Tool contract 4.1.0

The 3.13.0 guards, landed on the field that needed them (feedback #61, #62, #63, the 2026-09-19 regression run: every 3.13.0 item held, and three new defects were a guard sitting beside the number rather than on it).

• battles_performance (group_by: week) and battles_trends weeks carry trophy_mode_battles (the Trophy Road and Path of Legends battles played) beside trophy_battles, which counts only the ones the game reported a trophy delta for. A loss standing on an arena floor reports none, so trophy_battles was lower than the ladder games played, only losses dropped out, and a rate over it flattered exactly the floored player the response had already named. The docs now list it among the denominators, and a note names the weeks where the two counts differ.
• battles_levels lists EVERY step of monthly_trend where the modal arena changes or the mean starting trophies move by 200 or more under player.population_changes[] ({from_month, to_month, arena_changed, from_arena, to_arena, from_trophies, to_trophies, trophy_delta}; empty when the population held), and the note names all of them. It named the earliest step only and stayed quiet on a later arena crossing, the harder confound and the one #55 was filed for.
• battles_decks carries excluded {duels, no_deck}: a duel has no single deck and was silently outside total_battles_in_window and every share_of_battles (the docs listed the duel-excluding tools and left this one out). The identity total_battles_in_window + excluded.duels + excluded.no_deck = battles_performance.battles now holds over the same window, total_battles_in_window is summed over every deck-bearing battle rather than the first hundred decks, and a note says what sits outside the rows whenever the window held a duel.

Additive; no migration.