Model-freshness: flag stale/retired models in cascades from live provider list

[juniperbevensee]

Task for the gardener. Model IDs change often (Anthropic retires older ones); the create-new wizard and existing-harness cascade editor can offer models that are retired → agent crash-loops on restart. Add live freshness checking. Anthropic-first, extensible.

Design (ready to build)

1. lib/services/model-freshness.ts — fetchLiveModels('anthropic'): find an Anthropic key via the keys service, call GET /v1/models (auth x-api-key + anthropic-version: 2023-06-01, no beta header; paginate), cache to disk with ~1h TTL. Fail-soft: no key / network error → return null, callers fall back to the static MODEL_CATALOG unflagged (never block the UI).
2. GET /api/models/live?provider=anthropic → { models, fetchedAt }.
3. Freshness signal = list membership. Retired models simply disappear from /v1/models (404 on retrieve) — there is no "deprecated" flag in the response. So: catalog entries absent from the live list → mark retired: true; live IDs not in MODEL_CATALOG → offer as suggestable/new.
4. Consume in both the create-new wizard model step and the existing-harness cascade editor (app/api/harnesses/[id]/models) — show a "retired — pick a current model" warning on stale entries.
5. Provider-keyed so openai/google can be added later; wire Anthropic only now.

Notes / guardrails

• Verify model IDs against the live list, never hardcode from training data.
• The models endpoint object carries id, display_name, created_at, and (since Mar 2026) max_input_tokens/max_tokens/capabilities.
• RGTDD + independent audit before merge, per the repo's dev flow.

Context: raised 2026-07-02; deferred here from a live session so it lands as durable, gardener-pickup-able work rather than dying in chat.