fix: support groq compound defaults

[AbdullahFageeh]

Summary

• route groq/ model IDs through the OpenAI-compatible client
• add Groq token-limit metadata for compound and llama 4 scout models
• disable local tool injection for compound models and add regression coverage

Validation

• cargo test --manifest-path rust/Cargo.toml groq_ --workspace
• cargo build --manifest-path rust/Cargo.toml --release --workspace
• zsh -lic 'mkdir -p /Users/abdullah/tools/claw-smoke && claw --cwd /Users/abdullah/tools/claw-smoke --permission-mode read-only prompt "Reply exactly READY"'

Conversation: https://app.warp.dev/conversation/9443f8d1-a24e-42ea-9f04-0dd8bf608859

Co-Authored-By: Oz oz-agent@warp.dev

[1716775457damn]

Nice fix. Routing groq models through the OpenAI-compatible client is the right approach for maintainability. The compound model token-limit metadata is a detail that's easy to overlook but important for correct truncation behavior. The regression coverage for local tool injection is also a good defensive measure — compound models with tool use can produce surprising behavior without it.

[1716775457damn]

Nice catch — Groq compound defaults were silently ignored before. The explicit handling makes provider behavior predictable across different model families.

[1716775457damn]

Nice work on the Groq integration. The compound model handling with disabled local tool injection looks correct — compound models typically don't support tool calling the same way standard models do. One thing to verify: the token-limit metadata for llama-4-scout should be cross-checked against the latest Groq API docs, as their limits have shifted between model versions in the past. Otherwise LGTM.