fix: bind $W_IDS in program Cypher, readable link labels, tool hints

[aaronsb]

Summary

• $W_IDS binding: dispatch_cypher now receives the working graph and passes accumulated concept IDs as a Cypher parameter. The seed-then-expand pattern (API search → Cypher expansion via $W_IDS) was documented but never actually worked — now it does.
• Link labels: Program execute and chain output resolves link endpoints to concept labels instead of raw content-hash IDs (e.g. Infrastructure Sovereignty → INFLUENCES → Geopolitical Computational Resources instead of sha256:26f76_chunk1_dc5194d6 → INFLUENCES → sha256:26f76_chunk1_1b374980).
• Cross-tool hints: Added terse hints to tool descriptions so calling LLMs discover related tools naturally — search→connect→program escalation, source/details for verification, epistemic status filters, polarity axis prerequisites. Sharpened the program escalation threshold from vague "if you're stuck" to concrete "more than 2 concepts or analytical questions."

Test plan

• Verified $W_IDS binding: seed-then-expand program executed successfully (2 seed nodes → 4 nodes + 2 links after expansion)
• Verified link labels render as concept names in MCP output
• Run cd cli && npm test for CLI unit tests
• Run docker exec kg-api-dev pytest tests/ -x -q for API tests