Stacked on #5 — merge #5 (fix/battery-control-layer) first; this PR's base is set to that branch so it only shows its own changes. After #5 merges, retarget/rebase this PR onto dev.

Problem

1. Switch arrays double-compensate. The switch-array hysteresis acted on the residual from before this cycle's numeric-array correction. An export that the numeric arrays had just fully curtailed could still toggle a switch array off in the same cycle — both actuators correct the same watts and the grid swings to import.
2. Integrator windup against a saturated plant. The PID integrator kept accumulating while no actuator could absorb the output (all arrays in their settling window or at their limits, no switch toggled). With persistent import and PV at max, the I-term wound up towards the output clamp and caused an overshoot the moment headroom returned (cloud passing, settling expiring).

Fix

• _distribute_to_numeric_arrays now returns the unabsorbed remainder; the watts actually absorbed are fed forward into the residual used for switch-array decisions, and each switch toggle also updates the residual for subsequent switches.
• When the PID output moved nothing (no numeric absorption, no switch action), the integrator is frozen for the next cycle — same mechanism already used for deadband and calibration.

Tests

tests/test_control_loop_hardening.py: switch array stays on when numeric curtailment already absorbed the export, still toggles when nothing else can absorb, integrator stops growing while saturated, and control resumes normally when headroom returns. Full suite: 139 passed.

https://claude.ai/code/session_01RUWpwxbGsgR3PoLHLq4Djz

Generated by Claude Code