Refactor agent tool effects and centralize tool catalog - #17

Chart kinds, tool families and canonical-text projection were spread
across hand-written name lists. The durable projection kept canonical
text for only eight chart kinds, so taiyi, astrology and almanac
results fell back to raw JSON, and tool cards guessed icons from name
fragments.

- tool-catalog.ts is the browser-safe source for chart identity, tool
  families and canonical-text projection.
- The projection, chart tools, presentation helpers and tool cards read
  it; presentation types are keyed by the catalog's chart kinds, so a
  kind without a canonical JSON type fails to compile.
- Tests guard that every Core and registered Agent tool has a family
  and that all eleven chart kinds keep canonical text end to end.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GzFwaF8xWLhKF9YvLdD34m

- Keep Core canonical text on dayun, pillar resolution, ziwei horoscope
  and flying-star results so their views can fall back to it.
- remember_user_fact returns only the memory's id, title, content, kind
  and update time; ownership and fingerprint columns no longer reach the
  model or the transcript.
- Collaboration progress records each member's actual reasoning effort
  and configured focus.
- The public run view carries the settled billing state.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GzFwaF8xWLhKF9YvLdD34m

…boards

Replicate the "Taibu Agent 对话工作台" boards for the session header,
composer, empty state, model menu and question panel.

- Header: 56px bar with title menu (rename, export, delete), Agent and
  collaboration badge, run status with elapsed time or collaboration
  progress, workspace button with result count, and a more menu holding
  the reading width (default now 700px).
- Composer: single shell with "+", @, Agent/对话 switch, web toggle,
  context ring, model trigger and send/stop; state-dependent note,
  credit strip, saved-chart chips and suggestion cards on the hero.
- Model menu: tier labels from /api/models, segmented reasoning, style
  and mode controls, collaboration members as chips with one editor.
- Question dock replaces the modal dialog in the composer position:
  step navigation with short headers, review before submit, and
  "稍后回答" folding into a strip that hands the input back.
- Shared conversation export helper for header, /export and workspace.
- Tests: export helper, collaboration progress, question labels,
  composer copy, model tiers; browser fixture and live scripts follow
  the new dock, menu and labels.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GzFwaF8xWLhKF9YvLdD34m

Message: status pill icons per state, a folded process line that names
its steps, step rows with marker, summary, note, measured time and
family tag that open their view in the workspace, and a current-step
mark shared with the reading cards. Plan updates stay in the task list,
answered questions are not repeated, and reasoning alone shows as a
thinking line. Steps are numbered once across the task.

Readings become one-line cards (pillars or glyph, title, summary, a
workspace button that reads 正在工作区显示 while shown); outside the
workspace they expand in a dialog. Collaboration shows members side by
side with model, effort and state, and a member left running when the
task ended reads 未完成. The answer gets its own typography and caret,
本任务成果 lists saved readings and names the intermediate ones, and
evidence and message actions share one row.

Workspace process section: the step being looked at shows as 第 N 步
with its arguments and result to copy, its execution time apart from
waiting, and whether a replay reuses the cached result.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GzFwaF8xWLhKF9YvLdD34m

… suggestions

- Number evidence once per conversation: search results and chart focuses
  carry a server-stamped cite; answers render [n] as badges that open the
  cited source in the workspace and show which paragraph cited it.
- Rebuild the evidence view to the AgentResearch board: source filter,
  numbered cards, highlighted excerpts, empty-search notices.
- parallel_analysis accepts a title and up to 4 topics; members end with a
  <positions> block, the workspace shows live member steps, call counts,
  durations and a divergence table for option topics.
- The final answer may end with a <followups> block; it is stripped from the
  stored text and shown as suggestion chips above the composer.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GzFwaF8xWLhKF9YvLdD34m