[FEATURE] Explicit local-first precedence for synced skills

Feature Description
Add an explicit precedence model for skill resolution so synced external skills do not silently outrank workspace-local skills. OpenSkills already syncs installed skills into AGENTS.md for agents to discover and use, but the current model is not explicit about how authority should work when both a local skill and a synced external skill exist with the same name.

Use Case
Many agent setups use AGENTS.md as an early project instruction source, and Codex explicitly reads AGENTS.md before doing work. At the same time, Claude Code’s memory model encourages keeping root bootstrap context concise and loading more detailed guidance on demand. In mixed-agent environments, synced skill discovery is useful, but repo-local workflow and governance behavior should still remain authoritative. Without an explicit precedence rule, generic external skills can shadow local project intent. OpenAI Developers

Proposed Implementation
Please support and/or document a resolution order like:

1. workspace-local skills
2. user-local managed skills
3. synced/remote registry skills as fallback only

Possible implementation options:

• document local-vs-synced precedence rules
• add configurable resolver order
• support a “local-only” mode for selected skill names
• make sync output clearly distinguish local/project skills from external fallback skills

This fits well with OpenSkills’ existing support for configurable sync output paths and local/symlink-based workflows.

Compatibility Consideration
Please confirm:

• This feature maintains compatibility with Anthropic's SKILL.md specification
• This feature works across multiple AI agents (not just Claude Code)
• This feature aligns with the simplicity goal (.claude/skills only)

Alternatives Considered
Current workarounds include:

• manually avoiding name collisions
• using naming prefixes for external skills
• relying on project bootstrap files to restate precedence rules

These help, but they do not solve the underlying resolver-authority problem at the OpenSkills layer.

Additional Context
This request is about precedence and authority, not about reducing remote skill support. Remote skills are still valuable. The goal is simply to ensure synced external skills improve discoverability without overriding repo-local policy by default. OpenSkills positions itself as a universal skills loader for agents that can read AGENTS.md, which makes a documented precedence model especially important.