Let arranging operators return a caller-chosen agent

[antiguru]

arrange_core, reduce_with_tactic, and arrange_from_upsert build their trace and hard-wire the reader they return to TraceAgent, which hides the trace inside its TraceBox. A trace that shares state with something outside the operator, such as a handle through which other threads observe its batches, can only hand that state back through TraceAgent::trace_box_unstable.

This adds an Agent trait for the reader an operator returns. Agent::new takes the trace by value and returns the reader and a TraceWriter, and TraceAgent implements it by delegating to TraceAgent::new. arrange_core_with_agent, reduce_with_tactic_and_agent, and arrange_from_upsert_with_agent return Arranged<A> for a caller's agent, which can take the trace's shared state before wrapping a TraceAgent. The existing functions delegate with TraceAgent, so no caller changes. The writer stays concrete, since TraceWriter::new and TraceBox::new are already public.

This is the alternative to #915, which injects a trace factory instead. Here the handle becomes part of the returned type rather than leaving through a closure. Open points for discussion:

• Fixing TraceWriter<Self::Trace> in the trait rules out agents that intercept writes. Adding a writer associated type later would break implementers.
• Reduce and upsert read and compact through their own clone of the agent, so an agent must honour the TraceReader contract of the trace it wraps.
• Only the lowest-level entry points get variants. reduce_abelian, reduce_core, and reduce_trace still return TraceAgent.

Posted by Claude Code.

🤖 Generated with Claude Code

https://claude.ai/code/session_015tLhSbZdXrTSK2KwSocT59