* fix(coding-agent): keep refinement passes from evicting the session prompt cache Refinement review and planning run with their own prompt prefixes, so issuing them on the session model replaces the provider prefix-cache entry for the session; the next session request then re-reads its whole context. Add an auxiliaryModel setting that routes those passes to a different model when configured, falling back to the session model. Also send anthropic-style cache_control markers for anthropic/* Prime Inference models so prompt caching engages on gateways that pass them through to the upstream Anthropic API, matching the Anthropic cache pricing the catalog already applies to those entries. * fix(coding-agent): harden auxiliary refinement model review findings Treat malformed persisted auxiliaryModel values (non-string) as unset so refinement falls back to the session model instead of throwing, and drop caught error details from the fallback warning since auth-stack errors can embed credential material (CodeQL js/clear-text-logging). * refactor(coding-agent): drop unused setAuxiliaryModel setter Cursor review flagged the setter as speculative surface: no product or test code calls it, so the auxiliary model is configured only through settings files and read via getAuxiliaryModel.