Add Groq agent tool cassette regressions

[gold-silver-copper]

Summary

Adds a broad cassette-backed Groq regression suite for Rig's OpenAI-compatible Groq chat-completions path.

New scenarios under tests/providers/groq/agent_tool_sessions.rs cover:

• long multi-turn sequential tool loops with five local tools;
• streaming/non-streaming parity for equivalent multi-tool sessions;
• multiple tool calls in a single assistant turn with parallel_tool_calls;
• raw streaming tool-call deltas and final object-valued tool calls;
• long caller-owned history replay with system/user/assistant/tool-result messages;
• ToolChoice::{Auto, Required, Specific, None} wire behavior;
• JSON object response-format roundtrip;
• native JSON-schema structured output roundtrip;
• low-latency text streaming termination and final usage accounting.

Models used

• llama-3.3-70b-versatile: primary stable Groq-hosted chat/tool model for tool loops, streaming, JSON object mode, usage, and metadata.
• openai/gpt-oss-20b: JSON-schema structured-output scenario because Groq rejects json_schema response format for llama-3.3-70b-versatile.

Inspiration references

Reviewed these implementation/test patterns in many_rigs/inspirations:

• vercel-ai-sdk/packages/groq/src/groq-chat-language-model.ts
• vercel-ai-sdk/packages/groq/src/groq-chat-language-model.test.ts
• vercel-ai-sdk/packages/groq/src/groq-prepare-tools.ts
• vercel-ai-sdk/packages/groq/src/groq-prepare-tools.test.ts
• pydantic-ai/pydantic_ai_slim/pydantic_ai/models/groq.py
• pydantic-ai/tests/models/test_groq.py
• pydantic-ai/tests/providers/test_groq.py
• langchain/libs/partners/groq/langchain_groq/chat_models.py
• langchain/libs/partners/groq/tests/integration_tests/test_chat_models.py

Those references motivated coverage for specific tool choice, Groq response_format, streaming x_groq.usage, reasoning-token usage details, and Groq timing metadata.

Bugs found and fixed

• Groq max_tokens was not forwarded in chat-completions requests.
• Groq ToolChoice::Specific incorrectly reused the generic OpenAI Chat Completions conversion, which rejected specific tool choices even though Groq supports a single function-choice object.
• Rig warned and ignored output_schema for Groq; Groq now maps Rig output schemas to response_format: { type: "json_schema", ... }.
• Groq streaming usage can arrive via x_groq.usage; Rig now reads that path in addition to terminal usage chunks.
• Groq/OpenAI-compatible raw usage now preserves completion_tokens, completion_tokens_details.reasoning_tokens, and Groq timing fields (queue_time, prompt_time, completion_time, total_time) and maps output token usage from completion_tokens when present.
• Boxed Copilot's chat raw-response enum variant to keep CI clippy green under -D warnings after the latest toolchain started flagging large_enum_variant.

Validation

Record-mode commands run with GROQ_API_KEY from the environment:

RIG_PROVIDER_TEST_MODE=record cargo test -p rig --all-features --test groq sequential_complex_tool_calls_nonstreaming -- --nocapture --test-threads=1
RIG_PROVIDER_TEST_MODE=record cargo test -p rig --all-features --test groq agent_tool_sessions -- --nocapture --test-threads=1
RIG_PROVIDER_TEST_MODE=record cargo test -p rig --all-features --test groq sequential_complex_tool_calls_streaming -- --nocapture --test-threads=1
RIG_PROVIDER_TEST_MODE=record cargo test -p rig --all-features --test groq tool_choice_auto_required_specific_and_none -- --nocapture --test-threads=1
RIG_PROVIDER_TEST_MODE=record cargo test -p rig --all-features --test groq long_history_replay_with_tool_result_continuation -- --nocapture --test-threads=1
RIG_PROVIDER_TEST_MODE=record cargo test -p rig --all-features --test groq json_schema_structured_output_roundtrip -- --nocapture --test-threads=1

Replay/safety/unit checks:

env -u GROQ_API_KEY cargo test -p rig --all-features --test groq agent_tool_sessions -- --nocapture --test-threads=1
cargo test -p rig --all-features --test groq cassette_safety -- --nocapture --test-threads=1
env -u GROQ_API_KEY cargo test -p rig --all-features --test groq groq -- --nocapture --test-threads=1
cargo test -p rig-core providers::groq -- --nocapture
cargo fmt --check
cargo clippy --all-targets --all-features -- -D warnings

cargo test was also attempted and reached the pre-existing unrelated OpenAI cassette mismatch in openai::cassette::permission_control::permission_control_streaming_example.

Manual cassette inspection

• Confirmed Groq cassette requests use the expected /openai/v1/chat/completions path and expected request bodies.
• Confirmed no API keys, bearer tokens, cookies, account identifiers, or rate-limit error payloads are present in tests/cassettes/groq/.
• Confirmed cassettes preserve Groq response metadata such as system_fingerprint, usage, x_groq, and timing fields where the live API surfaced them.
• Confirmed replay passes without GROQ_API_KEY.

Non-goals / known gaps

• Does not migrate the older ignored Groq smoke tests to cassettes.
• Does not cover Groq transcription or native browser-search/tooling.
• Does not record Groq model-rejection behavior beyond using openai/gpt-oss-20b for JSON-schema output after observing llama-3.3-70b-versatile reject that response format.