usersim eval scores a repeated trajectory_id once per row, silently #40

What problem are you trying to solve?

One trajectory_id stands for one conversation. A writer that keeps its own ids, such as an extension
running a multi-session benchmark, can add a second row for a conversation, for example after
restarting a crashed session. Nothing flags it:

• usersim eval evaluates every row of the run (cli/evaluate.py), so that conversation counts
  twice in the scores.
• usersim select keeps both rows of one run, each joined to the first row's evaluation
  (selection/engine.py:354), and across several runs keeps the first row per id
  (selection/io.py:104), which may be the crashed attempt.

What would you like to happen?

The smaller scope agreed when this was first raised:

• usersim eval warns when a trajectory_id appears more than once in the run it evaluates, naming
  the ids. Rows are still evaluated as they are.
• The docs state the convention: an id is derived from what the conversation stands for (for a
  multi-session benchmark: member, arc and session), not from the attempt, and a restarted
  conversation's row replaces the earlier one.

Which part of the project?

The evaluator CLI and the output docs.