fix: prevent abort in local inference

[KubeCat]

Fix a bug "Rust cannot catch foreign exceptions" abort that occurs when using Local Inference with models that have long Jinja2 chat templates (e.g. Qwen3 GGUF models).

The problem

1. apply_chat_template() calls llama_chat_apply_template() which internally does a C++ map .at() lookup on the template string.
2. When the string is a full Jinja2 template (not a short name like "chatml"), this throws std::out_of_range.
3. The exception is normally caught in C++, but when goose's other native dependencies are linked into the binary, the C++ exception-handling ABI breaks and the exception propagates across the Rust FFI boundary, causing a process abort.

The solution

Switch from apply_chat_template() to apply_chat_template_with_tools_oaicompat() in inference_emulated_tools.rs

Why?

This function has its own C++ try-catch wrapper. When called with tools=None, it produces the same prompt. This aligns the emulated tools path with the native tools path, which already uses the oaicompat variant.

Testing

A few integration tests were improved, the integration tests for the local inference had been abandoned and marked with #[ignore], this is problematic because I could not understand very clearly how to actually test my changes. The changes in testing are as follows:

1. Fix broken model ID in local_inference_integration.rs, add TEST_MODEL env var override, removed perf assertion since it would never pass because the model was never truly in a cold start state.
2. Separate cold vs warm perf benchmark with TEST_MODEL env var support in local_inference_perf.rs

Type of Change

• Feature
• Bug fix
• Refactor / Code quality
• Performance improvement
• Documentation
• Tests
• Security fix
• Build / Release
• Other (specify below)

AI Assistance

• This PR was created or reviewed with AI assistance
• 56 unit tests pass (parsing, engine, hf_models)
• Integration tests pass with Llama 1B and Qwen3 32B, on both CPU and CUDA
• Qwen3 32B was the crashing model now works

Testing

• Tested end-to-end with bartowski/Qwen_Qwen3-32B-GGUF:Q4_K_M via test_provider_configuration —
  previously aborted, now passes.
• Verified the crash is specific to the goose binary (standalone llama-cpp-2 tests pass because the
  C++ ABI isn't disturbed by additional native deps).
• Verified that short template names ("chatml") never triggered the crash — only full Jinja2
  template strings.
• Existing unit tests for emulated tools parsing, tool parsing, and inference engine all continue to
  pass (they test downstream of the template step).

Related Issues

• Related to bug(windows): MSVC link failure in goosed when v8 and llama-cpp are linked together (LNK2038/LNK2005/LNK1169) #7410 same root cause (C++ exception-handling ABI clash between llama-cpp and other native deps). That issue manifests as a Windows MSVC link failure; this one manifests as a Linux runtime abort.
• Follow-up to Fix Windows MSVC linking issues #7511 which fixed the Windows MSVC link-time side of the v8/llama-cpp C++ ABI conflict (/FORCE:MULTIPLE for duplicate std::exception_ptr symbols). This PR addresses the Linux runtime side where the exception links but doesn't propagate correctly across the Rust FFI boundary.
• Discovered an issue where CUDA will crash when context_size is null in registry which falls through to n_ctx_train, despite the fact that estimate_max_context_for_memory should cap it.

Screenshots/Demos (for UX changes)

After: Model loads and generates normally.

With CUDA:

[jh-block]

Thank you for this!