Manage Manual AI Models gains a leaf toggle to designate one non-default
manual model as the "economy" model, used for cheap background judgements
(command-safety checks, screen-idle detection) in place of the main chat
model. The designation is stored in kPreferenceKeyAIEconomyModelName and
resolved by ScreenWatchPoller, which prefers it (used verbatim, with its
own url/api/key) over the catalog's automatic same-vendor economy pointer.
The toggle is disabled for the default row since a model cannot be both.

Manual models now win over a built-in of the same name everywhere a model
name is resolved (request routing, capability gating, provider binding,
and the pickers), so a user can proxy a known model through a custom URL
by reusing its name.

Because a manual config could not express several per-model catalog
fields, cloning a preset silently dropped them. Fix that class of bug:

- Picking a preset now fills in the model name too, not just the other
  fields.
- Add "Supports temperature" and "Configurable thinking" checkboxes,
  seeded from the preset. A stored value wins; when absent the value is
  inherited from the same-named built-in, so existing clones self-heal
  without being re-created. Same-named clones also inherit the built-in's
  reasoning-effort and service-tier options, which have no editor UI.
- Honor supportsTemperature in the OpenAI and DeepSeek completion request
  builders, not just Anthropic. This is a no-op for every shipping model
  (all are true on those paths) but stops a temperature-rejecting manual
  or future model from sending a field the API 400s on.

Also swap the economy-row table marker from a colored emoji to a
monochrome leaf SF Symbol.