Compose conversation cards by payload

[VolkerChristian]

Summary\n\n- compose conversation cards directly from their typed payload variants instead of maintaining parallel card-kind construction, update, and comparison switches\n- colocate each payload's construction and update logic while retaining the existing shared ConversationCard ownership and folding behavior\n- reconcile user-message image thumbnails by path so text-only updates retain unchanged thumbnails\n- preserve generated-image reload behavior because an image may become available later at the same path\n\n## Measurements\n\nRelease build, median of five representative runs:\n\n| Scenario | Before | After |\n| --- | ---: | ---: |\n| Mixed card construction | 302.625 us | 303.405 us |\n| Agent-message update | 12.6386 us | 12.6158 us |\n| 80-card no-op reconciliation | 2.39619 us | 2.39586 us |\n| One changed card among 80 | 2310.98 us | 2308.48 us |\n| Text update with unchanged image | 7613.13 us | 11.5443 us |\n\nThe unchanged-thumbnail path is approximately 660x faster. The remaining differences are measurement noise.\n\n## Simplification\n\n- removes three parallel decision trees in favor of payload-directed composition\n- production code is net 12 lines smaller\n- adds no controllers, repositories, caches, event systems, or duplicated application state\n- keeps card identity, folding, command output, generated-image, and LocalPrompt-to-UserMessage behavior intact\n\n## Tests\n\n- focused conversation-card tests passed after each commit\n- complete suite: 7/7 passed, 0 failures\n\nFocused coverage includes unchanged image retention, path replacement, deletion, recreation, and unavailable-image placeholders.\n\n## Scope\n\nThis PR intentionally excludes:\n\n- PR8: information/approval orange-card behavior and placement\n- PR9: background-activity spinner/notification\n- PR10: intermittent first-response card overlap/layout settlement\n- protocol changes\n- child-thread ownership or hierarchy changes\n- unrelated styling and features\n