Support reviewable per-skill permissions

[LadyBluenotes]

Problem

intent.skills currently permits packages as whole sources. Consumers cannot select one skill from a package without permitting its siblings, so they must use a broader package permission and express the narrower choice through exclusions.

User outcome

Consumers can explicitly permit selected skills from a package while preserving clear package and source-kind decisions.

In scope

• Extend the permission model to support selected skills within a package.
• Preserve package-level entries, npm and workspace source-kind matching, wildcard behavior, nearest non-null declaration inheritance, and intent.exclude precedence.
• Preserve strict parsing: malformed entries fail the complete permission list and are reported together.
• Keep package-level permissions valid and behaviorally compatible.
• Apply per-skill permissions consistently to list, load, stale, install --map, and hook catalogs.
• Extend the current redaction rule to per-skill permissions: humans can explicitly review hidden candidates outside agent sessions, while agent sessions receive counts without hidden source or skill identities.
• Document the final permission grammar and its interaction with package-level entries, wildcards, inheritance, and exclusions.

Out of scope

• Interactive permission editing.
• Content locks, hashes, or byte-level verification.
• Git sources.

Acceptance criteria

• A consumer can permit selected skills from a package without permitting every skill in that package.
• Existing package-level npm and workspace entries retain their current meaning.
• Source-kind matching remains enforced for package-level and per-skill permissions.
• intent.exclude continues to take precedence over permissions, including skill-level exclusions.
• Permission inheritance retains the current nearest non-null declaration behavior.
• Malformed entries fail the complete permission list and report all malformed entries together.
• list, load, stale, install --map, and hook catalogs enforce the same per-skill decision.
• Human review outside agent sessions can identify hidden candidates.
• Agent sessions receive hidden counts without hidden source or skill identities.
• Documentation includes valid examples and rejects ambiguous or unsupported forms.

Dependencies

None.

Open decision

Choose and document the per-skill permission syntax, including whether it extends the existing string grammar or uses a structured entry. The syntax must remain unambiguous for package names, source kinds, wildcards, and future source kinds.

Prior exploration

#204 explored this capability inside a broader installer, lockfile, and delivery change. This issue defines only the per-skill permission boundary and does not adopt that branch wholesale.