feat(rid): screen one-hour spikes; refetch and redo-from to repair the store

RID's report sometimes carries a wrong value for one hour and corrects it
later. P.1 read 10.6 m at 2026-10-25 00:00 between 1.04 and 1.07 (RID now
says 1.06); the store kept it, because a report is re-fetched only while it
has holes, and the model issued a 10.8 m forecast for that hour.

- rid::judge / rid::screen: a reading that leaves the previous one (within
  3 h) by more than 1 m (2 m on the flashy tributaries P.76, P.77, P.81,
  P.82, P.84, P.85, P.87) while the next reading comes back near the
  previous level is a spike: not stored, a stored copy is deleted. Limbs
  and steps the next reading confirms are kept; the newest reading waits
  for the next one when it jumps more than twice the limit. On 2018-2026
  history: 16 spikes found (slipped decimals, sign flips, -5/-6/-9 m
  placeholders), no genuine change rejected.
- Every store goes through it (rid::store: run_cycle, fill-gaps, refetch).
- `ping-collector refetch FROM [TO]`: re-fetch reports regardless of holes
  (corrects revised readings, fills the discharge of level-only evening
  hours, removes stored spikes) and print the catch-up command when
  readings changed.
- `catch-up-forecasts --redo-from T`: delete the model forecasts issued
  from T on, then recompute them.
- Tests: judge/screen unit tests; a DB test of refetch, the live hold and
  the forecast deletion. Checked end to end against RID on a scratch DB.
- Docs: DATA_SOURCES (spikes, level-only hours), FLOOD_FORECASTING,
  rust/README.

docs/DATA_SOURCES.md

@@ -23,10 +23,29 @@ The only source persisted to the database and used by the ML pipeline.
 | Params | `DW[UtokID]=1`, `DW[BasinID]=6` (Ping), `DW[TimeCurrent]=<Buddhist-calendar date>`, `rows=100` |
 | Variables | Water level (m, gauge datum), discharge (m³/s, `'***'` = malformed), discharge % of channel capacity |
 | Stations | 16 P-series gauges (P.1 anchor at Nawarat Bridge; see `src/data/stations.json`) |
-| Client | `rust/collector/src/rid.rs` (`ping-collector`; `fill-gaps` re-fetches old reports) |
+| Client | `rust/collector/src/rid.rs` (`ping-collector`; `fill-gaps` re-fetches old reports with holes, `refetch FROM [TO]` any reports) |
 
 Human-facing page: <https://hyd-app-db.rid.go.th/hydro1h.html>
 
+**Spikes.** The report sometimes carries a wrong value for one hour and corrects
+it later: a slipped decimal point (P.1 10.6 m between 1.04 and 1.07 on 2026-10-25
+00:00, later 1.06; P.85 0.101 for 1.1), a flipped sign (P.4A ±1.18, P.67 ±0.6) or
+a placeholder (−5, −6, −9, 50 m). The collector re-fetches a report only while it
+has holes, so a later correction never arrived. Since 2026-10-26 every batch is
+screened before it is stored (`rid::judge`, `rid::screen`): a reading that leaves
+the previous one (within 3 h) by more than 1 m (2 m on the flashy tributaries
+P.76, P.77, P.81, P.82, P.84, P.85, P.87) while the next reading comes back near
+the previous level is not stored, and a stored copy is deleted. A rising or falling
+limb, or a step the next reading confirms, is kept. The newest reading of a gauge
+waits for the next one when it jumps more than twice the limit. On the 2018–2026
+history the rule finds 16 spikes and would have rejected no genuine change (the
+largest genuine 1-h changes: 1.1 m on the mainstem, 3.1 m at P.76). Two wrong hours
+in a row are not caught.
+
+Evening hours often arrive level-only (no discharge) and gain the discharge
+later; the API's history fills a missing discharge with 20 × level (an old
+placeholder, kept for compatibility). `refetch` brings in RID's discharge.
+
 ---
 
 ## 2. ThaiWater / HII ecosystem

docs/FLOOD_FORECASTING.md

@@ -643,7 +643,10 @@ Open-Meteo's estimate of rain that has since fallen, as in training, where
 the forward rain is observed rain. Such rows are written only where no model
 row exists; `created_at` tells them apart (more than an hour after `as_of`).
 `ping-collector catch-up-forecasts --hours N` does the same further back (up
-to 60 days).
+to 60 days); with `--redo-from T` it first deletes the model forecasts issued
+from T on, to recompute them after `ping-collector refetch` corrected the
+readings they came from (a bad reading reaches the features for 72 h through
+`roll_max_72`).
