Latest V2 improvement to dev: verifier hardening, QD evolution improvements, and workflow runner isolation

[Fosowl]

Summary

This PR merges the mimosa_v2 development branch into dev, bringing ~85 commits of verifier hardening, quality-diversity evolution improvements, workflow runner isolation, and CLI robustness fixes.

What's Changed

Verifier & Grounding

• Per-claim verification scripts now use AST parsing, file-first analysis, and vacuous-antecedent rejection
• Source B extracts literal identifiers from goal body; Source C scans image artefacts with PIL; Source E flags dataset-sentinel leakage in selection
• Bounded retry on verifier-side failures with per-claim failure handling
• Grounding retriever now plans multi-step targeted literature sub-searches (dropped regex cues)
• Reduced verifier temperature and shorter prompt gradients for more consistent outputs

Evolution & QD

• Failure-fingerprint descriptor centered for QD novelty computation
• Genotype-embedding novelty + length penalty in selection
• Plateau-based stagnation replaces embedding stagnation in variation engine
• Reduced crossover rate for more stable evolution
• Per-goal evolution trees rendered into workflow folders

Workflow Runner & Orchestrator

• Managed venv isolation: workflow dependencies installed into isolated virtual environments
• max_tokens capped at 16384 to prevent token-limit crashes
• Random workflow temperature sampling for diversity

CLI & UX

• Hardened JSON parsing across all CLI tools (gibberish-tolerant prompts)
• Improved error handling in evaluation_cli, memory_chat_cli, and onboard_cli
• CSV recovery prompts now respect user-typed values
• Execution sandbox stdout truncation fixed

Agent & Misc

• Smolagent additional_authorized_imports updated
• Agent max steps bumped to 129
• Python 3.12 runner specification enforced
• Docs synced with multi-agent evolution claims

Breaking Changes

None expected — all changes are additive or internal refactoring.

Testing

• Verifier per-claim scripts pass on test goals
• Workflow runner venv creation tested locally
• CLI error handling verified with malformed inputs
• Evolution pipeline runs end-to-end without stagnation false-positives

Contribution checklist

Please confirm the following before requesting review:

• [x ] I have read CONTRIBUTING.md, docs/licensing-notes.md, and the repository license information (LICENSE, NOTICE) for this repository (Apache License 2.0).
• [x ] I have read docs/cla-process.md and understand the CLA workflow for contributions.
• [x ] I confirm that I have the legal right to submit this contribution.
• [x ] I have signed the required short Individual Contributor Agreement, or I will complete it if requested.
• [x ] If this contribution is made in the course of employment or under institutional IP rules, I understand that maintainers may also request the optional employer authorization.
• [x ] I am not knowingly submitting code that is incompatible with this repository’s Apache 2.0 licensing terms and CLA requirements.