C++ RAG integration and function calling

[Copilot]

Implements the two remaining near-term features in the C++ cognitive kernel: retrieval-augmented generation and LLM tool/function calling.

RAG (cpp/src/rag/)

• Self-contained vector store — no external embedding model required
• 1024-dim FNV-1a hash-projection feature space with L2-normalised term-frequency vectors and cosine similarity retrieval
• Configurable chunk size + overlap; file ingestion via rag_add_file()
• rag_build_context() returns a prompt-ready context block for direct injection before llm_chat_completion()

rag_store_t store = rag_create(NULL);
rag_add_file(store, "docs/manual.txt");
char* ctx = rag_build_context(store, "How do I configure X?", 4);
// prepend ctx to system prompt
free(ctx);
rag_destroy(store);

Function calling (cpp/src/function/)

• Tool registry with C-callback and popen-based shell-command backends
• Injects tool descriptions into the system prompt automatically
• Protocol: model emits <tool_call>{"name":"…","args":{…}}</tool_call>; dispatcher parses, invokes, and returns <tool_result>…</tool_result> for re-injection

function_registry_t reg = function_registry_create();
function_register_shell(reg, "calc", "Evaluate math", "echo $(({args}))");
char* tools_prompt = function_get_tools_prompt(reg);  // inject into system msg

char result[4096];
function_dispatch(reg, model_output, result, sizeof(result));
// inject result as next user message

Tests

12 new ctest cases (5 RAG, 7 function); total suite goes from 15 → 27 passing.