feat: type-level BM25 search corpus and search_types()

[jakerobinson]

Summary

Adds a fourth BM25 search surface to the rsc-client: a type-level corpus that aggregates all field vocabulary for each GraphQL type into a single document. Exposes it via search_types(search), mirroring the existing search_operations / search_fields pattern.

Why this matters: Operations like allClusterConnection and slaDomains return complex types (Cluster, GlobalSlaReply) with 50+ fields each. When an agent searches for "CDM clusters" or "retention-locked SLAs", BM25 can't reliably surface these operations via the ops corpus — the type's rich vocabulary gets diluted across 228-token operation documents and length-normalized into oblivion. The type corpus solves this by indexing each type's complete vocabulary in one place and resolving back to operations.

Changes

• mcp_indexer.py — build_types_bm25_corpus() as a fourth indexer step. Tokens: camelCase type/field name splits with org→organization and vm→virtual machine synonym expansion, first-8-word field descriptions, and one level of Connection field traversal to include child type vocabulary. Op resolution covers direct return-type, Connection→node, and interface→implementor paths.

• index.py — search_types(search) with lazy BM25Okapi loading. Returns up to 10 results with name, ops, and score keys.

• __init__.py — exports search_types alongside existing exports.

• mcp_types_bm25_corpus.json — 1,562 reachable types, 1.3MB.

• schema-update.yml — adds mcp_types_bm25_corpus.json to the git add in CI so it's committed on each schema update.

Test plan

• 8 new tests in tests/test_search_types.py
• search_types("cluster") returns Cluster type with cluster operations in ops
• Scores > 0 for relevant queries
• Results capped at 10
• CI schema-update workflow commits the regenerated corpus on schema changes