Report benchmark results

Arguments: $ARGUMENTS.

1. Run crucible report $ARGUMENTS from the repo. With a task id it compares every configuration for that task; without one it reports across tasks, grouped by task kind, with a recommendation per kind. Flags: --batch, --json, --kind, --min-attempts, --threshold.
2. If it says there are no results, tell the user to run /crucible:run <task-id> first. Do not start a run yourself; it spends usage.
3. Summarise in plain language, not by pasting the table alone:
   • the recommendation and the evidence behind it (success rate with its 95% interval, cost per success, median time), and on how many tasks it was compared (the cross-task report prints Compared on N of M tasks attempted by every config);
   • the runner-up, and the trade-off against it;
   • caveats: few attempts or a wide interval, Unknown results, Stopped attempts (a limit ended them), Judge≠check disagreements, unpriced attempts, a warning that a config pooled several models, or a warning that the task changed between batches. When the report says "compared on N of M tasks" with N below M, or "no common task", relay it. If no config meets the bar, say so and name the best candidates instead of recommending one. If it prints Recommendation: none; no task was attempted by every config, say there is no common task to compare on and suggest running the same task with every config; do not pick a winner.