goodwe_controller/DECISIONS/PRICE-02-the-verdict-skips-ended-periods.md

@@ -80,8 +80,10 @@ tested the first of them only because its plan began at midnight.
 
 On real 15-minute days the trade runs longer: about 1.5 to 2.5 times the
 minutes spent charging or discharging, with fewer on/off switches (for example
-78 -> 68, 92 -> 72). Through a modelled 10 kWh battery that is 27.5 -> 42.6 kWh
-over 16 days (2026-09-02 to 09-17, a decision every minute, refresh at :00 or
-:10, tomorrow's prices known always or from 13:00; reviewer's sim). Charge to
+78 -> 68, 92 -> 72), over 16 days (2026-09-02 to 09-17, a decision every
+minute, refresh at :00 or :10, tomorrow's prices known always or from 13:00;
+reviewer's sim). Through a modelled 10 kWh battery that is 27.5 -> 42.6 kWh
+(refresh at :00, tomorrow always known); 19.6-31.9 -> 43.4-48.7 kWh in the
+other three variants. Charge to
 discharge reversals are unchanged (10 -> 10, 13 -> 13). More cycling is the
 price of not pausing.

goodwe_controller/test_arbitrage.py

@@ -2093,12 +2093,14 @@ def replay_real(days, threshold, unc, *, soc0=50.0, floor=20.0, ceiling=95.0):
     it and never a day on its own. (An earlier version built one plan per
     calendar day with no tomorrow: after the cheap hour no cheap hour was left
     ahead, and the ledger shrank to a few trades that no real site would see.)
+    An open buy is kept over midnight: with tomorrow in view most buys are
+    sold the next day or later (46 of 51 at the shipped threshold).
     """
     soc = soc0
     trades, holdings, bad, neg_sells = [], [], [], 0
+    open_buy = None
     for di, day in enumerate(days):
         nxt = days[di + 1] if di + 1 < len(days) else []
-        open_buy = None
         for i, p in enumerate(day):
             ahead = day[i:] + (nxt if brussels(p.start).hour >= 13 else [])
             plan = TariffPlan(