test: run peer learning on the configured contributing node

[erikleblansch]

Change

Correct the disposable peer-learning topology to run the learning and republication participant on the already configured contributing R3 node. The original Client remains a distinct final named-content consumer.

The previous 6a782752 run correctly rejected 600 publication handoffs from a Client fixture whose content contribution was disabled. This change fixes that fixture role/endpoint mismatch; it does not weaken production contribution, policy, expiry or signature guards.

The one combined train-loop scenario still requires the original signed-catalog A-to-B selection and two real training cycles, cold protected source retrieval, measured peer adoption, five real peer-learning workers, successful signed successor contribution, the unchanged final Client named import/inference, packet evidence and complete cleanup.

Exact candidate and dependency

Head 68314466670704798bd8e57107127ea9f20902b3, tree fb8d5b84bcb201abd82c044099ce25578fdf0e7d; parent 342b8a80.
Depends on the preceding document-synthesis integration in #134. This PR targets current main directly so its own Quality and complete CodeQL checks run normally. Its incremental fix contains ten fixture/documentation files and no production guard changes.

Verification

• Four new focused tests, eighteen capture checks, nine smoke tests and the peer-helper self-test: PASS.
• Relevant shell syntax, ShellCheck, diff check and independent integration review: PASS.
• No model execution or development-host network changes.
• Exactly one agent-train-loop VM on this source is being requested; the complete combined outcome remains pending.
• The compute: synthesize public documents through hierarchical peer inference #134 document VM remains bound to its original 342b8a80 source and is not duplicated or relabelled.
• Own Quality and complete CodeQL remain pending.

Limits

Historical #130, #131 and #133 failures remain failures until the corresponding unchanged gates actually pass on a corrected source. Prior training and signature evidence is partial, not proof of successful republication. This is a disposable topology correction, not a new production role default, broader policy permission, better-answer guarantee, private offload or completed B05/alpha claim. No merge is requested while the required live outcomes or checks are pending.