docs(grading): align reference answers and artifacts

[christso]

Summary

• update public eval-authoring docs and examples to use vars.expected_output plus explicit llm-rubric consumption
• document public grading artifacts as pass/score/reason with recursive component_results, optional assertion, named_scores, and metadata
• update Promptfoo parity, README/SDK README, Autoevals/batch/custom-grader docs, and AI-facing eval-writer guidance to avoid legacy authoring/result shapes

Validation

• git diff --check
• bun run validate:examples (108/108)
• bun --filter @agentv/web build
• bun --filter @agentv/web check:routes

Note: bun install was run in this worktree only after the first docs build failed with astro: command not found; it hydrated dependencies and did not modify tracked files.

Tracker

• av-kfik.28.7
• final live provider + real LLM grader dogfood remains tracked by av-kfik.16

[cloudflare-workers-and-pages]

Deploying agentv with    Cloudflare Pages

Latest commit:	f54dd8d
Status:	✅  Deploy successful!
Preview URL:	https://dc8d517f.agentv.pages.dev
Branch Preview URL:	https://docs-av-kfik-28-7-grading-au.agentv.pages.dev

View logs