feat(autoresearch): supervise real GAN experiments

[FluffyAIcode]

Summary

• add the missing top-level AutoResearch supervisor
• use a real Gemma strategy-agent inference to select and rewrite candidate strategy
• deploy candidate chunk configuration to the real allens LaunchAgent and verify the process args
• clear allens and Primary caches before every experiment
• run one real full-context Generator/Critic/Proof-Ledger GAN experiment
• evaluate with immutable prepare.py hard constraints
• compare candidates lexicographically by unresolved proof obligations then cold Critic Prefill
• atomically keep or restore candidate, Agent state, Proof Ledger, and deployed worker config
• append every experiment to auditable results.tsv
• fail and stop on proposal, deploy, telemetry, GAN, evaluator, or restore errors; no fallback path

Candidate surface

Each candidate contains a target proof obligation, falsifiable hypothesis, distinct Generator/Critic directives, and Prefill chunk strategy. Full context, final-only snapshots, and no fallback remain hard invariants.

Tests

• executable candidate rendering and strict validation
• lexicographic proof/performance keep decisions
• append-only results and best-baseline selection
• existing evaluator, GAN state machine, ledger, and distributed paths
• targeted suite — 23 passed
• full platform-neutral suite — 888 passed, 1 skipped
• Python compile, shell syntax, and diff checks pass

Real acceptance plan

After merge, stop the direct GAN auto-loop at a completed checkpoint and run one supervisor iteration against Primary + allens. Acceptance requires a real candidate proposal, verified deployment, cold-cache GAN report, evaluator pass, results.tsv row, and observed keep/revert decision.

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.