v1.2.3 — six user-facing bugs found by the docs audit

Six user-facing bugs, found by checking every claim in the docs against the
code. None changes stored data.

Fixed

• get_session_timeline with tail returned the middle of long sessions.
  It read the first 5,000 events in ascending order and sliced their end, so
  on a longer session "how did it end" came back with events from the middle.
  It now reads newest-first under the same cap and returns them in ascending
  order. One session in the author's archive has 8,240 events.
• setup suggested a command that no longer exists. Its closing hint was
  longhand recap --days 7, removed at 1.0. It now suggests longhand status --days 7, the documented successor.
• setup said it installed the MCP server "for Claude Code".
  longhand mcp install writes Claude Desktop's config only. The step now
  says Claude Desktop, and prints the command that registers Longhand with
  Claude Code (claude mcp add longhand -s user -- longhand mcp-server), the
  same one doctor recommends.
• Recall and a tool description pointed Claude at a retired tool. Recall's
  three narrative footers and get_session_timeline's description said to
  use search_in_context, retired at 1.0. They now name search with
  session_id and context_events.
• The reconcile tool's schema said fix defaults to true. The handler
  has defaulted to a dry run since 1.0. The schema now agrees, so an agent
  that omits fix knows it gets a report, not a repair.
• --skip-analysis claimed to populate "SQLite only". Events are always
  stored and embedded; the flag skips the analysis pass (episodes, segments,
  and session/project vectors). The setup and ingest help now says so.

Internal

• Three new checks guard the text the program shows at runtime, the way the
  README checks guard the docs: every longhand <command> it suggests must be
  a registered command, and no tool description or recall footer may send
  Claude to a retired tool.
• The deeper findings from the same audit are filed as #96–#107.