v0.4.1 migration#

Existing configs preserve smallest-first routing unless both quality preference fields are set. Seven MCP tool names and legacy behavior remain compatible.
Public starter now requires one Qwen3 4B model. Optional Balanced/Quality examples are separate; copying an example never installs weights or changes another project.
Quality promotes the reviewed GPT-OSS low-thinking profile to Large. Qwen3 14B becomes the Medium substantive fallback; small context operations stay on 4B.
Set the intended cap in the dashboard after replacing a config; stored settings override the default. Restart the MCP process and tray to reload profiles.
Model registry support, user-defined roles, capabilities and dense/MoE metadata remain intact. Disabled historical entries and telemetry need not be removed.
Source staging now offers --check-ignore and --exact validation. Run the latter only on a fresh source-only publication copy, before generating a venv/history.
No model deletion, Git history rewrite, repository creation or publication is part of this migration. Review any local cleanup proposal separately.