BontaFlowStack 0.5.0

• Added the question-presentation preference: prefer-panel (with chat fallback) or chat, using existing user/project/task preference storage.
• Conversational setup offers the choice after technical checks when no effective choice exists; unanswered questions do not block setup or save a default.
• All skills load effective preferences once per chat; bfs-plan-tune can change/reset the presentation and refresh it for subsequent questions.
• Shortened all 29 skills and synchronized descriptions with the catalog. Shared instructions define accepted basis, checkpoint, evidence and waiting; optional capability, guard and delivery details load through conditional references.
• Reduced combined instruction text from 11127 to 9921 words (10.8%), counting the relocated reference text. This measures text size, not runtime cost or latency.
• Clarified mode-specific memory and deployment steps, removed duplicate output requirements, and checked local reference links and section anchors.
• Fixed setup to read the guard reference before any technical checks, including core-only setup. Missing capabilities no longer suppress questions about required task inputs.
• Validation: 35 automated checks, 19 matched native A/B cases, two desktop setup chats with human panel/chat answers, and five focused sessions for the follow-up fixes. See native results and limits.
• Native guard enforcement, actual provider deployment and the remaining manual cases are not certified by those runs. After updating, start a new chat and review changed hook definitions through Codex's native trust interface when requested.

Release checks: 35/35 passed; source ZIP inventory, SHA-256, versions and freshly extracted entrypoints passed. Built from clean commit f75bbd5dd074d114ca67ddafcc82bd144807bc1e.