Add Ollama provider for local code review + retune default temperature

[Airwhale]

Summary

• Adds a third --provider ollama path that POSTs to a local Ollama server's OpenAI-compatible /v1/chat/completions endpoint. No API key, no token cost, code never leaves the machine.
• Refactors model aliases into per-provider tables (MODEL_ALIASES_BY_PROVIDER); adds local / local-pro aliases for Ollama; preserves OpenRouter aliases unchanged; surfaces cross-provider mismatches as typed CONFIG errors that name the right --provider.
• Lowers default temperature from 0.5 to 0.3 after cross-model integration testing observed google/gemini-2.5-pro hallucinating a HIGH-severity finding referencing a CLI flag that does not exist in the code (the proposed fix would have crashed the runner with AttributeError).
• Updates README.md, .env.example, and docs/llm-code-review-runbook.md for the new three-provider model and the retuned temperature default. Adds a "Local vs cloud" section documenting the opposing failure modes (cloud hallucinates, local under-reports).

What the integration adds

New surface	Default	Purpose
--provider ollama	—	Third provider option
--ollama-host URL	$OLLAMA_HOST or http://localhost:11434	Override for non-default ports / remote Ollama / WSL without localhost mirroring
OLLAMA_MODEL env	qwen3-coder:30b	Default model when --provider ollama is selected
OLLAMA_TIMEOUT env	1800 (seconds)	HTTP timeout. Default 30 min absorbs CPU cold-starts + thorough reviews
--model local alias	qwen3-coder:30b	30B MoE coder, ~3.3B active params (quality/speed sweet spot on CPU)
--model local-pro alias	qwen3-coder-next	80B/3B MoE — higher quality, larger download

Why qwen3-coder:30b is the default

It is a Mixture-of-Experts model with 3.3B active parameters per token despite being 30B total. On CPU, active-param count drives inference speed, not total params — so this model runs roughly as fast as a 3B dense model while having the quality of a 30B model. For users with a GPU, the speed advantage is less pronounced but the quality benefit holds.

Why the temperature changed

Cross-model integration testing on the same 31 K-char diff (the diff of this PR, prior to commit) produced opposite failure modes:

• qwen3-coder:30b via local Ollama: 3 minutes, clean "no issues found." Likely missed legitimate nits but did not fabricate any.
• google/gemini-2.5-pro via OpenRouter at the previous temperature=0.5: 80 seconds, one HIGH-severity finding referencing args.timeout and --timeout flag. Neither the flag nor the attribute exist in the codebase. Verified by grep returning zero matches. The proposed fix would have crashed the runner with AttributeError: 'Namespace' object has no attribute 'timeout'.

The hallucination at 0.5 motivated reducing temperature to 0.3 — tighter than 0.5 to cut hallucination, looser than the original 0.2 to keep "more findings than too few." The DEFAULT_TEMPERATURE constant docstring carries the full history.

Typed error model for Ollama

The new call_ollama function distinguishes:

• Connection refused → ConfigError (exit 2) suggesting ollama serve and --ollama-host. No retry — config problem.
• HTTP 404 → ConfigError (exit 2) with the exact ollama pull <model> command. No retry — model not pulled.
• HTTP 429 / 413 / 5xx → routed through the shared _classify_http_error so rate-limit / overflow / transport are typed consistently with the cloud providers.
• Null content with finish_reason=length → ContextOverflow (exit 12). No SafetyRefusal branch — Ollama has no content-filter mode.

Test plan

• import review succeeds with all three providers and the new constants
• --help shows --provider {openrouter,gemini,ollama} and --ollama-host
• Smoke test: --provider ollama --base HEAD against a small diff with default model (qwen3-coder:30b) produces a clean review in under 1 minute
• Real review: --provider ollama --base HEAD against the entire 31 K-char diff of this PR returns a coherent summary + clean verdict in ~3 min
• Cross-provider review: same diff via --provider openrouter returns structured output (and surfaced the hallucination that prompted the temperature retune)
• Pull qwen3-coder-next and verify the local-pro alias dispatch (deferred — large download)
• CI integration with --provider ollama (deferred — needs Ollama in CI image)

Notes for reviewers

• The MODEL_ALIASES top-level dict is preserved as a backwards-compat view over MODEL_ALIASES_BY_PROVIDER["openrouter"] in case any external importer relied on it. If you're confident nothing does, the alias can be deleted in a follow-up.
• The 1800-second default Ollama timeout is intentionally generous. If you want faster failure on stuck inference, override with OLLAMA_TIMEOUT per env.
• The "Local vs cloud" section in both README and runbook documents the opposing failure modes empirically rather than asserting them — the evidence is the cross-model test in the PR's own diff.

🤖 Generated with Claude Code

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[Airwhale]

Apologies, this was opened against upstream by mistake : the changes are fork-specific to my local fork and accidentally opened the PR upstream. Thank you to the authors and maintainers of this useful tool! Been amazing to adjust it to my own workflow, local LLM's, etc...