V2: Reconcile provider catalogs with /models endpoints

Problem

V2 should reconcile each provider's model catalog with the provider's /models endpoint instead of relying only on a static or stale catalog.

Today this can fail in both directions:

• models exposed by the provider may be missing from OpenCode
• models no longer exposed or usable by the provider may remain selectable

That leaves new models unavailable and stale or irrelevant models visible until hardcoded metadata catches up.

Desired behavior

• Fetch model availability from supported providers' /models endpoints.
• Merge endpoint results with OpenCode metadata needed for capabilities, limits, variants, and display names.
• Drop models that the authoritative endpoint does not expose or marks unavailable for the current account or organization.
• Preserve explicitly configured custom models where appropriate.
• Define fallback behavior when discovery is unsupported, unauthenticated, malformed, or temporarily unavailable.
• Refresh without making startup or model selection depend indefinitely on a slow provider.

Acceptance criteria

• A newly exposed provider model becomes selectable without a static catalog update when the endpoint supplies enough information.
• A model removed from the endpoint is no longer offered unless explicitly configured or covered by a documented fallback.
• Account- or organization-scoped model lists are respected.
• Endpoint failures retain a usable bounded fallback and surface diagnostics without silently replacing a known-good catalog with an empty list.
• Tests cover additions, removals, explicit overrides, scoped availability, malformed responses, and request failure/timeout.

Related scope

This is broader than generic OpenAI-compatible custom-provider discovery: it covers V2's model-catalog reconciliation and filtering semantics across providers that expose authoritative model-list endpoints.