feat(collector): forecast and ntfy in ping-collector; ThaiWater keeps alerts current at night

Phase 2b/2c of docs/RUST_REWRITE.md. With FORECAST_IN_COLLECTOR=true (shared
.env) ping-collector, after each cycle that brought newer RID or HII readings
and at least hourly: loads the last 336 h (RID + HII mirror fill), fetches the
rain forecast, runs every exported model (rust/forecast), stores
forecast_history and forecast_current, and sends the ntfy notifications.
The Python API then serves forecast_current as /forecast (computing in-process
only when it is older than FORECAST_MAX_AGE_SECONDS) and neither forecasts nor
notifies.

ntfy (collector/src/notify.rs): a port of src/notify.py with the same topics,
texts, hysteresis, capacity guard and notification_state keys, checked word
for word on 751 messages generated by Python (testdata/gen_notify_cases.py).
The change: for stations whose ThaiWater twin is the same sensor, a reading
newer than RID's drives the level alerts (marked as ThaiWater telemetry), and
the feed counts as stale only when neither source has a reading for 3 h. RID
alone lagging is reported once after RID_STALE_ALERT_HOURS (18). This ends the
nightly "gauge feed stale" / "recovered" pairs caused by RID's late reports.

Commands: `forecast [--json] [--save] [--now] [--rain-file]`,
`notify --dry-run`. scripts/forecast_parity.py compares Python and the
collector on the live database (same moment, same rain); locally 48/48 rows
identical. install_collector.sh: models export + that check on every install,
--cutover-forecast (refuses on a failed check) and --rollback-forecast.

Trees are packed into 24-byte nodes: 70 MB peak with all 15 models (was 140).
Forecast values are written to NUMERIC as shortest round-trip text, so they
round as psycopg2's do. DB test schemas get a counter (parallel tests collided
on Windows' coarse clock).

rust/forecast/src/hgb.rs

@@ -7,19 +7,23 @@
 use anyhow::{Result, bail};
 use serde::Deserialize;
 
-fn thresholds<'de, D: serde::Deserializer<'de>>(d: D) -> Result<Vec<f64>, D::Error> {
-    let raw: Vec<Option<f64>> = Deserialize::deserialize(d)?;
-    Ok(raw
-        .into_iter()
-        .map(|t| t.unwrap_or(f64::INFINITY))
-        .collect())
+/// One node, packed (24 bytes): `x` is the split threshold of an internal
+/// node and the output of a leaf.
+#[derive(Debug, Clone, Copy)]
+struct Node {
+    x: f64,
+    left: u32,
+    right: u32,
+    feature: u16,
+    leaf: bool,
+    missing_left: bool,
 }
 
-#[derive(Debug, Deserialize)]
-pub struct Tree {
+/// The exported arrays, as in sklearn's predictor node table.
+#[derive(Deserialize)]
+struct RawTree {
     feature: Vec<usize>,
-    #[serde(deserialize_with = "thresholds")]
-    threshold: Vec<f64>,
+    threshold: Vec<Option<f64>>,
     missing_left: Vec<bool>,
     left: Vec<usize>,
     right: Vec<usize>,
@@ -27,44 +31,81 @@ pub struct Tree {
     value: Vec<f64>,
 }
 
-impl Tree {
-    #[inline]
-    fn predict(&self, x: &[f64]) -> f64 {
-        let mut n = 0;
-        while !self.leaf[n] {
-            let v = x[self.feature[n]];
-            let go_left = if v.is_nan() {
-                self.missing_left[n]
-            } else {
-                v <= self.threshold[n]
-            };
-            n = if go_left { self.left[n] } else { self.right[n] };
-        }
-        self.value[n]
-    }
+#[derive(Debug)]
+pub struct Tree {
+    nodes: Vec<Node>,
+}
 
-    fn validate(&self, n_features: usize) -> Result<()> {
-        let n = self.value.len();
+impl<'de> Deserialize<'de> for Tree {
+    fn deserialize<D: serde::Deserializer<'de>>(d: D) -> Result<Self, D::Error> {
+        use serde::de::Error;
+        let r = RawTree::deserialize(d)?;
+        let n = r.value.len();
         if n == 0
             || [
-                self.feature.len(),
-                self.threshold.len(),
-                self.missing_left.len(),
-                self.left.len(),
-                self.right.len(),
-                self.leaf.len(),
+                r.feature.len(),
+                r.threshold.len(),
+                r.missing_left.len(),
+                r.left.len(),
+                r.right.len(),
+                r.leaf.len(),
             ]
             .iter()
             .any(|&l| l != n)
         {
-            bail!("tree arrays have different lengths");
+            return Err(D::Error::custom("tree arrays have different lengths"));
         }
-        for i in 0..n {
-            if !self.leaf[i]
-                && (self.feature[i] >= n_features || self.left[i] >= n || self.right[i] >= n)
-            {
-                bail!("tree node {i} points outside the tree or the feature vector");
-            }
+        let nodes = (0..n)
+            .map(|i| {
+                let leaf = r.leaf[i];
+                if !leaf && (r.left[i] >= n || r.right[i] >= n || r.feature[i] > u16::MAX as usize)
+                {
+                    return Err(D::Error::custom(format!(
+                        "tree node {i} points outside the tree"
+                    )));
+                }
+                Ok(Node {
+                    // +inf thresholds ("missing vs everything") are exported as null
+                    x: if leaf {
+                        r.value[i]
+                    } else {
+                        r.threshold[i].unwrap_or(f64::INFINITY)
+                    },
+                    left: r.left[i] as u32,
+                    right: r.right[i] as u32,
+                    feature: r.feature[i] as u16,
+                    leaf,
+                    missing_left: r.missing_left[i],
+                })
+            })
+            .collect::<Result<_, _>>()?;
+        Ok(Tree { nodes })
+    }
+}
+
+impl Tree {
+    #[inline]
+    fn predict(&self, x: &[f64]) -> f64 {
+        let mut node = &self.nodes[0];
+        while !node.leaf {
+            let v = x[node.feature as usize];
+            let go_left = if v.is_nan() {
+                node.missing_left
+            } else {
+                v <= node.x
+            };
+            node = &self.nodes[if go_left { node.left } else { node.right } as usize];
+        }
+        node.x
+    }
+
+    fn validate(&self, n_features: usize) -> Result<()> {
+        if let Some(i) = self
+            .nodes
+            .iter()
+            .position(|n| !n.leaf && n.feature as usize >= n_features)
+        {
+            bail!("tree node {i} reads feature outside the feature vector");
         }
         Ok(())
     }

rust/forecast/src/hii.rs

@@ -55,6 +55,8 @@ pub struct FillInfo {
     pub station_code: String,
     pub offset: f64,
     pub filled_hours: usize,
+    /// The newest filled hour and its level.
+    pub latest: Option<(NaiveDateTime, f64)>,
 }
 
 /// pandas `Series.median` (numpy: mean of the two middle values when even).
@@ -117,6 +119,11 @@ pub fn fill(rid: &[Row], hii: &[HiiRow]) -> (Vec<Row>, Vec<FillInfo>) {
                 station_code: code.to_string(),
                 offset,
                 filled_hours: fills.len() - before,
+                latest: fills
+                    .iter()
+                    .skip(before)
+                    .max_by_key(|r| r.timestamp)
+                    .map(|r| (r.timestamp, r.water_level)),
             });
         }
     }