 
 A model row in `forecast_history` is never replaced by a heuristic one. While
 RID is silent overnight each hourly run re-issues the 21:00 reading, and once

rust/README.md

@@ -22,12 +22,17 @@ ping-collector serve [--listen A] [--no-analytics]
 ping-collector once [--dry-run]     # one cycle of every source; --dry-run writes nothing
                                     # and compares the RID fetch with the database
 ping-collector fill-gaps <days|all> # re-fetch RID reports with holes
+ping-collector refetch FROM [TO]    # re-fetch the RID reports of FROM..=TO (YYYY-MM-DD) even
+                                    # without holes, through the spike screening: corrects
+                                    # readings RID revised, deletes stored spikes; prints the
+                                    # catch-up-forecasts command when readings changed
 ping-collector check-db             # connect, create missing tables, row counts
 ping-collector forecast [--json] [--save] [--now T] [--rain-file F]
                                     # forecast from the database and print it
-ping-collector catch-up-forecasts [--hours N]
+ping-collector catch-up-forecasts [--hours N] [--redo-from T]
                                     # forecast the hours whose readings came late (RID's
-                                    # overnight batch) as a run then would have; default 336 h
+                                    # overnight batch) as a run then would have; default 336 h;
+                                    # --redo-from first deletes the model forecasts from T on
 ping-collector notify --dry-run     # the readings the notifier judges, and what it would send
 ```
 

rust/collector/src/db.rs

@@ -237,6 +237,62 @@ pub async fn save_rid(pool: &PgPool, readings: &[Reading], stations: &Stations)
     Ok(rows.len())
 }
 
+/// Stored RID levels of report columns `nums` in [start, end], for the
+/// screening (rid::screen).
+pub async fn rid_readings_between(
+    pool: &PgPool,
+    nums: &[u32],
+    start: NaiveDateTime,
+    end: NaiveDateTime,
+) -> Result<Vec<crate::rid::StoredReading>> {
+    let rows: Vec<(NaiveDateTime, i32, f64, Option<f64>)> = sqlx::query_as(
+        "SELECT timestamp, station_id, water_level::float8, discharge::float8
+         FROM water_measurements
+         WHERE station_id = ANY($1) AND timestamp >= $2 AND timestamp <= $3
+           AND water_level IS NOT NULL",
+    )
+    .bind(nums.iter().map(|&n| n as i32).collect::<Vec<_>>())
+    .bind(start)
+    .bind(end)
+    .fetch_all(pool)
+    .await
+    .context("reading stored RID levels")?;
+    Ok(rows
+        .into_iter()
+        .map(|(t, n, l, q)| (t, n as u32, l, q))
+        .collect())
+}
+
+/// Delete stored RID readings (timestamp, report column).
+pub async fn delete_rid_readings(pool: &PgPool, keys: &[(NaiveDateTime, u32)]) -> Result<usize> {
+    if keys.is_empty() {
+        return Ok(0);
+    }
+    let r = sqlx::query(
+        "DELETE FROM water_measurements w
+         USING UNNEST($1::timestamp[], $2::int4[]) AS k(ts, sid)
+         WHERE w.timestamp = k.ts AND w.station_id = k.sid",
+    )
+    .bind(keys.iter().map(|(t, _)| *t).collect::<Vec<_>>())
+    .bind(keys.iter().map(|(_, n)| *n as i32).collect::<Vec<_>>())
+    .execute(pool)
+    .await
+    .context("deleting RID spikes")?;
+    Ok(r.rows_affected() as usize)
+}
+
+/// Delete the model rows of forecast_history issued at or after `from`, so
+/// a catch-up recomputes them (after the readings they were computed from
+/// changed).
+pub async fn delete_model_forecasts_since(pool: &PgPool, from: NaiveDateTime) -> Result<usize> {
+    let r = sqlx::query("DELETE FROM forecast_history WHERE as_of >= $1 AND source = 'model'")
+        .bind(from)
+        .execute(pool)
+        .await
+        .context("deleting forecast_history rows")?;
+    Ok(r.rows_affected() as usize)
+}
+
 /// Day -> hours (0-23) that have at least one RID reading, for [start, end].
 pub async fn recorded_hours_by_day(
     pool: &PgPool,

rust/collector/src/db_tests.rs

@@ -191,6 +191,147 @@ async fn yesterdays_late_hours_are_recovered_after_midnight() {
     db.drop().await;
 }
 
+/// A report source with fixed P.1 levels per hour of one report date.
