Tariff dashboard engine over-states TOU tariffs (Agile +38%) — bills consumption × avg rate instead of import × half-hourly rate

[albinati]

Symptom

GET /api/v1/tariffs/dashboard's projection for the user's current
half-hourly tariff (AGILE-24-10-01) was £108.27/month, while the
real realised cost for the SAME month was £78.43. The engine over-
states Agile by ~38 %.

=== ACTUAL realised May cost (from /energy/period — real half-hourly) ===
  import £69.43 + standing £19.29 - export £10.29 = NET £78.43

=== Tariffs dashboard projection for Agile (same month) ===
  total = £108.27
  unit_rate avg = 29.59p · standing = 62p/d

Effect: the comparison table showed Agile costing MORE than BG Fixed
v58 (£80.56), when in fact the user beats BG every month. Users on
solar+battery+TOU tariffs see a comparison that contradicts their
real bills.

Likely cause

The engine appears to use a flat-rate model:

total_pence = consumption_kwh × unit_rate_pence + days × standing_per_day

For flat tariffs (BG Fixed, Cosy Fixed) this is accurate.
For half-hourly tariffs (Agile, Intelligent Flux) it's wrong because:

1. It bills CONSUMPTION (total load) instead of IMPORT (what actually
   came from the grid) — ignoring battery + solar self-use.
2. It uses the AVERAGE rate instead of the per-slot rate — ignoring the
   user's deliberate strategy of charging the battery during cheap
   slots and discharging during peak slots.

The dashboard's pricing field already distinguishes
half_hourly vs time_of_use vs flat — the engine could
branch on this and replay the half-hour profile against the tariff's
own half-hourly rates (analogous to _realised_import_pence in
src/analytics/pnl.py but indexed by the candidate tariff's
agile_rates rather than the active one).

Workaround shipped (UI)

src/components/home/TariffComparisonWidget.tsx now overrides the
current row's projected total with the realised cost from
/energy/period?period=month (real half-hourly × measured import).
Other rows keep the engine projection, with foot text noting the basis
difference:

Your current tariff = real realised cost (half-hourly Agile ×
measured grid import). Others = engine projection (flat avg rate ×
usage, no battery-arbitrage credit), so they over-state TOU tariffs.

Asks

1. Backend tariff engine — replay measured grid_import_kw half-hour
   buckets against each tariff's own half-hourly rates (when
   pricing in {"half_hourly","time_of_use"}) instead of
   consumption × avg rate. Use db.half_hourly_grid_import_kwh_for_day
   the same way _realised_import_pence does. Standing + export
   remain straightforward.
2. Where a TOU tariff has time-of-use bands (Octopus Go cheap nights
   etc.), replay each slot against the band the slot falls into.
3. Remove the UI workaround once the engine reports realistic numbers.

Related

• src/api/main.py:3104 — /tariffs/dashboard endpoint
• src/energy/tariff_engine.py — likely engine location
• PR feat(ui): rebuild as Vite + Preact SPA — landing, cockpit, forecast, settings #418 — UI workaround (c7293b1)