Implement lazy resource loading (progressive disclosure)

[bug-ops]

Context

Audit finding AUDIT-005 from spec conformance review.

Problem

Agent Skills spec mandates progressive disclosure:

1. Metadata (~100 tokens): loaded at startup for all skills
2. Instructions (<5000 tokens): loaded when skill activated
3. Resources (as needed): loaded only on demand

Zeph implements steps 1 & 2, but prompt.rs eagerly loads ALL reference files on skill activation, violating progressive disclosure. This causes context bloat and unnecessary I/O.

Fix

1. Remove eager reference injection from format_skills_prompt()
2. Implement on-demand resource loading API (e.g. get_skill_resource(skill, path))
3. Agent requests specific references when needed

References

• Spec: https://agentskills.io/specification.md#progressive-disclosure
• Files: crates/zeph-skills/src/prompt.rs, crates/zeph-skills/src/resource.rs
• Parent audit: Align SKILL.md parser with agentskills.io specification #686