feat(linkml-engineering): vendor tested Neo4j-targeting generators; add generated skill inventory

[costezki]

Summary

LinkML → Neo4j generators (tools/linkml_neo4j/), driven by openspec/changes/linkml-neo4j-generators:

• Fixed two confirmed bugs found by actually running the generators against a real schema (vendored from hulubul-broker, CC-BY-4.0, attributed): a relationship whose LinkML range is an abstract class resolved to an undefined Python class in the generated neomodel code; neomodel's reserved id/deleted/element_id attribute names crashed class definition outright (the schema's identifier slot is conventionally named id).
• Enriched the Cypher generator: existence constraints for required-multivalued properties (previously silently skipped), a --profile community/--profile full split (Community-safe subset vs. the unverified-on-Community reference output), and schema-driven CREATE INDEX via an opt-in neo4j_index annotation.
• Vendored both generators as tested repo tooling, not skill assets — a Skill's home is knowledge (spec/skill-repo-governance.md), not a maintained codebase, so linkml-engineering documents and cites them rather than owning the code.
• project-setup gains a real --neo4j scaffold.sh flag that projects a pinned, refreshable copy into a consuming repo — verified by actually running it (dry-run, real copy, skip-on-rerun, force-refresh, and confirming no scripts/ appears without the flag).

Generated skill inventory (docs/skill-inventory.md, tools/skill_inventory.py):

• A generated (never hand-edited) map of all 22 skills: bundle containment + purpose classification as one Mermaid diagram, and a second diagram for skill-to-skill relations distinguishing a mechanically-parsed "depends on" (from each skill's own Delegates text) from the weaker "related" (the rest of the Related list).
• Freshness enforced by tests/test_skill_inventory.py (fails make test on drift), same pattern as the existing .opencode/ drift gate.
• AGENTS.md documents the one manual step (make skill-inventory after adding/editing a skill) vs. what's automated (drift detection, everything else).

Test plan

• make validate (lint + test): 92 passed, 0 failed, 7 deselected (opt-in Docker-marked integration tests).
• Both Mermaid diagrams compiled with mermaid-cli and visually reviewed as PNGs, not just eyeballed as text.
• scaffold.sh --neo4j exercised end-to-end against a scratch directory (dry-run, real copy byte-identical to source, skip/force-refresh behavior, and confirmed --neo4j-less runs create no scripts/).
• The Neo4j Community/Enterprise constraint-support spike (tests/test_linkml_neo4j_integration.py, pytest -m docker) is written but not executed — this sandbox had no reachable Docker daemon. Needs a real run before treating its "passed" status as verified; see the generator's own module docstring for the current caveat.

🤖 Generated with Claude Code