fix(battery): stop scheduled batteries and the PID from double-correcting (oscillation), add SoC limits and reactive fallback

[bvweerd]

Problem

The scheduled-battery correction (filtered × capacity share, summed over batteries = the full grid error) and the PID both corrected the same grid error every cycle. Combined loop gain ≈ 2 → the closed loop diverges. Simulated with an ideal battery, schedule = 0, initial import 1000 W:

c01 grid= -1000  sp= 60   bat_target= 1000
c02 grid= +2000  sp= 40   bat_target=-1000
c04 grid= +5000  sp= 20   bat_target=-3000
c06 grid= +8000  sp=  0   bat_target=-5000   ← sustained limit cycle
c07 grid= -7000  sp=100   bat_target=+5000   ← battery bang-bangs at full power, every 5 s

Two more issues in the scheduled path:

• No SoC limits: the correction could push a full battery into charge or an empty one into discharge — SoC gating only existed in the reactive layer.
• Silent degradation on optimizer failure: schedule_w = read_power_w(...) or 0.0 turned an unavailable/unknown/non-numeric schedule sensor into schedule 0, leaving the battery under pure P-on-grid control with no SoC protection and no notification.

Fix

• The commanded-but-not-yet-measured scheduled battery response is fed forward into the PID residual (same 30 s trust-window mechanism as reactive batteries), so the PID only handles what the battery will not absorb. The oscillation scenario now converges within the deadband (closed-loop regression test included).
• SoC limits bound the corrected target: no charging at/above max_soc, no discharging at/below min_soc.
• Optimizer sensor unavailable → reactive fallback: the battery drops into the SoC-aware incremental reactive layer, with a warning logged once until the sensor recovers. The four existing fallback tests were updated to document the new (safer) behaviour.
• Scheduled setpoint writes are skipped when unchanged (< 1 W), matching the reactive layer.

Known limitation (unchanged, by design): the proportional correction steers the grid toward zero, so a schedule that intentionally imports/exports (e.g. cheap night charging) is partially cancelled by the correction. Fixing that properly needs a grid-plan signal from the optimizer alongside the battery schedule — happy to pick that up separately if desired.

Tests

New: closed-loop no-oscillation regression, SoC max blocks scheduled charging, SoC min blocks scheduled discharge, unchanged target not rewritten. Full suite: 243 passed.

https://claude.ai/code/session_01RUWpwxbGsgR3PoLHLq4Djz

---

Generated by Claude Code