fix(agents): require full-context semantic Critic review

[FluffyAIcode]

Summary

• pass the complete Generator response verbatim into the Gemma Critic
• remove extractive evidence windows, token budgets, sampling, and summarization
• enforce review_scope=full, equal full/context token counts, and zero omitted tokens before Critic inference
• instruct Critic to reason over the response as one semantic argument
• expose full review scope and context coverage in benchmark metrics/dashboard
• extend strict allens Prefill timeout from 900 to 3600 seconds for full responses

Semantic contract

A global Critic score is forbidden unless the full Generator output is present. There is no partial-review score, chunk score, weighted score, or semantic fallback.

Performance consequence

Full-context review restores semantic coverage at the cost of cold Prefill latency on the current 16GB allens Gemma worker. Heartbeats remain visible while the strict remote job runs.

Tests

• full-context identity and zero-omission invariants
• strict timeout launchd configuration
• targeted Agent GAN/launchd suite — 15 passed
• full platform-neutral suite — 859 passed, 1 skipped
• Python compile and diff checks pass

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.