+struct ScriptedRid(Mutex<HashMap<NaiveDate, Vec<(i64, f64)>>>);
+
+impl RidSource for ScriptedRid {
+    async fn fetch_report(&self, date: NaiveDate) -> Result<Vec<Reading>> {
+        let script = self
+            .0
+            .lock()
+            .unwrap()
+            .get(&date)
+            .cloned()
+            .unwrap_or_default();
+        Ok(script
+            .into_iter()
+            .map(|(h, level)| Reading {
+                timestamp: date.and_time(NaiveTime::MIN) + Duration::hours(h),
+                station_num: 8,
+                water_level: level,
+                discharge: Some(45.0),
+                discharge_percent: None,
+            })
+            .collect())
+    }
+}
+
+async fn p1_levels(pool: &PgPool) -> Vec<(NaiveDateTime, f64, Option<f64>)> {
+    sqlx::query_as(
+        "SELECT timestamp, water_level::float8, discharge::float8 FROM water_measurements
+         WHERE station_id = 8 ORDER BY timestamp",
+    )
+    .fetch_all(pool)
+    .await
+    .unwrap()
+}
+
+#[tokio::test]
+async fn spikes_are_screened_and_refetch_repairs_the_store() {
+    let Some(db) = TestDb::new().await else {
+        return;
+    };
+    let stations = Stations::bundled();
+    let day = d(2026, 10, 24);
+    // what the store held before screening existed: the slip, level-only
+    db::save_rid(
+        &db.pool,
+        &[
+            Reading {
+                timestamp: t(2026, 10, 24, 23, 0),
+                station_num: 8,
+                water_level: 1.04,
+                discharge: None,
+                discharge_percent: None,
+            },
+            Reading {
+                timestamp: t(2026, 10, 25, 0, 0),
+                station_num: 8,
+                water_level: 10.6,
+                discharge: None,
+                discharge_percent: None,
+            },
+            Reading {
+                timestamp: t(2026, 10, 25, 1, 0),
+                station_num: 8,
+                water_level: 1.07,
+                discharge: Some(45.0),
+                discharge_percent: None,
+            },
+        ],
+        &stations,
+    )
+    .await
+    .unwrap();
+
+    // RID still serves the slip: refetch deletes it and fills the discharge
+    let src = ScriptedRid(Mutex::new(HashMap::from([(
+        day,
+        vec![(23, 1.04), (24, 10.6)],
+    )])));
+    let s = rid::refetch(&src, &db.pool, &stations, day, day)
+        .await
+        .unwrap();
+    assert_eq!((s.saved, s.spikes, s.deleted), (1, 1, 1));
+    assert_eq!(s.first_change, Some(t(2026, 10, 24, 23, 0)));
+    assert_eq!(
+        p1_levels(&db.pool).await,
+        vec![
+            (t(2026, 10, 24, 23, 0), 1.04, Some(45.0)),
+            (t(2026, 10, 25, 1, 0), 1.07, Some(45.0)),
+        ]
+    );
+
+    // RID corrected it: the hour comes back with the right value
+    src.0
+        .lock()
+        .unwrap()
+        .insert(day, vec![(23, 1.04), (24, 1.06)]);
+    let s = rid::refetch(&src, &db.pool, &stations, day, day)
+        .await
+        .unwrap();
+    assert_eq!((s.saved, s.spikes, s.deleted, s.changed), (2, 0, 0, 1));
+    assert_eq!(s.first_change, Some(t(2026, 10, 25, 0, 0)));
+    assert_eq!(p1_levels(&db.pool).await.len(), 3);
+
+    // a live cycle: the newest hour jumps and waits for the next one
+    let today = d(2026, 10, 25);
+    src.0
+        .lock()
+        .unwrap()
+        .insert(today, vec![(1, 1.07), (2, 9.9)]);
+    // (yesterday's report comes along: it still has holes)
+    let o = rid::run_cycle(&src, &db.pool, &stations, t(2026, 10, 25, 2, 10)).await;
+    assert_eq!((o.fetched, o.saved), (4, 3));
+    assert_eq!(
+        p1_levels(&db.pool).await.last().unwrap().0,
+        t(2026, 10, 25, 1, 0)
+    );
+    // 03:00 is back at 1.1: 02:00 was a spike and never stored
+    src.0
+        .lock()
+        .unwrap()
+        .insert(today, vec![(1, 1.07), (2, 9.9), (3, 1.1)]);
+    let o = rid::run_cycle(&src, &db.pool, &stations, t(2026, 10, 25, 3, 10)).await;
+    assert_eq!((o.fetched, o.saved), (5, 4));
+    let levels: Vec<f64> = p1_levels(&db.pool).await.iter().map(|r| r.1).collect();
+    assert_eq!(levels, vec![1.04, 1.06, 1.07, 1.1]);
+
+    // the forecasts issued from the changed hours can be redone
+    let mut f = forecast_row("P.1", 24, 0.9);
+    f.as_of = t(2026, 10, 25, 0, 0);
+    let mut g = forecast_row("P.1", 24, 0.0);
+    g.as_of = t(2026, 10, 24, 23, 0);
+    db::save_forecast_history(&db.pool, &[f, g]).await.unwrap();
+    assert_eq!(
+        db::delete_model_forecasts_since(&db.pool, t(2026, 10, 25, 0, 0))
+            .await
+            .unwrap(),
+        1
+    );
