fix: prevent server crash when clients send 'tools' param (fixes #530)

[octo-patch]

Fixes #530

Problem

When clients like Open WebUI send a POST /v1/chat/completions request with a tools parameter (standard OpenAI function-calling API), the llama-server binary crashes with SIGABRT instead of returning a proper error response:

terminate called after throwing an instance of 'std::runtime_error'
  what():  Unsupported param: tools
Error occurred while running command: ... died with <Signals.SIGABRT: 6>

Root Cause

In examples/server/server.cpp, the handle_chat_completions function calls oaicompat_completion_params_parse(), which throws std::runtime_error for unsupported parameters (tools, tool_choice). This exception is not caught locally, so it propagates past httplib's global exception handler and calls std::terminate(), crashing the server process.

Solution

Wrap the oaicompat_completion_params_parse() call in a try-catch block that:

1. Catches any std::exception thrown during parameter parsing
2. Returns an HTTP 400 (invalid request) error response with the error message
3. Returns early, keeping the server running for subsequent requests

The change is minimal (7 lines in examples/server/server.cpp) and does not affect normal request handling.

Changes

• Updated 3rdparty/llama.cpp submodule to octo-patch/llama.cpp at commit 7525084 which adds the try-catch fix in examples/server/server.cpp

Testing

Verified that the fix pattern matches how other error conditions are handled in the same file (e.g., the --embeddings check at the top of handle_chat_completions uses the same res_error + return pattern).

After this fix, sending a request with tools will return HTTP 400 with a clear error message instead of crashing the server.