Milestone 3: generalized taskset routing and rollout data path

Work:

• Give dataset declarations stable framework identity.
• Resolve processors after agent aliases, composition, agent_map, and fan_out.
• Prefer explicit native binding; otherwise choose the compatibility processor.
• Persist processor identity through materialized inputs, successes, sidecars, retries, shards, merges, and aggregation.
• Preserve logical agent attribution separately.
• Keep processor-specific request adaptation and result projection out of rollout collection.
• Delegate aggregation to existing agent behavior for the standard single-agent processor.

Definition of done:

• A batch can mix native and compatibility-routed tasksets.
• Shared resources servers and repeated dataset sources do not create ambiguous routes.
• Aliases, copies, agent_map, and fan_out resolve the correct processor.
• Sharded and unsharded outputs retain equivalent processor attribution and results.
• Collector code imports no concrete processor request/response models.