+    db.drop().await;
+}
+
 #[tokio::test]
 async fn rid_upsert_revises_values_and_rounds_like_numeric() {
     let Some(db) = TestDb::new().await else {

rust/collector/src/main.rs

@@ -62,6 +62,14 @@ enum Command {
     },
     /// Re-fetch RID reports that have holes: a number of days back, or "all"
     FillGaps { days: String },
+    /// Re-fetch the RID reports of FROM..=TO (YYYY-MM-DD; day D's report holds
+    /// D 01:00 to D+1 00:00) even without holes, through the spike screening:
+    /// corrects readings RID has revised and removes stored spikes
+    Refetch {
+        from: chrono::NaiveDate,
+        /// Last report date [default: FROM]
+        to: Option<chrono::NaiveDate>,
+    },
     /// Connect, create missing tables, print row counts and the newest RID reading
     CheckDb,
     /// Forecast from the database now and print it; stores nothing without --save
@@ -87,6 +95,11 @@ enum Command {
         /// How many hours back to look
         #[arg(long, default_value_t = 336)]
         hours: i64,
+        /// First delete the model forecasts issued from this local time on
+        /// (e.g. 2026-10-25T00:00) so they are recomputed: after `refetch`
+        /// changed the readings they came from
+        #[arg(long)]
+        redo_from: Option<String>,
     },
     /// Show the readings the notifier judges and the ntfy transitions due now
     Notify {
@@ -212,7 +225,43 @@ async fn run(cli: Cli) -> Result<()> {
             let rain = rain_file.as_deref().map(read_rain_file).transpose()?;
             collector.forecast(json, save, now, rain).await
         }
-        Command::CatchUpForecasts { hours } => collector.catch_up_forecasts(hours).await,
+        Command::CatchUpForecasts { hours, redo_from } => {
+            let redo_from = redo_from.as_deref().map(parse_local).transpose()?;
+            collector.catch_up_forecasts(hours, redo_from).await
+        }
+        Command::Refetch { from, to } => {
+            let to = to.unwrap_or(from);
+            if to < from {
+                bail!("refetch: TO ({to}) is before FROM ({from})");
+            }
+            db::ensure_schema(&collector.pool).await?;
+            let s = rid::refetch(
+                &collector.rid,
+                &collector.pool,
+                &collector.stations(),
+                from,
+                to,
+            )
+            .await?;
+            println!(
+                "{} readings saved ({} new or changed), {} spikes rejected, {} stored spikes deleted, {} held for want of the next reading",
+                s.saved,
+                s.changed - s.deleted,
+                s.spikes,
+                s.deleted,
+                s.held
+            );
+            if let Some(t) = s.first_change {
+                let hours = (util::bangkok_now() - t).num_hours() + 2;
+                if hours <= 24 * 60 {
+                    println!(
+                        "forecasts issued since {t} came from other readings; recompute them with\n  ping-collector catch-up-forecasts --hours {hours} --redo-from {}",
+                        t.format("%Y-%m-%dT%H:%M")
+                    );
+                }
+            }
+            Ok(())
+        }
         Command::Notify { dry_run } => collector.notify(dry_run).await,
         Command::CheckDb => {
             db::ensure_schema(&collector.pool).await?;

rust/collector/src/rid.rs

@@ -9,6 +9,12 @@
 //! every reading it gets (not only "the current hour"), and re-fetches
 //! yesterday's report while the store still has holes in it: hours 22-24 of a
 //! day are reachable only through that day's report.
+//!
+//! Before storing, readings are screened for single-hour spikes (see
+//! [`judge`]): RID's report sometimes carries a slipped decimal point, a
+//! flipped sign or a placeholder for one hour (P.1 10.6 m between 1.04 and
+//! 1.07 on 2026-10-25 00:00), which would raise a false alarm and
+//! distort the forecast features for three days.
 
 use crate::db;
 use crate::stations::Stations;
@@ -17,7 +23,7 @@ use anyhow::{Context, Result};
 use chrono::{Datelike, Duration, NaiveDate, NaiveDateTime, NaiveTime};
 use serde_json::Value;
 use sqlx::PgPool;
-use std::collections::{BTreeSet, HashMap, HashSet};
+use std::collections::{BTreeMap, BTreeSet, HashMap, HashSet};
 use std::future::Future;
 
 pub const N_STATIONS: u32 = 16;
@@ -96,6 +102,235 @@ pub fn parse_report(report_date: NaiveDate, payload: &Value) -> Vec<Reading> {
     out
 }
 
+// ---- Screening -------------------------------------------------------------
+
+/// Neighbours of a reading are looked for this many hours before and after it.
