Add Agent Skills support to improve dn context estimates

[bpmooch]

Goal

Add initial Agent Skills support to dn, starting with more accurate dn context estimates.

dn context check currently reports inherited AGENTS.md context. Skills add context through progressive disclosure: every discovered skill contributes catalog metadata, while a skill's full SKILL.md instructions contribute only when that skill is activated. The command should represent those costs separately instead of assuming all installed skills are loaded eagerly.

Initial scope

• Discover project- and user-level skills using the locations and precedence described by the Agent Skills client implementation guide, including .agents/skills/ interoperability paths.
• Parse the required name and description metadata from each SKILL.md and retain its source path.
• Include the discovered skill catalog in dn context check byte and optional Claude token estimates.
• Report catalog overhead separately from inherited AGENTS.md instructions.
• Provide a way to include selected activated skills in an estimate, accounting for their full SKILL.md instructions without loading unrelated skill bodies.
• Expose the same breakdown in human-readable and --json output.
• Surface skipped, malformed, and shadowed skills without making cosmetic validation failures fatal.

Acceptance criteria

• dn context check <target> discovers the skills available to that target using deterministic project-over-user precedence.
• The default estimate includes inherited instructions and the tier-1 skill catalog, but not every skill body.
• Callers can request an estimate for one or more activated skills, and each selected skill is counted once.
• Human-readable output distinguishes inherited instructions, skill catalog metadata, and activated skill instructions.
• JSON output provides stable structured fields for the same categories and totals.
• Tests cover discovery scopes, name collisions, malformed metadata, no available skills, and activated-skill deduplication.
• docs/subcommands.md documents the new behavior and clarifies that results remain estimates of client-specific context construction.

Follow-up work

Full runtime skills support can build on this discovery and parsing layer: catalog disclosure to models, model- or user-driven activation, bundled resource access, permissions, and preserving activated instructions during context compaction.