v0.1.35

Added

• Named provider registry — okf/config.py now ships built-in provider presets (anthropic, openai, deepseek, gemini, glm, openrouter, dashscope, minimax, ollama, lmstudio, local) with default base_urls. Users reference by name instead of repeating URL+model.
• Per-mode provider routing — enrich.description, enrich.deep, enrich.security, and enrich.semantic_related each resolve their own provider independently. Resolution cascade: enrich.{mode}.{key} → providers.{name}.{key} → llm.{key}.
• LLM enrichment extensions — enrich_concept() now also sets design_pattern and LLM-suggested semantic tags (merged, deduped). New enrich_concept_deep() second pass (gated behind enrich.deep.enabled) reads actual source body via source_lines to produce usage_example, side_effects, security, complexity.
• Deterministic deprecation detection — _detect_deprecation() uses regex over docstring+decorators, no LLM call needed.
• _get() multi-level dot-notation — now supports arbitrary depth (e.g. enrich.deep.provider resolves correctly).
• Anthropic SDK support — _resolve_client() uses anthropic.Anthropic for "anthropic" provider, openai.OpenAI for all others.
• 7 new unit tests — provider resolution cascade, per-mode overrides, built-in provider lookups, multi-level _get.

Changed

• config.py DEFAULTS now includes providers.* and enrich.* sections. Default LLM provider changed to "local".
• load() uses copy.deepcopy(DEFAULTS) instead of mutable reference merge — prevents config file loads from mutating module-level defaults (fixes test isolation).
• enrich_concept() now also accepts inheritance context in prompt, max_tokens bumped 300→400.
• Version bumped to 0.1.35.

---