelSwitched
  confirmation for the requesting client.
• agent_loop: new AgentMessage::SwitchModel routed to a live worker's
  dedicated model_switch_tx channel; dropped (no worker spawn) when the
  thread isn't running since the selection is already persisted.
• thread_worker: the current model lives in a slot
  (Option<(ModelRef, Arc<dyn ModelProvider>)>). Each completion round takes
  the model out of the slot and the round future owns it (async move
  wrapper around process_batch + inner cancel forwarding via an outer
  oneshot), returning it when the round ends. Applying a switch fills the
  slot, so a mid-round switch simply drops the old model when the round
  finishes — in-flight completions always finish on the model they started
  with, and the next round uses the latest selection (switch channel is also
  drained right before each round starts). Switches are handled in biased
  select! arms both while a completion is active and while idle-waiting,
  broadcasting ModelSwitched to the thread's subscribers on apply.
  context_window updates with the switch so auto-compaction thresholds
  follow the active model. process_batch itself is untouched.
• Protocol: DaemonMessage::ModelSwitched { thread_id, model_name, context_window, provider_id } (+ remote-prefix handling); documented
  SwitchModel semantics.
• InMemoryConversationStore::has_thread for validating switch targets.
• Unit tests: switch while waiting for an async tool result applies to the
  next completion; switch during an in-flight completion leaves it
  undisturbed and applies to the next round.

TUI

• DaemonMessage::ModelSwitched updates the status-line pill
  (model/provider/context window) via a new terminal::ModelSwitched channel
  when it targets the viewed thread (the /model picker was already wired to
  send SwitchModel).
• New e2e tests (vt100 + real daemon + mock providers):
  switch_model_mid_session (picker → switch persists while idle → next
  worker boot uses the new provider) and switch_model_while_busy (switch
  mid-stream; old model finishes, follow-up hits the new one), with screen
  snapshots.

Desktop web UI

• The model pill (top right) now drops down on hover with the list of
  available models (provider, name, context size, active marker); clicking
  one sends SwitchModel for the currently viewed thread. The pill and a
  "Switched model to …" transcript line update from the ModelSwitched
  confirmation (no optimistic update).
• infinity-ui protocol types: added ModelRef and ModelSwitched, fixed
  the stale SwitchModel/CreateSession client types.
• New Playwright e2e switch_model_mid_session (two mock providers): hover
  dropdown, select, confirmation line, next request routed to the second
  provider; two new golden screenshots.

[cloudflare-workers-and-pages]

Deploying infinity with    Cloudflare Pages

Latest commit:	ebd2980
Status:	✅  Deploy successful!
Preview URL:	https://5481d7ab.infinity-dc7.pages.dev
Branch Preview URL:	https://sandbox-40cde97a-42ef-4dfd-a.infinity-dc7.pages.dev

View logs