+const SCREEN_WINDOW_H: i64 = 3;
+
+/// Gauges on flashy tributaries: genuine 1-h changes of 1.3-3.1 m are on
+/// record (P.76 0.09 -> 3.18 m on 2026-05-29). Elsewhere the largest genuine
+/// 1-h change since 2018 is 1.1 m (P.1: 0.56 m).
+const FLASHY: &[&str] = &["P.76", "P.77", "P.81", "P.82", "P.84", "P.85", "P.87"];
+
+/// A change from the previous reading larger than this (m) needs the next
+/// reading to tell a spike from a real change.
+pub fn jump_limit(code: &str) -> f64 {
+    if FLASHY.contains(&code) { 2.0 } else { 1.0 }
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub enum Verdict {
+    Keep,
+    /// The newest reading jumped more than twice the limit: wait for the next
+    /// one (an hour at most while RID is current) before storing it.
+    Hold,
+    /// Out and back within an hour or two: not stored.
+    Spike,
+}
+
+/// `prev`: the last accepted level within [`SCREEN_WINDOW_H`] before the
+/// reading, `next`: the following reading within the window, if any.
+///
+/// A spike leaves the previous level by more than `limit` and the next
+/// reading comes back, closer to the previous level than half of either
+/// jump. A rising or falling limb (prev < level < next) or a step the next
+/// reading confirms is kept. 2018-2026 history: 16 spikes (sign flips,
+/// slipped decimal points, -5/-6/-9 m placeholders) and no genuine change
+/// would have been rejected.
+pub fn judge(limit: f64, prev: Option<f64>, level: f64, next: Option<f64>) -> Verdict {
+    let Some(p) = prev else {
+        return Verdict::Keep;
+    };
+    let out = (level - p).abs();
+    if out <= limit {
+        return Verdict::Keep;
+    }
+    let Some(n) = next else {
+        return if out > 2.0 * limit {
+            Verdict::Hold
+        } else {
+            Verdict::Keep
+        };
+    };
+    let back = (level - n).abs();
+    if back > limit && (n - p).abs() < out.min(back) / 2.0 {
+        Verdict::Spike
+    } else {
+        Verdict::Keep
+    }
+}
+
+/// A reading already in the database: (timestamp, report column, level, discharge).
+pub type StoredReading = (NaiveDateTime, u32, f64, Option<f64>);
+
+#[derive(Debug, Default)]
+pub struct Screened {
+    pub keep: Vec<Reading>,
+    pub held: Vec<Reading>,
+    pub spikes: Vec<Reading>,
+    /// Stored rows at a spike's hour that are spikes themselves: delete.
+    pub delete: Vec<(NaiveDateTime, u32)>,
+}
+
+/// Screen a batch against itself and the stored readings around it. A batch
+/// reading replaces the stored one at the same hour; stored readings count
+/// as accepted. `code` maps a report column to its station code.
+pub fn screen(
+    readings: &[Reading],
+    stored: &[StoredReading],
+    code: impl Fn(u32) -> String,
+) -> Screened {
+    #[derive(Clone, Copy)]
+    enum Src {
+        Batch(usize),
+        Stored,
+    }
+    let mut out = Screened::default();
+    let mut by_station: BTreeMap<u32, BTreeMap<NaiveDateTime, (f64, Src)>> = BTreeMap::new();
+    for (t, n, level, _) in stored {
+        by_station
+            .entry(*n)
+            .or_default()
+            .insert(*t, (*level, Src::Stored));
+    }
+    let stored_at: HashMap<(NaiveDateTime, u32), f64> =
+        stored.iter().map(|(t, n, l, _)| ((*t, *n), *l)).collect();
+    // the last copy of an hour wins, as in the upsert
+    for (i, r) in readings.iter().enumerate() {
+        by_station
+            .entry(r.station_num)
+            .or_default()
+            .insert(r.timestamp, (r.water_level, Src::Batch(i)));
+    }
+    let window = Duration::hours(SCREEN_WINDOW_H);
+    for (n, series) in &by_station {
+        let limit = jump_limit(&code(*n));
+        let mut last: Option<(NaiveDateTime, f64)> = None;
+        let entries: Vec<_> = series.iter().map(|(t, v)| (*t, *v)).collect();
+        for (k, &(t, (level, src))) in entries.iter().enumerate() {
+            let prev = last.filter(|(pt, _)| t - *pt <= window).map(|(_, l)| l);
+            let i = match src {
+                Src::Stored => {
+                    last = Some((t, level));
+                    continue;
+                }
+                Src::Batch(i) => i,
+            };
+            let next = entries
+                .get(k + 1)
+                .filter(|(nt, _)| *nt - t <= window)
+                .map(|(_, (l, _))| *l);
+            match judge(limit, prev, level, next) {
+                Verdict::Keep => {
+                    out.keep.push(readings[i].clone());
+                    last = Some((t, level));
+                }
+                Verdict::Hold => out.held.push(readings[i].clone()),
+                Verdict::Spike => {
+                    out.spikes.push(readings[i].clone());
+                    if let Some(&old) = stored_at.get(&(t, *n)) {
+                        if judge(limit, prev, old, next) == Verdict::Spike {
+                            out.delete.push((t, *n));
+                        } else {
+                            last = Some((t, old));
+                        }
+                    }
+                }
+            }
+        }
+    }
+    out.keep.sort_by_key(|r| (r.timestamp, r.station_num));
+    out
+}
+
+#[derive(Debug, Default)]
+pub struct StoreOutcome {
+    pub saved: usize,
+    pub held: usize,
+    pub spikes: usize,
+    pub deleted: usize,
+    /// Saved readings that were new or differ from the stored ones, plus deletions.
