Round-2 QLoRA tool-use fine-tune of swiss-ai/Apertus-v1.1-4B-Instruct, quantized to Q5_K_M for llama.cpp (converted from f16, never quant→quant).

Round-2 changes vs round 1: direct weather/fetch tool selection (no forced chaining), notes answers stay grounded in note hits, date anchoring (Current date: <Weekday>, <YYYY-MM-DD> as the last system-prompt line — the runtime must inject it), and a higher share of clarifying questions.

Intended to run with its paired system prompt and tools.gbnf grammar (lazy-triggered on <tool_call>); sampling temp 0.8, top_p 0.9, repeat_penalty 1.0, context 4096.