perf(feel): adaptive chat drip, low-effort wiki thinking, lazy ExcelJS

Measured investigation found the backend near-idle (~2ms local TTFB); the slow
FEEL came from the frontend, not compute.

- useAssistantChat: replace the fixed 8 chars/tick (~500 c/s) drip ceiling with
  an adaptive drain (max(8, ceil(gap/12))) so a fast stream catches up within a
  few hundred ms. A ~4000-char answer now finishes painting in ~2s, not ~8s.
- Thread thinkingEffort through StreamChatParams/runLLMStream; default stays
  'high' for interactive chat, wiki builder now runs 'low' - no user watches the
  background thought stream across its up-to-40 tool iterations.
- exportToExcel: lazy import('exceljs') inside the export handler instead of a
  static top-level import, removing the ~270KB chunk from the tabular route's
  initial bundle.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>