Cross-examination pipeline pattern (propose → attack → rebut)

Support a workflow where one agent proposes a solution and another attacks it, with configurable rounds of rebuttal. The attacker tries to refute the proposer's work; a judge rules on what survived.

Why

Self-review is weak. A model is measurably worse at finding holes in its own reasoning than at attacking another's. Cross-provider examination (e.g. one model proposes, a different model attacks) decorrelates errors further.

Status

Unblocked by #174 for fixed-round implementations: agents can now receive each other's prior outputs through downstream step context. DAG loops are still not supported, so the practical first version should model a fixed number of rounds as sequential steps.

Example flow

1. Proposer agent produces a solution
2. Attacker agent receives the proposal and tries to refute it
3. Proposer receives the attack and rebuts
4. Repeat for a configured fixed number of rounds
5. Judge agent rules on survived/conceded/open points

Initial scope

Start with fixed-round sequential steps. A later dynamic-loop design can be tracked separately if repeated rounds need to continue until convergence.

Start by tracing the downstream step context described in the issue and the existing sequential-step execution path enabled by #174. Model the proposer, attacker, rebuttal, and judge as a fixed sequence of steps with configurable rounds; done means the workflow runs without DAG loops and passes the project's relevant tests.