refactor: clear Konjo DRY, dead-code, and complexity gates

[konjoinfinity]

Summary

A full Konjo quality check surfaced three failing gates on main — DRY (merge-blocking, 67 violations), complexity (4 functions over CLAUDE.md's ≤15 limit), and dead code. This PR clears all three with behavior-preserving refactors. No production logic changed.

DRY (67 → 0)

• generate/generator.py — extracted _BaseGenerator (shared async stream() bridge + _format_prompt) and _OpenAIChatGenerator. OpenAIGenerator/SquishGenerator now inherit one generate/generate_stream implementation instead of triplicating it.
• ingest/chunkers.py — extracted _EncoderBackedChunker mixin for the lazy-encoder load shared by SemanticSplitter and LateChunker.
• Route tests — added tests/unit/_route_harness.py (SettingsStub, GeneratorStub, make_app, clear_proxy_env, default result fixtures); the crag / self-rag / decomposition / timeout route tests now import it instead of redefining the same scaffolding.

Dead code (vulture)

• Dropped the unused tracer_name param from cache.tracing.cache_span (get_tracer() takes no name; no caller passed it).
• Renamed VectorStoreAdapter.upsert param vectors → embeddings to match the concrete QdrantStore.upsert signature.

Complexity (all konjoai/ functions now ≤ 15)

Function	Before	After
suspicious.scan_for_suspicious	22	14
analytics.compute_analytics	19	12
crag.CRAGEvaluator.run	16	7
cache route _kmeans_cluster	16	A

Verification

• pytest — 1287 passed, 7 skipped
• Coverage — 83.3% (≥80 gate)
• DRY — 0 · vulture — clean · complexity — 0 functions >15 · ruff check/format — clean

Not included

• Repo-wide ruff format drift (131 files) is driven by a newer ruff than the repo was formatted with and is continue-on-error in CI — left out to keep this diff focused; worth a separate version-pin PR.
• Docstring coverage (interrogate 64% in konjoai/) — non-blocking in CI (80, continue-on-error); not addressed here.

🤖 Generated with Claude Code

---

Generated by Claude Code