[Yaphet2015]

Implements spec v2 (goal + card weights), keeping the canvas almost unchanged and putting the goal experience in the Agent panel.

What changed

• Onboarding: ordinary chat questions (3–5, one at a time, with tappable chips) collect deliverable, success criteria, constraints and experience into a structured plan.goal; the last message shows a goal summary and the existing "开始第 1 站" button.
• Cards: a 来源 · 截至 YYYY-MM line under the body; station progress dots (filled = mastered, ring = read, none = not started).
• Weights: background / core / hard cards look the same and differ only in the footer (「懂了」 / choice question / question + input). A wrong answer promotes one level; the new footer fades in over 240ms.
• Agent panel: goal pin row under the header (editable by saying 「改目标…」, tint sweep on change). After a station, numeric gaps are calculated in code from the model's measured values (no model-written numbers), followed by a focus route diff and a collapsible 「我的产出物 · n/5 块」 attachment that is expanded only on the newest message.
• Copy cleanup: removed the 「4 拍」/「第 N 拍」 labels.
• Older canvases load with default values.

Verification (Linux box)

• tsc (renderer + electron) pass
• Vitest: 50 files / 467 tests pass
• vite build pass
• Playwright: goal-weights, chat, kickoff, flow: 26/26 pass, with light/dark screenshots

Not verified: real LLM calls, Electron packaging and macOS-only behavior.