Fix demo screens: evidence panel, answer wording, counter counts, sidebar new project and project restore

[hyperdrivebeep]

Four screen and project fixes found while recording the demo.

• Answer evidence opens again. Conversation reads return a summary view of a result, and the evidence reader compared that summary with the full stored result, so "근거 원문" always reported a mismatch. It now accepts a shown result whose fields all match the stored one; a real server/manifest mismatch is still reported. (Same change as Open answer evidence when the screen holds the summary view of a result #8, which this replaces.)
• Answer wording no longer contradicts the trace table. The answer's own check list is called "답변 점검 항목" instead of "평가기준", so "10개 중 충족 0" no longer reads against "기준 6개 · 충족 2". The hypothesis card shows "아직 대조 안 한 근거 63/66" instead of a bare "미평가 63"; expected results show each hypothesis's statement, and "근거 없는 예측" becomes "자료로 확인되지 않은 예상". Projects whose trace carries the synthetic-data notice show a one-line notice above the conversation.
• Counter-evidence counts are not hidden. When a hypothesis has counter sources, the card and the answer line show the count ("반박 3") even if no separate counter search was recorded; "-" stays only when there are none and no search ran.
• Sidebar "새 프로젝트" works, and archived projects can be restored. The sidebar button now opens the create form from any screen. New command project/restore returns an archived project to the lifecycle recorded before the archive (or DRAFT/ACTIVE by the activation rule when the record cannot say), records a new revision, and copies no data. The web sidebar gains "보관한 프로젝트 보기" with a restore button, and the archive confirmation now says the project can be restored.

Contract changes: project/restore added to the public method catalog (368→369), atomicity paths and registration fixtures updated; notifications map to project/statusChanged.

Validation (clean checkout of this branch): ruff, architecture, module budget, protocol parity, atomicity and source manifest checks pass; web tests, lint, typecheck, build, README quick-start smoke and the full Python suite are run locally, with results added to this PR when they finish. No model calls were made.