Fix what the reviews of questions, tickets and the team found

[danielehrhardt]

Questions and approvals

• An answer counts as read the moment Claude starts replying, not when the
  run ends: a crash later can no longer hand an approval over twice.
• Stopping the run that was to read an answer drops it, so no later turn
  is told "Do it now" for a step the human stopped (queued messages that
  take over still carry it).
• A continued run whose own note already carries the answer doesn't get a
  second copy; a lost session's retry keeps it.
• A reply sent while the agent is still asking waits a moment and counts
  as the answer (chat, ticket, Slack/Telegram/Teams owner).
• A follow-up in a platform chat that asks a question posts its answer once.
• An automation's reminder replaces the earlier notice, so answering
  settles it.

Tickets

• A manager's note no longer leaves a waiting ticket in progress for good;
  its notes count per ticket and belong to no run of the ticket.
• Cancelling a waiting coding ticket's follow-up delivers it like a run's
  end: push and pull request.
• A follow-up an earlier run set is cancelled once a later run finished
  the work, instead of keeping the ticket waiting.
• Deleting a waiting ticket no longer delivers it on the way out.
• Ticket cost counts each run's own cost: runs from before the
  session-cost fix only count what they added to the chat's session.
  Totals are added up again once, and after a restore (with who filed the
  ticket and why it is blocked).
• The reply box drops its "Sending…" row once sent; the phone stops
  saying "Working on it" for tickets that wait.
• "Open the pull request" is recorded once, not every round.

Team

• The handoff card follows exactly the run its result names.
• A continued run that waits for a free slot stays "queued"; Activity's
  live list and follow-ups count running runs only.
• Reporting loops are found however long the line.
• "On its plate" shows waits for Claude's usage limit (runs list their
  pause).
• Any stop by the human clears "Last run failed" (chat, board, platform,
  a paused run).
• An agent can't set a lead that reads secrets or controls the computer
  when it couldn't hand that agent work itself; agents aren't steered to
  hand decisions to a lead that manages agents.
• A sender name read from a message is shown as another agent's, never
  as the human.

Checks

• pnpm typecheck: all packages pass
• Core tests: 1338 pass, 10 skip, 0 fail (with main merged in)
• Desktop build and phone type check pass
• New regression tests: questions (stop drops an unread approval; answer settled at the first reply), tickets (manager note on a waiting ticket, deleting a waiting ticket, session-total costs), team (loops longer than 16, a secret-reading agent as lead)