fix(desktop): structured Fallback Models editor + out-of-catalog model visibility

[MarkVLK]

Problem

In Settings → Model, two things were confusing or broken:

1. Fallback Models rendered [object Object], [object Object]. fallback_providers is a list of {provider, model} objects, but it was shown through the generic list config field, which does value.join(', ') and stringifies each object.
2. The model box was blank whenever the configured model wasn't in the selected provider's advertised catalog — e.g. a Nous‑proxied openai/* id, an LM Studio / custom‑endpoint model, or any overridden id. Radix <Select> renders nothing for a value with no matching item, so the page hid what was actually configured.

Both surfaced from a real case where a primary model was misconfigured (a provider/model mismatch that 404'd and silently fell back) and the UI gave no usable signal about what was set or why.

Changes

• New FallbackModelsField — a structured provider + model row editor (add / remove) for fallback_providers, sourced from the same getGlobalModelOptions() the composer model picker uses. ConfigField renders it for that key instead of the generic list input. Half‑filled rows are kept in local state so the config autosave never persists a partial {provider, model: ''}, and an out‑of‑catalog model stays selectable so existing custom entries still render.
• Out‑of‑catalog model visibility — the main (and auxiliary) model <Select> now includes the current value as a selectable item even when it isn't in the provider's catalog, and shows a ⚠ note when the configured model isn't in the catalog, so the user knows calls may fall back.
• i18n — settings.model.{fallbackAdd, fallbackEmpty, notInCatalog} (en + zh; ja / zh‑hant fall back to English via defineLocale).
• Tests — fallback-models-field.test.tsx: each entry renders as its own row (never [object Object]), removing a row emits the remaining entries, adding a blank row never persists a partial pair, and the empty‑state hint.

Old

New

Verification

• npm run typecheck and eslint clean on the changed files (the two remaining react-hooks/exhaustive-deps warnings in config-settings.tsx are pre‑existing in untouched useEffects).
• vitest — the 4 new fallback-models-field tests pass. model-settings.test.tsx is unchanged by this PR; its one failing case fails identically on main (pre‑existing, unrelated).

Follow-up (not in this PR)

A "running on a fallback" banner: when the default model is failing and the agent is silently serving from the fallback_providers chain, surface it (e.g. "Default nous/… unavailable — running copilot/…"). Doing this correctly needs a backend fallback_activated signal on session.info — a pure effective‑vs‑configured comparison would false‑positive on intentional per‑session model switches — plus making the agent's self‑report name the effective model. Tracked separately.

🤖 Generated with Claude Code

[teknium1]

Salvaged the remaining structured fallback-editor scope into a current-main PR. The out-of-catalog visibility half was already implemented independently, so the salvage keeps only the object-list editor, tests, and profile/config reload hardening while preserving your authorship.