Make BlockSchema.cached_jsonschema default to None and add clear_schema_cache and clear_all_schema_caches (recursive) so JSON schemas can be invalidated and regenerated. Update modules (rest_api, admin llm_routes, executor llm_registry_init) to import BlockSchema from backend.blocks._base so the new cache-clearing API is used when refreshing LLM costs/discriminator mappings. Also switch cache sentinel from {} to None to avoid truthiness preventing regeneration.