feat: skill overlay Phase 2 — inspect, negation, sealed, policy engine

[mrwogu]

Context

Phase 1 of Skill Overlay/Extends was implemented in #198. This issue tracks Phase 2 and Phase 3 items from the design spec.

Spec: docs/superpowers/specs/2026-03-30-skill-overlay-extends-design.md (section 13)
Review: docs/superpowers/specs/2026-03-30-skill-overlay-extends-review.md (section 10)

Phase 2 (post-MVP)

• Diamond dependency priority resolution — when two registries at the same layer extend the same skill, priority determined by registries: declaration order in promptscript.yaml. Requires tagging imported blocks with source registry index. (Spec §5.1)
• prs inspect <skill-name> --layers — debugging command showing each layer's contribution to a composed skill (Spec §13, Review §10)
• Semantic base/overlay validation — detect when Layer 2 changes SKILL.md structure in a way that breaks Layer 3 overlays (Spec §7.3, Review §6.1)
• Reference negation syntax — references: [!./references/deprecated.md] to remove a reference added by a lower layer (Spec §13)
• sealed / final property — block override of content by higher layers (Spec §13, Review §3.2)
• Overlay-aware suggestion rules — prs init suggests overlay skills when both base and BU registries are configured (Spec §8, §13)

Phase 3 (roadmap)

• Integrity hashes in lockfile — hash references from registries for supply chain security (Spec §13, Review §4.3)
• Policy engine — validate that extensions comply with organizational rules (Spec §13, Review §10)

Enterprise context

Requested by Comarch for their 4-layer registry architecture (company → product → BU → project). See original proposal: /Users/wogu/Work/Comarch/promptscript-architecture-proposal.md.