feat(marketplace): kg-swarm skills and agents

[ohdearquant]

Summary

Adds the kg-swarm coordination pattern to the kg marketplace plugin:

• 2 new skills: skills/gap/SKILL.md (gap-analyst rubric — rank-1 frontier, density, orphans, missing relations) + skills/expand/SKILL.md (per-concept neighborhood exploration, citations mandatory).
• 5 new agents: agents/gap-analyst.md, agents/digester.md, agents/expander.md, agents/polisher.md, agents/librarian.md — each with per-role pickup/handoff protocols.
• README updates documenting the swarm pipeline + 6-agent table.

How it works

The 5 agents form a closed loop coordinated through the GTD substrate:

gap-analyst → expander → polisher → gap-analyst (next cycle)
                                      ↑
                                  librarian (curator)
                                      ↑
                                  digester (ingest)

Each agent reads tasks assigned to it from the substrate, processes, then assigns follow-up tasks to the next agent. No central orchestrator needed once the loop is seeded.

Notes for review

This is one of three PRs split out of the original #72. The other two PRs are:

• fix(runtime): update(properties=...) merges instead of replacing (#71) #74 fix/update-merge-semantics-code — runtime update(properties) deep-merge fix
• style/adr-deno-fmt — pre-existing deno fmt drift fix on ADRs 029, 035-041

No runtime/code changes — pure additions in marketplace/kg/ (skills + agents + README).

Test plan

• All new files lint-clean (deno fmt --check marketplace/kg/)
• Manual dogfood: ran 2 full swarm cycles via the gap → expand → polish loop against the live kg KG. Cycle 2 surfaced BUG: update() returns ok=true but properties writes silently do not persist (polisher cycle-1 → cycle-2 evidence) #71 as a side-effect (which fix(runtime): update(properties=...) merges instead of replacing (#71) #74 fixes).

🤖 Generated with Claude Code