Add Prime Agent host-agent support

[yangwuan55]

What

Add Prime Agent as a third host agent alongside Claude Code and Codex. publish.sh would accept --agent prime and render the runtime layout for Prime Agent.

Why

Prime Agent is a terminal coding/research harness that reads AGENTS.md (or CLAUDE.md) as its context file. It already consumes .agents/skills/, but the existing Codex layout is not a good fit for it:

• The runtime manifest renders the asset-skill call as $asset-gen — Codex's skill-invocation syntax. Prime Agent does not understand $asset-gen; its own command is /skill:asset-gen.
• The Codex layout generates agents/openai.yaml metadata per skill, which Prime Agent ignores (it reads SKILL.md frontmatter directly).
• Prime Agent has its own native project skill location, .prime/agent/skills/, that a published repo could target for a clean, unambiguous layout.

Without this, a Prime Agent user gets a manifest they can't act on and no first-class publishing path.

Why not something simpler?

The simplest option is "just tell Prime Agent users to run --agent codex". But that leaves the $asset-gen call in the manifest, which the Prime Agent model can't invoke, and ships Codex-only metadata. publish.sh is already a per-agent render switch — adding a prime branch is the smallest change that makes the published repo correct for Prime Agent, mirroring how codex was added.

Additional context

• Prime Agent reads AGENTS.md and discovers skills from .prime/agent/skills/ and .agents/skills/ (project level, walking up from cwd).
• Skill command syntax: /skill:asset-gen.
• Verified end-to-end: publish.sh --engine godot --agent prime produces AGENTS.md + .prime/agent/skills/asset-gen/, and a Prime Agent session in the published repo discovers the asset-gen skill.