skillgen: tell extraction subagents not to emit JSON-Schema keyword nodes

[harshslr]

Adds one rule to the extraction subagent prompt: don't emit nodes whose label is a bare JSON-Schema keyword.

Why

A JSON-Schema contract file has no named entities at the keyword level, but nothing in the prompt says so. A subagent handed one emits a node per description / type / $ref / required / properties / enum occurrence — and those recur once per definition, so they arrive in bulk.

On a real corpus (92 files, 669k words) a single 160 KB event-contract schema contributed 362 such nodes and 726 edges — 13% of the graph — which Louvain then grouped into ~12 communities whose entire membership is description ~ $ref ~ type. They swamped the community list and skewed the god-node ranking.

What changed

One rule in tools/skillgen/fragments/references/shared/extraction-spec.md and its compact variant:

JSON-Schema / contract files (*.schema.json, events-*.json, anything with a top-level $schema or $defs): extract the domain the schema describes - event types, payload names, enum values, field semantics, the definitions it composes. NEVER emit a node whose label is a bare schema keyword […]. Link a definition straight to its fields and values (envelope contains event_id, dataClassification contains CUI), never via the keyword that nested them.

The other 28 files are generated: python -m tools.skillgen then python -m tools.skillgen --bless. 30 files, 30 insertions, no deletions.

Relationship to #2255

Same node explosion, other half of the problem. #2255 is the AST path — json_config._is_config_json() treats any JSON with a top-level $schema as a config manifest and walks the whole keyword tree, so the nodes appear even with semantic extraction switched off. That needs a code fix. This PR only hardens the semantic pass, and is useful independently: a subagent can produce these nodes from a schema file the AST extractor correctly skipped.

Testing

tests/test_skillgen.py, tests/test_skill_version_warning.py, tests/test_claude_md.py — 81 passed. The --bless step was required; test_check_passes catches an un-blessed snapshot, which is how I found it.

Note for maintainers

main is ~730 files behind v8 (the default branch) — it has no graphify/extractors/ or tools/skillgen/. Anyone reading the repo tree from main sees a skeleton. Might be worth deleting or clearly marking it; I lost a few minutes to it while researching #2255.