fix(forecast): no nightly gaps in the issued model forecasts

The P.1 chart's "24 h peak, as issued" line had a gap every night and a
dip at 21:00. Two causes:

- A run forecasts from the newest reading only. RID publishes 22:00..06:00
  in one batch in the morning (and now and then one hour late), so those
  hours were never forecast from. After each run the collector now
  forecasts the hours of the last 48 h (two weeks on the first run after a
  start) that have a RID reading but no model forecast, each as a run at
  that hour would have: gauges, HII fill and dam state up to that hour.
  Rain is fetched once, so a late hour's rain_fc24 partly holds rain that
  has since fallen (as in training). Only missing (hour, station) pairs are
  written. `ping-collector catch-up-forecasts --hours N` goes further back.
- Overnight each hourly run re-issues the 21:00 reading, after 6 h as the
  heuristic, which overwrote the 21:00 model row. A heuristic row no longer
  replaces a model row in forecast_history.

The dashboard draws only model rows on the issued line.

Checked on a local copy of 18 days of RID data: the catch-up of an hour
matches `ping-collector forecast --now <hour>`; 288 hours took 20 s
(debug build). Tests for both database rules.

docs/DASHBOARD_REDESIGN.md

@@ -111,6 +111,7 @@ One screen, no page scroll on a desktop:
 - **Flood zones:** shown automatically while P.1 is at or above a stage, unless hidden by the user.
 - **Basin stations:** ThaiWater's classes for the level as a share of the bank height (critically low ≤ 10 %, low ≤ 30 %, normal ≤ 70 %, high ≤ 100 %, over the bank). On the map and in the list over the bank is danger, high is watch, the rest normal; no data after 24 h. Gauges RID already lists are shown once, as RID gauges. The layer is off until the list is opened or the button pressed.
 - **History:** above 2,500 readings (30 d and longer for most gauges) the chart shows the daily mean with the daily low–high band, so a flood peak is not averaged away; gaps longer than 3.5 h (2.5 days when daily) break the line rather than bridging RID's nightly pause. Short ranges reload with every refresh, long ones at most hourly. "All" asks for `limit=100000`: `/measurements/history` returns the oldest rows first and stops at its 50,000 default, which made the old dashboard's "All time" for P.1 end in October 2024.
+- **Issued forecasts:** the history's violet line is the model's 24 h peak for every hour (`/api/forecast/history`, `source = model` only: a heuristic row is the level itself and drew a dip into the line). The nightly gaps were RID's overnight batch; the collector now forecasts those hours when they arrive (docs/FLOOD_FORECASTING.md, Serving).
 - **Forecast accuracy:** loaded when the section is opened (a cold `/api/forecast/skill` query takes up to 20 s), cached for 15 minutes; bias and persistence error are in each row's tooltip.
 - **Replay of October 2024:** `src/static/flood-2024-all.json` (hourly readings of the 16 gauges, 15 September – 12 October 2024, and the 24 h warning chance and peak of a model trained on data before 1 September 2024). The page's clock, verdict, list, markers, flood zones and panel follow the replay; the panel's chart shows the season on fixed axes up to the replay hour. It plays hour by hour while the model alerts or P.1 is above 3.50 m, faster through quiet days, and can be paused and dragged. Rain, the basin stations and the 5-minute refresh are off during the replay, since they are today's data. A violet bar (REPLAY, date, P.1 level, the model's message) and the top bar's "Replay time" replace the live data age.
 - **Simulation:** `?demo_level=4.4` pins P.1 at 4.40 m; `?demo_rise=1` raises it by 5 cm every 0.7 s from the real level to the 2024 record (5.30 m). An amber SIMULATION bar names the real level; the top bar and the P.1 panel say "simulated". Only P.1 is simulated: the forecasts and other gauges stay live.