feat: OpenAI-compatible inference backend alongside Ollama

[wikithoughts]

What

Adds PW_INFERENCE_BACKEND=ollama|openai + PW_INFERENCE_API_BASE to council/ollama.py, the single shared HTTP client used by worker, researcher, judge, batch, and rerank. Since every one of those callers already goes through this one module's base() / generate() / smallest_chat_model(), this gives all five a new backend option with zero changes to those callers.

• Default (ollama, or unset): behavior unchanged — /api/generate + /api/tags, PW_OLLAMA_BASE.
• PW_INFERENCE_BACKEND=openai: speaks the OpenAI-compatible /v1/chat/completions + /v1/models shape that both llama.cpp-server and LM Studio expose out of the box (2026), resolving the endpoint from PW_INFERENCE_API_BASE (no hardcoded default — there's no sane guess for an arbitrary local port the way localhost:11434 is for Ollama).

generate()'s public signature and return shape (tuple[str, int] of (text, tokens)) are unchanged; internally it now dispatches to _generate_ollama/_generate_openai. The openai path mirrors the existing Ollama path's error-tolerance contract exactly (0 tokens / empty text on a missing field, never a crash).

Out of scope (intentional)

Other call sites that talk to Ollama directly (council/local.py, council/library.py, council/doctor.py, council/net/agent.py, council/net/baseline.py, council/operator.py) are not routed through this dispatch and remain Ollama-only — documented as a known limitation in the module's docstring, not an oversight. D1 (no token), D4 (no proxied traffic), and D18 (no browser automation) are unaffected — this is purely a local-inference-runtime choice.

Config

Registered PW_INFERENCE_BACKEND and PW_INFERENCE_API_BASE in council/config.py's KNOWN registry so they're discoverable via pw config list.

Tests

New tests/test_ollama.py (module had zero dedicated tests before this PR): base() resolution on both backends, generate()'s request shape on both backends (mocked requests.post, asserted on URL/body), resolve_timeout() precedence, and the two new KNOWN keys.

Verification

• python -m py_compile council/ollama.py council/config.py tests/test_ollama.py — clean
• ruff check . — all checks passed
• pytest tests/ -q — 521 passed, 1 skipped, 0 failed (full suite, including everything touching council/ollama.py / council.config.KNOWN)
• No new third-party dependency — only stdlib os and the already-core requests package, so the core-install-no-extras CI smoke job is unaffected.

🤖 Generated with Claude Code