Status of the official gguf → q4nx conversion tool announced in v0.9.26 — and supported path for fine-tuned Qwen3/Qwen3.5? #828

Hi — in the v0.9.26 release notes (Jan 2026), the "Runtime Restructure for Fine-Tuned Models" section said a gguf → q4nx conversion tool was "almost ready." We're now on v1.0.8+ and I can't find the official tool or docs. Could you clarify:

1. Did the official conversion tool ship (under what name/docs URL), or was the plan superseded?
2. Our use case: fine-tunes (merged LoRA/full SFT) of catalog architectures — text Qwen3 and Qwen3.5 at 0.6B–9B. If self-serve conversion is supported, which families/sizes, GGUF quant formats, and context lengths are in scope?
3. The community Q4NX_Converter + flm-add path (convert GGUF to .q4nx, register via model_list.json) currently works with the restructured runtime. Is that a configuration you consider supported/stable across releases, or best-effort?
4. Relatedly: can Lemonade register a user-converted q4nx package under the flm recipe, or is the flm recipe limited to catalog packages?
5. If self-serve isn't the intended route: is the FastFlow SDK "Customize" tier (IRON/HRX) meant for this, or is there a partner conversion program? Any timeline you can share?

Context: we run a small local fleet on Ryzen AI (XDNA2) and want our fine-tuned variants on the NPU seat; the stock catalog models already serve there via FLM. Happy to test pre-release tooling and report results. Thanks!