SkillCard and SkillCollection UX improvements

Improve how users author, discover, validate, and consume SkillCards and SkillCollections across image, git, and inline sources.

## Scope

- [ ] #153 — Decide the SkillCollection authoring model: generated versus hand-written SkillCards, duplicate detection, type defaults/overrides, and inline entry shape.
- [ ] #31 — Create owned child SkillCards for git-sourced SkillCollection entries, with lifecycle management. Coordinate implementation with the authoring decision in #153.
- [ ] #185 — Spike an internal container registry for non-image skill sources: resolve/validate inline and git content, package it as OCI images, and evaluate mounting immutable digests through ImageVolumes for SkillCard and SkillCollection consumers.

## Relationship between the work

#153 establishes the authoring experience; #31 implements the git-source child-card behavior. #185 investigates whether a common image delivery path can improve validation timing, consistency, and reuse for non-image sources. Consider these together so enumeration and source delivery do not impose conflicting authoring models.

The registry remains a spike, not an accepted implementation choice. Evaluate its operational cost and kubelet pull requirements against the benefits of ImageVolumes and the current ADR 0015 design. Any change to that design should follow the repository ADR conventions once the findings support a decision.

## Release scope

Decision from the v0.11.0 milestone review (2026-09-16): this work belongs in Next and does not block v0.11.0. Existing child issues retain their detailed acceptance criteria; this umbrella coordinates them.

Related work: #154 (private sources), #155 (disconnected mirroring), #172 (assembled-skill visibility), and #194 (readiness propagation). These are related considerations, not additional required children of this umbrella.

Start by reading the detailed acceptance criteria in #153 and #31, then review the registry spike in #185 and ADR 0015. Done means the authoring decision, git child-card behavior, and registry/ImageVolumes findings are coordinated, with any resulting design change following the repository ADR conventions.

Written by the indexing model from the issue text.