runLlmCallPipeline hardcoded a 24h TTL on every cache write; the host's
namespaced cache treats a per-call TTL as an override, so a DAG's
configured cacheTtlMs never applied to LLM cache entries (observed live:
a 7-day cacheTtlMs produced 1-day Redis expiries). The pipeline now
omits the TTL — expiry policy belongs to the adapter, which the host
wires from per-DAG config / host defaults. Contract documented on
ContextCacheAdapter.set; regression test pins the no-TTL call.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>