fix(skills): load symlinked skill slash commands

[wysie]

Summary

• Preserve lexical paths under trusted skill roots when loading skill slash commands
• Fix /skill-name invocations for skills symlinked under ~/.hermes/skills
• Add regression coverage for symlinked skill directories

Why

Skill slash commands cache the discovered skill directory. When that directory is a symlink, _load_skill_payload() resolved it before normalizing it for skill_view(). If the symlink target lives outside ~/.hermes/skills, the loader passed an absolute path through and skill_view() rejected it with Non-relative patterns are unsupported.

This keeps the trusted, visible path under the configured skills root intact before falling back to resolved-path handling.

Safety

• Only changes path normalization for skill slash/preload loading
• Does not make arbitrary external paths trusted
• Existing skill_view() platform, disabled-skill, collision, and warning behavior remains in place

Test Plan

• ./scripts/run_tests.sh tests/agent/test_skill_commands.py tests/tools/test_skills_tool.py -q
• ./scripts/run_tests.sh tests/agent/test_skill_commands.py::TestScanSkillCommands::test_loads_skill_invocation_from_symlinked_skill_dir -q

[alt-glitch]

Duplicate of #27662 — same fix for symlinked skill slash-command loading in _load_skill_payload(). Both normalize the lexical path under trusted skill roots before resolving symlinks.