Add OpenRouter agent tool cassette regressions

[gold-silver-copper]

Summary

• Adds a new OpenRouter cassette module, agent_tool_sessions, with 7 live-recorded regression scenarios for serious multi-turn/tool use.
• Covers sequential non-streaming and streaming agent loops with four ordered tools and complex arguments: empty objects, nested JSON, arrays, escaped strings, backslashes, newlines, and unicode.
• Covers multiple tool calls in one model turn for non-streaming and streaming, raw streamed tool-call deltas, long caller-owned chat-history replay with assistant tool calls/tool results, tool_choice required/none behavior, usage/finish reason/model preservation, routed provider preferences, and nested JSON-schema structured output.

Models/providers used

• openai/gpt-4o via OpenRouter/OpenAI for stable tool calling, multi-tool turns, streaming tool-call deltas, and long chat-history replay.
• google/gemini-2.5-flash via OpenRouter/Google AI Studio preference for inexpensive nested JSON-schema structured-output coverage with provider.require_parameters.

Inspiration references studied

• tests/README.md, existing tests/providers/openrouter/cassette/*, tests/cassettes/openrouter/*, crates/rig-core/src/providers/openrouter/*, and OpenAI-compatible code under crates/rig-core/src/providers/openai/*.
• tests/providers/openai/cassette/responses_sessions.rs for long-session/multi-tool regression shape.
• /Users/kisaczka/Desktop/code/many_rigs/inspirations/vercel-ai-sdk/packages/openai-compatible/src/chat/convert-to-openai-compatible-chat-messages.test.ts and packages/openai/src/chat/openai-chat-language-model.ts for OpenAI-compatible tool message/result, streamed tool delta, usage, and finish-reason handling.
• /Users/kisaczka/Desktop/code/many_rigs/inspirations/vercel-ai-sdk/packages/openai/src/responses/openai-responses-language-model.ts for Responses event/usage/tool continuation patterns.
• /Users/kisaczka/Desktop/code/many_rigs/inspirations/pydantic-ai/tests/models/test_openrouter.py and tests/providers/test_openrouter.py for OpenRouter provider options, reasoning/tool preservation, nested schemas, usage/provider details, and media/document edge cases.
• /Users/kisaczka/Desktop/code/many_rigs/inspirations/langchain/libs/partners/openrouter/tests/integration_tests/test_chat_models.py and test_standard.py for OpenRouter tool calling, structured output, multimodal support, usage, and multi-turn reasoning preservation.
• /Users/kisaczka/Desktop/code/many_rigs/inspirations/semantic-kernel/dotnet/src/IntegrationTests/Data/BaseTextSearchTests.cs for required function/tool-choice behavior in agent loops.

Bugs found/fixed

• No provider implementation bug was found while recording these cassettes.
• Test-side replay note: caller-owned historical tool-call IDs use the cassette redacted placeholder form so the existing cassette scrubber can replay generated-looking OpenAI tool-call IDs deterministically.

Recording/replay commands run

• Record mode, targeted new suite/tests using exported OPENROUTER_API_KEY only through RIG_PROVIDER_TEST_MODE=record, including:
  • RIG_PROVIDER_TEST_MODE=record cargo test -p rig --all-features --test openrouter agent_tool_sessions -- --nocapture --test-threads=1
• Replay/no-credential validation:
  • env -u OPENROUTER_API_KEY cargo test -p rig --all-features --test openrouter agent_tool_sessions -- --nocapture --test-threads=1
  • env -u OPENROUTER_API_KEY cargo test -p rig --all-features --test openrouter openrouter::cassette -- --nocapture --test-threads=1
  • env -u OPENROUTER_API_KEY cargo test -p rig --all-features --test openrouter cassette_safety -- --nocapture --test-threads=1
• Other validation:
  • cargo fmt --check
  • cargo clippy --all-targets --all-features
  • cargo test -p rig-core providers::openrouter -- --nocapture

Manual cassette inspection notes

• Inspected new YAMLs under tests/cassettes/openrouter/agent_tool_sessions/ for expected /chat/completions requests, parallel_tool_calls: false, tool_choice: required/none, streamed tool-call deltas, response usage, finish_reason/native_finish_reason, routed provider/model metadata, and nested response_format schema payloads.
• Ran cassette safety and an additional grep for API keys, bearer tokens, auth/cookie headers, and account/project-id shaped fields; no secrets or unrelated churn found.

Remaining known gaps / non-goals

• No new OpenRouter Responses API cassette beyond existing OpenAI Responses compatibility coverage; this PR focuses on Rig's OpenRouter chat-completions path.
• No new multimodal cassette in this module because existing OpenRouter cassette coverage already exercises image/audio/video/PDF paths.
• Broad cargo test without feature flags fails on an existing unrelated Gemini image-generation feature-gating compile issue (rig::image_generation and GEMINI_2_5_FLASH_IMAGE are gated behind the image feature); targeted OpenRouter and all-features clippy validations passed.