Interval HVAC disaggregation + balance-point temperature

[delabrcd]

Sub-issue of #76 (depends on interval ingest). Sharpens #47 (weather-normalization) and #67 (Kalman next-bill model). Related: #45.

What

Use interval usage correlated with hourly temperature to separate weather-driven HVAC load from everything else, at much finer resolution than the monthly degree-day regression.

• Balance-point temperature per fuel — the outdoor temp where heating/cooling kicks in — fit from interval-usage-vs-temp.
• Heating/cooling sensitivity (kWh or therms per degree) from the slope above/below the balance point.
• HVAC vs. baseload split: of a given period's usage, how much was temperature-driven vs. flat.
• Gas furnace efficiency drift — daily gas vs. temperature trend across a season.

Why interval data + dependency

Hourly usage-vs-temperature gives a clean balance-point fit that monthly data smears out. Dependency: this needs hourly weather to pair with interval usage — today we only store WeatherDaily (Open-Meteo). Add an hourly weather pull (Open-Meteo archive supports it) scoped to the account location, or note it as a prerequisite sub-task.

Implementation notes

• Regression / balance-point fit in app/src/lib/series.ts or prediction.ts, pure + tested.
• Feed the derived sensitivities back into the Proper baseload-aware weather model for the next-bill estimate #67 model as sharper priors (don't rip out the monthly fallback for short-history/pre-AMI accounts).
• Scatter (usage vs temp) + fitted balance point via chartSpec.ts.

Acceptance criteria

• Hourly weather available for the account (new pull or documented prerequisite).
• Balance-point temp + heating/cooling sensitivity per fuel, unit-tested.
• HVAC-vs-baseload split surfaced.
• Demonstrated improvement (or clean integration) with the Proper baseload-aware weather model for the next-bill estimate #67 next-bill estimate; monthly fallback intact.