docs(evals): plan AgentV eval authoring extensibility

[christso]

Summary

• Adds the AgentV eval authoring extensibility plan produced from the public demo and private framework-parity research.
• Grounds recommendations in comparable framework patterns from promptfoo, Braintrust, Phoenix, and Pi coding-agent skills/extensions.
• Keeps the plan conservative: improve docs/templates/adapters first, add only proven primitive schema gaps to core.

Key Decisions

• Keep AgentV core as a lightweight primitive execution layer.
• Use private parity artifacts for promptfoo/Braintrust/Phoenix comparison instead of public competitor examples.
• Treat Pi coding-agent style skills/extensions as the authoring-pattern inspiration for progressive-disclosure eval builders.
• Prioritize benchmark provenance docs, rubric criterion operator support, narrow templates/adapters, and trace/replay follow-up Beads.

Evidence

• Branch includes the plan at docs/plans/2026-06-06-001-agentv-eval-authoring-extensibility-plan.md.
• Related Beads record completion of private parity artifacts and follow-up work split under av-r0s.5.