Add evolution_plan.md: audit, roadmap, and prioritized reliability tasks

[arty-kk]

Motivation

• The repository's evaluator and acceptance flow were found to be inconsistent with project test conventions leading to false-negative candidate rejections and untested rollback paths.
• A consolidated plan is needed to align runtime test execution, enforce evaluator-selector contracts, and add tests for critical evolution/rollback behaviors.
• The document organizes findings, goals, and a prioritized incremental roadmap to stabilize core reliability and developer UX.

Description

• Add evolution_plan.md which captures a baseline architecture map, critical flows, prioritized pain points (P0/P1/P2), and a three-phase roadmap to address them.
• Introduce atomic task specs (EVO-001..EVO-005) recommending replacing unittest discover with pytest -q, adding evaluator→selector contract tests, adding deterministic rollback and ExperimentManager.run_async integration tests, and normalizing LLM task contracts.
• Include concrete steps, acceptance criteria, validation commands, and rollback strategies for each task to guide implementation and verification.
• Enumerate explicit non-goals to constrain scope and a stopping rule to ensure the audit's actionable outcomes are addressed.

Testing

• No automated tests were added in this change and no CI jobs were executed as part of this PR.
• The plan recommends running pytest -q and python -m compileall -q src as validation commands for subsequent code changes.

---

Codex Task