[FEATURE] Allow loading skills from custom directory in --bare mode

When running claude -p in --bare mode for programmatic/eval use cases, skills are completely disabled. Without --bare, personal config (hooks, MCP servers, global CLAUDE.md, memory) leaks into the session.

There is no way to get:

• Config isolation (no personal hooks/MCP/memory)
• Custom skills loaded from a specified directory
• In non-interactive -p mode

Use case

Running LLM eval experiments where the user agent needs:

• A clean environment (no personal config influencing behavior)
• Specific skills installed (to measure skill effectiveness)
• Non-interactive execution via -p

Proposed solution

A flag like --skills-dir /path/to/skills that works with --bare.
Or: --bare could respect $HOME/.claude/skills/ from an overridden
HOME (currently it does not — Claude Code reads config from the real
user home at process startup, ignoring HOME override).

Current workarounds

• --bare + --add-dir loads CLAUDE.md but not skills
• apiKeyHelper via --settings enables non-bare auth, but the full

personal config still loads (hooks, MCP, all skills, global CLAUDE.md)

• HOME override only affects subprocess env, not Claude Code's own

config resolution

Related issues

• #25771 (programmatic skill deployment)
• #30278 (scoped skill discovery per subagent)