Coding Wrapped 0.2.0

• Counted Claude Code facts accurately: injected rows (meta, compact
  summaries, local command output, interrupts, tool results) no longer count
  as prompts, and model turns, tokens, tool calls, and prompts repeated across
  content-block rows or resumed session files count once.
• Classified tools by exact Claude Code and Codex names before falling back to
  whole-word matching, so to-do and planning tools no longer count as edits or
  subagents, and MCP tools are judged by their tool name.
• Added synthetic evals for both cases; the portable-copy eval now ignores
  local node_modules.
• Kept Codex-injected context (IDE file context, retained instructions,
  AGENTS.md, browser context, approval payloads, history replays) out of prompt
  statistics.
• Replaced the Dashboard's template-only insight generator: Generate now copies
  a request for the agent, and the page picks up new batches on focus. Removed
  POST /api/insights/generate.
• Fallback illustrations follow each card's theme. Persistence requires four
  distinct themes and evidence that resolves to real metrics with matching
  values.
• Skipped session files last written before a range's cutoff, and stopped
  asking the agent to load the full practice catalog.
• Replaced the TTF fonts with WOFF2, including a GB2312 subset of Noto Sans SC
  (2 MB instead of 17.8 MB); the Skill folder shrinks from about 28 MB to 12 MB.
• Served index.html without caching so an upgraded Dashboard loads its new
  bundle immediately.