feat: chart-stats scenarios: Gekisou off, every rank, the Perfect play

[nichinichisou0609]

chart-stats gains the play scenarios a page needs: Gekisou off (a solo live), a Gekisou live at any ranks, and the Perfect play for a Just rate. The format stays ournotes-deck.chart-stats/2: fields are only added; every existing number and check deck is unchanged.

Changes (src/chartstats.rs)

• offSeeds (per chart, one seed): the live with Gekisou off, as Free Live and Challenge Live play it: JudgementStream::theoretical_best (every note Perfect at its time), seed 0, LiveModel::new + run. Same shape as a seed without the range fields: seed, score, weights[kind][position], check. A kind whose conditions read the Gekisou state fails as Unsupported without Gekisou and gets null weights instead of failing the chart; the off check deck draws only from the other kinds, on its own generator. A chart with more than three fevers (unplayable with Gekisou) is measured here too.
• Ranks: ranges[].rankBonusPercents (ranks 1..5 of the mission pattern; rankBonusPercent stays = [0]) and seeds[].rangeWeights[kind][position][range], the range score each existing measurement run gains per unit of power and factor (no extra runs). The rank bonus is trunc(rangeScore * p / 100), a fixed score in the frame of the range end that changes no factor and no note score, and a later range holds it at both ends of its score, so:
  • score_r = score - sum_i rankBonus_i + sum_i trunc(rangeScore_i * p_i(r_i) / 100) exactly (SeedStats::score_at_ranks);
  • weights_r[kind][k] = weights[kind][k] + sum_i (p_i(r_i) - p_i(1)) / 100 * rangeWeights[kind][k][i], within 2 * ranges / POWER per unit of factor (SeedStats::weights_at_ranks).
  • seeds[].rankCheck {ranks, exact, predicted, bound}: the seed's check deck at random ranks through new_gekisou_external + queue_gekisou_rank_confirmation; a deviation beyond the bound fails the chart, like check.
  • Guards: a kind whose condition, release or reset group has condition 7012 (confirmed rank) gets null range weights; a chart where a range's bonus frame lies inside another range's score frames (frame(start_j) < frame(end_i) <= frame(end_j), i.e. overlapping ranges) gets rangeWeights: null.
• Perfect play: seeds[].scorePerfect and seeds[].ranges[].rangeScorePerfect, one more no-skill run per seed on the same play with every Just judged Perfect (Just is only enabled inside the Just-count ranges, so nothing else changes). The Perfect play's rank 1 bonus follows from rangeScorePerfect.
• model gains ranks, perfect, off; the existing keys are unchanged.
• READMEs and the CLI usage describe the scenarios.

Cost per chart: +1 run per seed (Perfect play), +1 per seed with ranges (rank check), +1 + kinds x positions + 1 for Gekisou off.

Tests

• ranks_follow_linearly_on_the_rank_confirmation_path: with a rank table that moves every rank, random ranks on every seed: the no-skill score from the statistics equals the external rank path exactly; a unit 2000 effect at every position is within 6 points (2 per range); real decks at another power are within the bound; rank 1 everywhere gives score and weights bit for bit.
• off_seeds_match_the_per_order_model: the Gekisou off no-skill score equals the per-order model's; 40 random decks (life-conditioned skills included) at random powers within the flooring bound of the per-order model.
• more_than_three_fevers_cannot_be_played: such a chart still has offSeeds, equal to the per-order model's no-skill score.
• kinds_on_the_rank_or_the_gekisou_state: a 7012 kind has no range weights, a 7005 kind has no Gekisou off weights and stays out of that check deck.
• The Perfect play: ranges before the Just-count range score the same, the Just-count range less; a chart without Just has scorePerfect == score, and Gekisou off equals Gekisou on without fevers.
• Unit test of the overlapping-range guard.

cargo fmt --check, cargo clippy --all-targets -- -D warnings, cargo test and cargo test --release: 242 passed, 0 failed (6 ignored, as before).

Full TW data through nnnotes music-data (MetaSekaiLab/nnnotes#5, pinned at this commit; 340 charts, 1376 seeds): every chart measured, every check, off check and rank check within its bound; every old value equal to the file of 7e5d84b; at rank 1 everywhere the rank formulas give every seed's score and weights bit for bit; no kind or chart needed a guard. Expert, no skills: Gekisou off / on median 0.417 (0.211 to 0.633), all rank 5 / all rank 1 median 0.685 (0.610 to 0.782).

Generated with Claude Code