fix(loadmodel): bound the heating estimate by the house's own use

[frahlg]

Problem

The legacy load twin cannot lower a heating coefficient that is set too high:

• Update skips an hour's base sample whenever the heating estimate exceeds the measured load.
• The fit that moves the coefficient runs only in hours with eight trained days.

So an overstated coefficient never trains the hours it overstates, and never comes down. The home box had weather.heating_w_per_degc: 300: about 2.4 kW of assumed heat at 10 °C against a measured night load of 0.35–0.45 kW. The plan's night load was about 2 kW too high every night. (That setting is now removed on the box.)

Change

Heating is part of the house's load, so over a day's cold samples coef × Σ(18 °C − T) ≤ Σ load.

• Each local day with at least six hours of samples below 15 °C yields that bound.
• Once the last three such days all put the coefficient above their bound, it moves halfway toward the highest of them.
• The bound needs no trained hours. It only lowers the coefficient; the existing fit still raises it.
• The model saves the day's sums and the last three bounds in new fields. Old state loads with empty sums and the same feature hash, so no site loses what it has learned.

Three days, not one: a single day with a wood stove or the heating off has a low bound, and would otherwise halve a correct coefficient. Only the slow fit would raise it again.

Results (simulated, one sample a minute)

Case	Coefficient	02:00 error before → after
Home box: 300 W/°C prior, flat 400 W at 10 °C	175 → 112.5 → 81.25 → … → 50	2,650 → 650 W
Real heating: 300 W + 200 W/°C, start 150 or 200	unchanged	—
Real heating, start 400	falls to about 223, never below 200	—
Real heating, one day with the heating off	unchanged at 200	—

What this does not fix

Night hours still do not train in a house whose night load is below its daily average. The bound settles near the day's mean load ÷ mean ΔT, so the heating estimate stays above the night load and those buckets keep their prior. A day-shaped house without electric heating (mean 1,209 W, 300 W at night, 10 °C) improves from 2,750 to about 1,560 W too high at 02:00, but not to zero. #1491 stays open for that. Options for the owner:

• train skipped hours with a base of zero instead of skipping them;
• store each hour's mean load and temperature, so a coefficient change cannot shift trained hours;
• leave it: with fix(forecast): choose each forecast source from measured errors #1494 the planner uses Energyplan load whenever it measures clearly better.

Tests

• TestHeatingBoundLowersAnOverstatedPrior: the home box case steps at Thu, Fri and Sat midnight.
• TestHeatingBoundKeepsRealHeating, TestHeatingBoundIgnoresOneUnusualDay, TestHeatingBoundIgnoresMildDays, TestHeatingBoundNeedsSixColdHours.
• TestHeatingBoundSumsSurviveRestart, TestStateWithoutHeatingBoundStillLoads; TestFeatureHashPinned unchanged.
• A single-day bound fails three of these tests. make verify clean.

Refs #1491. One of four forecast fixes found while evaluating #1482, with #1492, #1493 and #1494. It shares no files with them. No forecast policy bump: #1494 already starts a new evaluation cohort in the same release.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-01T17:33:08.071708Z	742b3d0	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.