+    pub changed: usize,
+    /// The earliest hour whose stored data changed (forecasts from then on
+    /// were computed from other data).
+    pub first_change: Option<NaiveDateTime>,
+}
+
+/// Screen `readings` against the stored ones around them, then save them and
+/// delete stored spikes.
+pub async fn store(
+    pool: &PgPool,
+    readings: &[Reading],
+    stations: &Stations,
+) -> Result<StoreOutcome> {
+    let mut out = StoreOutcome::default();
+    let (Some(first), Some(last)) = (
+        readings.iter().map(|r| r.timestamp).min(),
+        readings.iter().map(|r| r.timestamp).max(),
+    ) else {
+        return Ok(out);
+    };
+    let window = Duration::hours(SCREEN_WINDOW_H);
+    let nums: Vec<u32> = readings
+        .iter()
+        .map(|r| r.station_num)
+        .collect::<BTreeSet<_>>()
+        .into_iter()
+        .collect();
+    let stored = db::rid_readings_between(pool, &nums, first - window, last + window).await?;
+    let code = |n: u32| stations.get(n).code;
+    let s = screen(readings, &stored, code);
+    for r in &s.spikes {
+        tracing::warn!(
+            "RID {} {}: {} m is a one-hour spike, not stored",
+            r.timestamp,
+            code(r.station_num),
+            r.water_level
+        );
+    }
+    for r in &s.held {
+        tracing::warn!(
+            "RID {} {}: {} m jumps from the previous reading; waiting for the next one",
+            r.timestamp,
+            code(r.station_num),
+            r.water_level
+        );
+    }
+    let before: HashMap<(NaiveDateTime, u32), (f64, Option<f64>)> = stored
+        .iter()
+        .map(|(t, n, l, q)| ((*t, *n), (*l, *q)))
+        .collect();
+    // NUMERIC(10,3) / (10,2): compare at the stored precision
+    let differs = |a: f64, b: f64, places: i32| {
+        let f = 10f64.powi(places);
+        (a * f).round() != (b * f).round()
+    };
+    let mut changed: Vec<NaiveDateTime> = s
+        .keep
+        .iter()
+        .filter(|r| match before.get(&(r.timestamp, r.station_num)) {
+            None => true,
+            Some((l, q)) => {
+                differs(*l, r.water_level, 3)
+                    || match (q, r.discharge) {
+                        (Some(a), Some(b)) => differs(*a, b, 2),
+                        (None, None) => false,
+                        _ => true,
+                    }
+            }
+        })
+        .map(|r| r.timestamp)
+        .collect();
+    changed.extend(s.delete.iter().map(|(t, _)| *t));
+    out.saved = db::save_rid(pool, &s.keep, stations).await?;
+    out.deleted = db::delete_rid_readings(pool, &s.delete).await?;
+    out.held = s.held.len();
+    out.spikes = s.spikes.len();
+    out.changed = changed.len();
+    out.first_change = changed.into_iter().min();
+    Ok(out)
+}
+
 /// Anything that can produce a parsed RID report (HTTP in production,
 /// canned rows in tests).
 pub trait RidSource {
@@ -248,12 +483,14 @@ pub async fn run_cycle(
         tracing::warn!("RID: no data fetched for {:?}", out.dates);
         return out;
     }
-    if let Err(e) = db::save_rid(pool, &readings, stations).await {
+    match store(pool, &readings, stations).await {
-        tracing::error!("RID save failed: {e:#}");
+        Ok(s) => out.saved = s.saved,
-        out.error = Some(format!("save failed: {e:#}"));
+        Err(e) => {
-        return out;
+            tracing::error!("RID save failed: {e:#}");
+            out.error = Some(format!("save failed: {e:#}"));
+            return out;
+        }
     }
-    out.saved = readings.len();
     let latest = readings
         .iter()
         .map(|r| r.timestamp)
@@ -308,8 +545,7 @@ pub async fn fill_gaps(
         }
         match src.fetch_report(*d).await {
             Ok(readings) if !readings.is_empty() => {
-                db::save_rid(pool, &readings, stations).await?;
+                filled += store(pool, &readings, stations).await?.saved;
-                filled += readings.len();
             }
             Ok(_) => {
                 empty += 1;
@@ -332,6 +568,52 @@ pub async fn fill_gaps(
     Ok(filled)
 }
 
+/// Re-fetch the reports of `from..=to` (report dates: day D's report holds
+/// D 01:00 to D+1 00:00) whether or not they have holes, and store them
+/// through the screening: corrects readings RID has revised since (and
+/// fills the discharge of level-only hours), removes stored spikes.
