test: expand test suite from 21 to 218 tests covering all modules

[s0j0hn]

Summary

• Add 13 new test files covering every previously-untested module: knowledge_graph, palace_graph, searcher, layers, dialect, entity_detector, entity_registry, general_extractor, room_detector_local, mcp_server, spellcheck, split_mega_files, version_consistency
• Fix resource leaks in test_config (temp dirs never cleaned, env var leak) and test_convo_miner (cleanup outside try/finally)
• Expand test_normalize from 3 to 15 tests covering all 5 chat formats, privacy export, Slack username fallback, mixed content, IOError path
• Expand test_convo_miner from 1 to 11 tests covering general extract mode, chunking edge cases, room detection, scan filtering
• Expand test_mcp_server with duplicate checking, graph traversal tools, graph stats
• Expand test_knowledge_graph with seed_from_entity_facts, NULL date queries, empty results
• Expand test_entity_registry with learn_from_text auto-discovery, disambiguation tie, mocked Wikipedia lookup
• Expand test_dialect with zettel encoding, file encoding with tunnels, count_tokens
• Add shared conftest.py with HOME isolation, tmp_dir, palace_path, populated_palace, config, kg, seeded_kg fixtures
• Simplify after code review: extract helpers, convert MCP patching to yield fixture, remove redundant assertions and comments

Before: 4 test files, 21 tests
After: 17 test files, 218 tests — all passing in ~8s

Test plan

• pytest tests/ -v — 218 passed, 0 failures
• ruff check tests/ — all checks passed
• ruff format --check tests/ — all files formatted
• No API keys or network access required

🤖 Generated with Claude Code

[s0j0hn]

Do what you want with this @bensig @milla-jovovich

[mvalentsev]

I opened #157 earlier with a smaller focused test file for general_extractor only — black-box coverage against the public extract_memories API (empty input, the five memory types at the default min_confidence, chunk index sequencing, the fenced-code-block skip, speaker-turn splitting with type assertions, and the return-dict shape). It might be worth looking at the two side by side: this one exercises some private helpers (_get_sentiment, _has_resolution, _is_code_line, _extract_prose, _split_into_segments) which #157 deliberately avoids to stay robust to refactors.

If either approach ends up getting merged, happy to rebase the other on top or drop the overlapping file so there's no conflict.

[bensig]

Thanks for the test expansion work. Coverage was bumped to 85% in #281 (tmuskal) which overlaps significantly. If there are tests here that aren't covered, a focused follow-up would be great.