feat: value the state of charge the device is expected to leave with

[andig]

closes #163

One optional field per battery lets the schedule weigh that a vehicle may leave before the horizon ends. Absent, the model is unchanged and every stored case solves as before.

• r_departure: probability of leaving in each time step, the remainder stays. p_a then values the state of charge the device is expected to leave with, the departure weighted sum of s[t], instead of s[-1]. A charge that may be cut short is worth less than one already in. One objective term, no new variables or rows.
• Once the probabilities add up to one the device has certainly left, and its charge and discharge bounds are zero from there on. Without that the cost neutral tie break could put free surplus into a vehicle that is gone: with p_E = 0 and charge_before_export it did.
• The cost of charging planned after an uncertain departure is still counted in full, so the schedule leans early rather than being exact in expectation.
• A p_demand past a certain departure is a conflicting request. The demand penalty wins and the schedule fills the vehicle before it leaves, at real cost. The caller should cut the demand where it cuts the presence.
• The reported objective value follows the same accounting.
• tools/bench_departure.py runs the stored cases against main, then with a 30 % departure chance spread over hours 6 to 12 on every battery without discharge. Plain: 5.20 s against 5.18 s total. Departure: 4.19 s, rows and columns unchanged, 020 and 028 solve faster with it. A certain departure at hour 12 over the 13 cases that reach it: 3.47 s before the bounds, 2.45 s with them, since the bounded steps leave the model in presolve. 023 drops from 0.69 to 0.04 s, 028 from 0.69 to 0.38 s.
• evcc has to send the field for this to take effect, from the session history or a default such as a 24 h half-life. Forecast uncertainty is the separate feat: price the grid at the mean over the forecast band #185.

🤖 Generated with Claude Code