+pub async fn refetch(
+    src: &impl RidSource,
+    pool: &PgPool,
+    stations: &Stations,
+    from: NaiveDate,
+    to: NaiveDate,
+) -> Result<StoreOutcome> {
+    let mut total = StoreOutcome::default();
+    let mut day = from;
+    while day <= to {
+        if day > from {
+            tokio::time::sleep(std::time::Duration::from_secs(1)).await; // be polite
+        }
+        match src.fetch_report(day).await {
+            Ok(readings) => {
+                let s = store(pool, &readings, stations).await?;
+                tracing::info!(
+                    "refetch {day}: {} fetched, {} saved ({} new or changed), {} spikes rejected, {} stored spikes deleted, {} held",
+                    readings.len(),
+                    s.saved,
+                    s.changed - s.deleted,
+                    s.spikes,
+                    s.deleted,
+                    s.held
+                );
+                total.saved += s.saved;
+                total.held += s.held;
+                total.spikes += s.spikes;
+                total.deleted += s.deleted;
+                total.changed += s.changed;
+                total.first_change = match (total.first_change, s.first_change) {
+                    (Some(a), Some(b)) => Some(a.min(b)),
+                    (a, b) => a.or(b),
+                };
+            }
+            Err(e) => tracing::warn!("refetch {day}: {e:#}"),
+        }
+        day += Duration::days(1);
+    }
+    Ok(total)
+}
+
 /// Next RID poll: `offset` minutes past the next hour when current, else in
 /// `retry` minutes (never later than that hourly poll).
 pub fn next_poll_time(
@@ -407,6 +689,104 @@ mod tests {
         assert!(!by.contains_key(&5)); // negative discharge drops the reading
     }
 
+    #[test]
+    fn judge_spikes_steps_and_limbs() {
+        use Verdict::*;
+        // the P.1 slip of 2026-10-25 00:00, a sign flip, a placeholder
+        assert_eq!(judge(1.0, Some(1.04), 10.6, Some(1.07)), Spike);
+        assert_eq!(judge(1.0, Some(-1.18), 1.18, Some(-1.18)), Spike);
+        assert_eq!(judge(2.0, Some(-0.05), -5.0, Some(-0.05)), Spike);
+        // small changes, no neighbour, a confirmed step, rising limbs
+        assert_eq!(judge(1.0, Some(1.0), 1.9, Some(1.0)), Keep);
+        assert_eq!(judge(1.0, None, 10.6, Some(1.07)), Keep);
+        assert_eq!(judge(1.0, Some(1.0), 2.5, Some(2.6)), Keep);
+        assert_eq!(judge(2.0, Some(0.04), 1.61, Some(2.67)), Keep);
+        assert_eq!(judge(1.0, Some(0.82), 1.78, Some(2.82)), Keep);
+        // a peak that falls back less than halfway in the next hour is real
+        assert_eq!(judge(1.0, Some(1.0), 3.0, Some(1.8)), Keep);
+        // the newest reading: a jump over twice the limit waits
+        assert_eq!(judge(1.0, Some(1.04), 10.6, None), Hold);
+        assert_eq!(judge(1.0, Some(1.0), 2.5, None), Keep);
+        assert_eq!(judge(2.0, Some(0.09), 3.18, None), Keep);
+        assert_eq!(jump_limit("P.76"), 2.0);
+        assert_eq!(jump_limit("P.1"), 1.0);
+    }
+
+    fn rd(h: u32, n: u32, level: f64) -> Reading {
+        Reading {
+            timestamp: t(2026, 10, 25, h, 0),
+            station_num: n,
+            water_level: level,
+            discharge: None,
+            discharge_percent: None,
+        }
+    }
+
+    fn code(n: u32) -> String {
+        if n == 8 {
+            "P.1".into()
+        } else {
+            format!("P.{}", 19 + n)
+        }
+    }
+
+    #[test]
+    fn screen_rejects_a_spike_and_deletes_its_stored_copy() {
+        let stored = vec![
+            (t(2026, 10, 24, 23, 0), 8, 1.04, None),
+            (t(2026, 10, 25, 0, 0), 8, 10.6, None),
+        ];
+        // the report re-fetched, RID still has the slip
+        let batch = vec![rd(0, 8, 10.6), rd(1, 8, 1.07), rd(2, 8, 1.09)];
+        let s = screen(&batch, &stored, code);
+        assert_eq!(s.spikes, vec![rd(0, 8, 10.6)]);
