cached_provider_model_ids and cached_fetch_api_models loaded the cache,
fetched live, then saved that pre-fetch snapshot unlocked, so a slower
prefetch worker, SWR thread or other process dropped rows written
meanwhile and the picker fell back to curated lists. _store_cache_entry
now always re-reads under the thread lock plus a cross-process flock.
Drop the prefetch worker's compensating update_provider_cache_entry
re-persist: it re-stamped rows the canonical path deliberately leaves
untouched (stale catalog served on a failed fetch looked fresh for a TTL),
stripped the curated-fallback flag, and keyed ollama as "custom".
