//! Bounded-memory online statistics: fixed histogram + Welford mean/variance.
//!
//! Unlike the original implementation (which pushed every sample into a
//! `Vec` and sorted at the end — ~30 MB/hour of RAM per metric at 1 kHz),
//! memory use here is constant, so multi-hour stability runs are safe.

pub struct OnlineStats {
    name: &'static str,
    /// Histogram bin width in ns.
    bin_ns: i64,
    /// Values below `low` or above `high` go to under/overflow counters but
    /// still feed min/max/mean/variance.
    low: i64,
    high: i64,
    bins: Vec<u64>,
    under: u64,
    over: u64,
    n: u64,
    min: i64,
    max: i64,
    mean: f64,
    m2: f64,
}

impl OnlineStats {
    /// Track values in `[low, high]` with `bin_ns` resolution (all in ns).
    pub fn new(name: &'static str, bin_ns: i64, low: i64, high: i64) -> Self {
        let n_bins = ((high - low) / bin_ns + 1).max(1) as usize;
        Self {
            name,
            bin_ns,
            low,
            high,
            bins: vec![0; n_bins],
            under: 0,
            over: 0,
            n: 0,
            min: i64::MAX,
            max: i64::MIN,
            mean: 0.0,
            m2: 0.0,
        }
    }

    pub fn push(&mut self, v: i64) {
        self.n += 1;
        if v < self.min {
            self.min = v;
        }
        if v > self.max {
            self.max = v;
        }
        // Welford
        let d = v as f64 - self.mean;
        self.mean += d / self.n as f64;
        self.m2 += d * (v as f64 - self.mean);

        if v < self.low {
            self.under += 1;
        } else if v > self.high {
            self.over += 1;
        } else {
            let idx = ((v - self.low) / self.bin_ns) as usize;
            let last = self.bins.len() - 1;
            self.bins[idx.min(last)] += 1;
        }
    }

    pub fn len(&self) -> u64 {
        self.n
    }

    fn percentile(&self, p: f64) -> i64 {
        if self.n == 0 {
            return 0;
        }
        let mut target = (self.n as f64 * p).ceil() as u64;
        target = target.max(1);
        // Mass below the histogram window counts first.
        if target <= self.under {
            return self.low;
        }
        target -= self.under;
        let mut acc = 0u64;
        for (i, &c) in self.bins.iter().enumerate() {
            acc += c;
            if acc >= target {
                return self.low + (i as i64 + 1) * self.bin_ns;
            }
        }
        self.high // fell into overflow region
    }

    pub fn report(&self, unit_div: i64, unit: &str) {
        if self.n == 0 {
            println!("{}: no samples", self.name);
            return;
        }
        let std = (self.m2 / self.n as f64).sqrt();
        let d = unit_div as f64;
        println!(
            "{}: n={} min={:.1} p50={:.1} p95={:.1} p99={:.1} p99.9={:.1} max={:.1} mean={:.2} std={:.2} ({}) under={} over={}",
            self.name,
            self.n,
            self.min as f64 / d,
            self.percentile(0.50) as f64 / d,
            self.percentile(0.95) as f64 / d,
            self.percentile(0.99) as f64 / d,
            self.percentile(0.999) as f64 / d,
            self.max as f64 / d,
            self.mean / d,
            std / d,
            unit,
            self.under,
            self.over,
        );
    }
}