Follow up to: #3576100 **Description:** AI Context selection is currently deterministic. Future smart selection may use an LLM to re-rank or filter candidate context items, but this should be designed separately from the RC1 deterministic pipeline. Key decisions to resolve: - Whether LLM ranking should run synchronously during selection or asynchronously ahead of time. - What candidate limit should be sent to the model. - How ranking calls should be cached. - What cache TTL and invalidation rules should apply. - How token and cost budgets should be enforced. - How failures should degrade back to deterministic ranking. - Whether ranking should be implemented as an event subscriber or a dedicated smart-selection plugin. This is suitable for post-RC1 because the deterministic selector can remain stable, and LLM ranking can plug into the selection pipeline later as an optional enhancement.