fix(agents): complete responses to EOS and ground critique

[FluffyAIcode]

Summary

• treat 64 tokens as a streaming chunk and continue Generate calls until model EOS
• add a separate 512-token safety cap; reaching it marks the stage incomplete and fails the Agent gate
• persist stop_reason, completion state, incomplete counts and stop-reason history in benchmark reports/UI
• pass Generator completion metadata to Critic
• prohibit Critic from calling an EOS-complete answer truncated or penalizing an honest statement that an open problem has no accepted proof
• retain explicit no-progress and server stop handling

Root cause

The interactive REPL used one 64-token Generate call, so both answers ended mid-sentence. Critic had no completion metadata and interpreted the client truncation as a reasoning failure; it also lacked an epistemic rule for unsolved mathematics.

Tests

• chunk continuation to EOS and explicit safety-limit tests
• incomplete stage gate and report aggregation tests
• Critic/REPL privacy and cache gate tests
• focused suite — 9 passed
• full platform-neutral suite — 852 passed, 1 skipped

Made with Cursor

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.