feat(distill): add OpenAI-compatible local LLM backend

[senna-lang]

Summary

Makes the distillation LLM provider-switchable so distillation can run against a local OpenAI-compatible endpoint (Ollama / LM Studio / llama.cpp-server / vLLM) instead of claude --print. Zero new dependencies (stdlib urllib), full backward compatibility — provider defaults to claude and existing configs behave identically.

Implements proposal openspec/changes/add-local-distill-backend (L1–L5).

What changed

• config (config.py): add distill.provider (claude|openai) and distill.base_url with validation — unknown provider, or openai without base_url, warns to stderr and falls back to claude.
• llm (llm.py): add frozen DistillBackend dataclass + from_config; turn call_claude into a dispatcher (existing subprocess path moved verbatim into _call_claude_cli). _call_openai POSTs to {base_url}/chat/completions with response_format=json_object, temperature=0, no Authorization header (local-only). Shared _strip_json_fence / _build_distill_prompt helpers.
• validation (llm.py): add _validate_palace + LLMValidationError with a one-shot regenerate on validation failure — also closes a pre-existing latent bug where raw["exchange_core"] was accessed unvalidated.
• wiring (distiller.py, cli/distill_cmd.py, cli/__init__.py): thread backend through distill_exchange / distill_all, built from config.
• docs: README (Distilling with a local LLM), CLAUDE.md, and the generated config.toml template document provider / base_url with Ollama/LM Studio examples.

Design notes

• One OpenAI-compatible path covers Ollama / LM Studio / llama.cpp-server / vLLM via base_url only.
• call_claude name kept as the dispatcher so the existing test seam (patch("codeatrium.distiller.call_claude")) is unchanged.
• No API key — Authorization is never sent; authenticated remote endpoints are out of scope.

Testing

• 266 tests pass; ruff and pyright clean (verified via pre-commit hook).
• New tests: config provider/base_url parsing & fallback; dispatcher routing; _call_openai URL/body/no-auth (urllib mocked, network-independent); fence stripping; validation + one-shot retry; backend pass-through with unchanged patch points.
• Live-verified against Ollama (qwen2.5-coder:14b): loci distill produces a valid palace object end-to-end (DB restored afterward — non-destructive).

Known limitations (follow-up)

• Small models can emit degenerate/malformed JSON (qwen2.5:14b looped; qwen2.5-coder:14b was 5/5 valid). Malformed JSON raises json.JSONDecodeError, which the one-shot retry (catching only LLMValidationError) does not recover — and temperature=0 makes retry deterministic anyway. Errors are isolated per-exchange and the exchange stays pending. Candidate for the planned model-benchmark + robustness follow-up.

Out of scope

Docs-only config.toml provider UI in loci init, embedding localization, authenticated remote endpoints, streaming/token accounting, Ollama native /api/generate.

🤖 Generated with Claude Code

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.