Add custom-endpoint support to the Anthropic provider

[sairohiththammana]

Mirrors the OpenAI slot: an optional base_url on the anthropic profile points the Anthropic SDK at any Messages-API-compatible gateway (LiteLLM, a corporate proxy). Blank keeps api.anthropic.com — and keeps the SDK's own ANTHROPIC_BASE_URL env support intact.

Anthropic needs less than OpenAI did: there is one wire either way (no Responses ⇄ Chat Completions switch), so the endpoint only swaps the host.

What it does need

• The stored value is the ROOT. The SDK appends /v1/messages itself, so a /v1 carried over from the OpenAI box — where it belongs — is trimmed rather than doubled into /v1/v1/messages.
• Fable/Mythos skip the beta server-side-fallback path on a custom endpoint. The beta header, the fallbacks param and the beta messages route are Anthropic-only; a gateway that doesn't know them 400s the whole turn. Losing the fallback there costs a refusal retry, not a working request.
• Test falls back to a one-token /v1/messages call when a gateway proxies messages but not /v1/models, and treats an Anthropic-shaped 4xx as proof the endpoint speaks the API. "Anthropic-shaped" means the top-level "type": "error" discriminator — a bare nested error object is the OpenAI/LiteLLM shape, and accepting it would green-light an OpenAI URL pasted into this box: Test would pass and every turn would fail afterwards.
• resolve_api_key grows a prefer_profile flag, set when an endpoint is configured, so a key typed alongside a gateway is never shadowed by an ambient ANTHROPIC_API_KEY meant for api.anthropic.com. Belt-and-braces on the router path (the registry already passes the profile key explicitly); it earns its keep for provider instances built from a SecretStore alone.

GUI

No new component. The endpoint row is descriptor-driven — ProviderForm renders it for any keyed provider that declares base_url — so the field appears in both Onboarding and Settings ▸ Models as soon as the descriptor has it. One behavior change: a stored endpoint now opens the disclosure instead of hiding behind the collapsed "Custom endpoint ⌄" link, since a configured gateway is the one thing a collapsed row would actively misrepresent. That applies to OpenAI too.

Testing

• pytest tests — 1181 passed, 1 skipped.
• New: endpoint normalization (/v1 trimming, blank → None), base_url reaching the SDK only when set, key precedence under a gateway, the beta-path skip, and four verify cases (custom endpoint URL, /v1/messages fallback, key rejection on the fallback, non-Anthropic host still failing — parametrized over a plain 404 page and an OpenAI-shaped error body).
• GUI: two ProviderForm tests covering the descriptor-driven endpoint row. Not run locally (no Node on the authoring machine) — CI's gui-unit job covers them.

Known limitation

Gateways that require Authorization: Bearer instead of x-api-key are not supported; the SDK sends x-api-key. Adding an auth-style toggle would be a second field and a second verify branch — deliberately deferred until a real gateway needs it.