skills: rename agent-eval to codegraph-lift; portable runners table

[bompus]

agent-eval names nothing about what the skill measures and collides with a growing family of unrelated eval harnesses. codegraph-lift names the actual comparison: agent behavior with CodeGraph vs plain grep/read.

What changed (rebased cleanly on upstream main; supersedes #1896, which was based on our fork tree and conflicted on unrelated CHANGELOG drift):

• .claude/skills/agent-eval/ → .claude/skills/codegraph-lift/ (skill + corpus), frontmatter and self-references updated.
• Live pointers updated (add-lang corpus path + invocation, one playbook line). Harness paths (scripts/agent-eval/), memory keys, and dated benchmark docs intentionally untouched.
• Adjacent one-liners: dropped a stale 0.7.10 version example, tmux prereq scoped to the interactive arm, paid-run disclosure on step 5.
• New Runners section: headless is the portable arm (per-host commands with verified/proven status); tmux arm marked Claude-TUI-specific; parse-*.mjs noted as Claude-format (other hosts need their own parser — follow-up, not this PR).

Validation: corpus.json re-parsed clean; no live /agent-eval or skills/agent-eval references remain outside harness paths, memory keys, and dated docs (verified by grep).

Review start: the renamed SKILL.md frontmatter + Runners section, then the add-lang diff.