feat: Workbench evidence preview and assistant navigation

[tiven-w]

Summary

• add a host-rendered file/PDF evidence preview with page-aware markers, PDF text-layer alignment, image-only fallback calibration, and localized English/zh-Hans UI
• add typed Workbench navigation commands so remote components can open assistant conversations in the ClawXpert right-side ChatKit instead of legacy chat routes
• harden background Assistant Task conversation/execution persistence and preserve RequestContext through tracing, with regression coverage for execution upsert and tracing context
• export additional shadcn UI primitives and a browser-safe Collaboration SDK client entry point

Why

Drawing review evidence previously relied on external PDF tabs or coarse normalized coordinates, so evidence markers could not be rendered or resumed reliably in the host. Remote component navigation also depended on string-only commands and could fall back to legacy conversation pages. Background assistant execution needed stable conversation/execution state and context propagation so created records remain in the correct user scope and ChatKit streams can be resumed.

Impact

• drawing review can open evidence directly in the host preview and keep markers aligned with the rendered PDF page
• assistant executions open in the existing right-side ChatKit and preserve stream/history access
• remote components gain typed navigation contracts and browser-safe collaboration imports
• preview labels, actions, loading states, and errors default to English and support zh-Hans

Validation

• corepack pnpm exec tsc -p apps/cloud/tsconfig.app.json --noEmit
• corepack pnpm nx build plugin-sdk
• targeted Assistant runtime, tracing, execution-upsert, navigation, and renderer regression tests are included
• translation JSON parsing and git diff --check
• local UI smoke test with a real PDF evidence record