+        assert_eq!(s.delete, vec![(t(2026, 10, 25, 0, 0), 8)]);
+        assert_eq!(s.keep, vec![rd(1, 8, 1.07), rd(2, 8, 1.09)]);
+        // RID corrected it: the new value replaces the stored spike
+        let s = screen(&[rd(0, 8, 1.06), rd(1, 8, 1.07)], &stored, code);
+        assert!(s.spikes.is_empty() && s.delete.is_empty());
+        assert_eq!(s.keep, vec![rd(0, 8, 1.06), rd(1, 8, 1.07)]);
+    }
+
+    #[test]
+    fn screen_keeps_a_good_stored_copy_and_judges_the_next_hour_by_it() {
+        let stored = vec![
+            (t(2026, 10, 25, 0, 0), 8, 1.06, None),
+            (t(2026, 10, 25, 1, 0), 8, 1.07, None),
+        ];
+        let s = screen(&[rd(1, 8, 10.7), rd(2, 8, 1.09)], &stored, code);
+        assert_eq!(s.spikes, vec![rd(1, 8, 10.7)]);
+        assert!(s.delete.is_empty());
+        assert_eq!(s.keep, vec![rd(2, 8, 1.09)]);
+    }
+
+    #[test]
+    fn screen_holds_the_newest_jump_until_the_next_reading() {
+        let stored = vec![(t(2026, 10, 25, 13, 0), 8, 1.04, None)];
+        let s = screen(&[rd(14, 8, 10.6)], &stored, code);
+        assert_eq!(s.held, vec![rd(14, 8, 10.6)]);
+        assert!(s.keep.is_empty() && s.spikes.is_empty());
+        // next poll: the report again, now with 15:00
+        let s = screen(&[rd(14, 8, 10.6), rd(15, 8, 1.06)], &stored, code);
+        assert_eq!(s.spikes, vec![rd(14, 8, 10.6)]);
+        assert_eq!(s.keep, vec![rd(15, 8, 1.06)]);
+        // a real jump the next reading confirms is stored then
+        let s = screen(&[rd(14, 8, 3.5), rd(15, 8, 3.6)], &stored, code);
+        assert_eq!(s.keep, vec![rd(14, 8, 3.5), rd(15, 8, 3.6)]);
+    }
+
+    #[test]
+    fn screen_window_and_stations_are_separate() {
+        // previous reading 4 h back: nothing to compare with
+        let stored = vec![(t(2026, 10, 25, 9, 0), 8, 1.0, None)];
+        let s = screen(&[rd(13, 8, 5.0), rd(13, 1, 0.5)], &stored, code);
+        assert_eq!(s.keep.len(), 2);
+        // a jump at one gauge does not involve the others
+        let stored = vec![(t(2026, 10, 25, 12, 0), 1, 0.1, None)];
+        let s = screen(&[rd(13, 8, 1.0), rd(13, 1, 0.12)], &stored, code);
+        assert_eq!(s.keep.len(), 2);
+    }
+
     #[test]
     fn missing_rows_key_is_empty() {
         assert!(parse_report(d(2026, 7, 22), &json!({"unexpected": true})).is_empty());

rust/collector/src/service.rs

@@ -400,10 +400,22 @@ impl Collector {
 
     /// `ping-collector catch-up-forecasts`: `forecasting::catch_up` over
     /// `hours`, e.g. once after an upgrade for the nights before it.
-    pub async fn catch_up_forecasts(&self, hours: i64) -> Result<()> {
+    pub async fn catch_up_forecasts(
+        &self,
+        hours: i64,
+        redo_from: Option<NaiveDateTime>,
+    ) -> Result<()> {
         if !(1..=24 * 60).contains(&hours) {
             bail!("--hours must be between 1 and 1440 (Open-Meteo keeps 92 days of past rain)");
         }
+        let now = bangkok_now();
+        if let Some(from) = redo_from
+            && from < floor_hour(now) - Duration::hours(hours)
+        {
+            bail!(
+                "--redo-from {from} is more than --hours {hours} back: those hours would not be recomputed"
+            );
+        }
         let mut store = ModelStore::new(&self.cfg.models_dir);
         let models = store.refresh()?;
         if models.is_empty() {
@@ -413,11 +425,15 @@ impl Collector {
             );
         }
         db::ensure_schema(&self.pool).await?;
+        if let Some(from) = redo_from {
+            let n = db::delete_model_forecasts_since(&self.pool, from).await?;
+            eprintln!("deleted {n} model forecasts issued since {from}; recomputing");
+        }
         let c = forecasting::catch_up(
             &self.pool,
             &self.rain,
             &models,
-            bangkok_now(),
+            now,
             hours,
             &Default::default(),
         )