feat(cli): Add cognee-cli demo with bundled zero-key graph

[Vasilije1990]

Description

First-session research shows the most common organic failure: the first search returns nothing (or an API-key error) before the user has ever seen cognee work. This PR ships proof of life: cognee-cli demo loads a pre-built knowledge graph and answers a search in ~5 seconds with no LLM key, no embedding provider, and no network access.

Linear issue key pending — the Linear MCP connection needs re-auth; the title will be updated with the COG reference before merge.

$ cognee-cli demo
Loading the bundled demo knowledge graph (no API key required)...
Success: Demo graph loaded into dataset 'demo' (37 nodes, 71 edges).

Query: Who works at Anthropic?
  1. ... Alice works at Anthropic in San Francisco. ...

This demo uses keyword search (CHUNKS_LEXICAL) — it needs no LLM and no embeddings. LLM answers over your own data need LLM_API_KEY set.
Next: cognee-cli search "your question" -t CHUNKS_LEXICAL -d demo

How it works

• Bundled COGX archive (cognee/cli/samples/demo_graph/, ~48 KB, ships in the wheel): built once by a maintainer via tools/build_demo_archive.py, which cognifies the existing quickstart sample in a scratch store, exports it with cognee.export(format="cogx"), and trims documents.jsonl — document records would route through add(), whose pipeline requires a working LLM connection. Chunk text survives as raw nodes (the exporter writes every DocumentChunk as one), which is exactly what lexical retrieval reads.
• Graph-only archive restore: remember(source, index_vectors=False) — reusing the kwarg the code route already declares — threads through the migration loader as add_data_points(graph_only=True) and sets skip_connection_test on the import pipeline. No vector engine is initialized, so no key is demanded. Works for both import shapes (streaming and buffered).
• Zero-cost search: queries run as CHUNKS_LEXICAL (BM25 over graph-stored chunks — no LLM, no embeddings, no network). Also fixes a standalone gap: CHUNKS_LEXICAL was missing from the CLI's SEARCH_TYPE_CHOICES, making it unreachable from cognee-cli search/recall entirely.
• Honest expectations: the command states it uses keyword search, that LLM answers over the user's own data need LLM_API_KEY, and prints copy-pastable next steps plus cleanup (forget --dataset demo). In --api-url mode the dispatcher's existing "run locally" error covers demo.

Test plan

• 24 new tests: bundled-archive integrity guards (ships in package, parses, carries no document records, contains lexical-searchable DocumentChunk raw nodes); graph_only threading through stream_graph_from_source, store_imported_graph, and the remember(index_vectors=...) mapping; demo-command behavior (graph-only import asserted, every search pinned to CHUNKS_LEXICAL, custom query/dataset flags, error wrapping).
• Full cognee/tests/unit/migration/ + cognee/tests/cli_tests/cli_unit_tests/ suites: 309 passed locally.
• End-to-end verified with LLM_API_KEY/OPENAI_API_KEY/EMBEDDING_API_KEY/ANTHROPIC_API_KEY all stripped from the environment in a scratch store: import + two searches complete in ~4.5 s with real chunk text returned.
• pre-commit (ruff check + format) clean.

🤖 Generated with Claude Code