TOU write failures silently marked as applied; dashboard crashes on unavailable battery_soc

[johanzander]

Summary

Debug bundle bess-debug-2026-07-22-120026.md (reporter restarted the UI and
saw GRID CHARGING / Battery First / Grid Charge ON for the current
period, but the physical inverter's TOU period list had no period covering
that time window — screenshots confirm the mismatch).

Root cause 1: failed TOU hardware writes are silently treated as success

GrowattMinController.write_schedule_to_hardware (core/bess/growatt_min_controller.py)
catches each per-segment hardware write exception internally and only logs
it — it never re-raises. BatterySystemManager._apply_schedule
(core/bess/battery_system_manager.py:2481-2507) already has retry logic
built for exactly this case (self._hardware_write_pending = True on
exception, forcing a retry next cycle), but it never sees the exception
because it was swallowed one layer down. So:

• 2026-07-22 11:30:13 — write of TOU segment 6 (15:30-15:44) is attempted.
• 2026-07-22 11:30:55 — fails after 4 retries with 500 Server Error from
  growatt_server/update_time_segment (HA core itself was returning 500s
  intermittently most of the day — external instability, not a BESS Manager
  bug).
• growatt_min_controller:1316 - FAILED: Failed to update TOU segment is
  logged, but write_schedule_to_hardware returns normally.
• _apply_schedule's except never fires, _hardware_write_pending is
  incorrectly cleared to False, "Schedule applied successfully" is
  logged.
• Because the in-memory controller (holding the intended schedule) had
  already been swapped in before the write (by design, so strategic intents
  stay available across a write failure), and schedule-diffing compares
  against this same in-memory intended state rather than a hardware
  readback, the failed write is now indistinguishable from a successful one
  on every later comparison — it's never retried, and the dashboard shows
  the intended schedule as fact.

Root cause 2: dashboard/inverter-status endpoints crash on unavailable battery_soc

Same debug bundle shows repeated Error generating dashboard data: unsupported operand type(s) for /: 'NoneType' and 'float' and Error getting inverter status: ... whenever sensor.battery_soc goes
unavailable (same HA-instability window).

• backend/api_dataclasses.py:953 — battery_soe = (battery_soc / 100.0) * battery_capacity — battery_soc comes from
  controller.get_battery_soc() (core/bess/ha_api_controller.py:1129),
  which returns None when the sensor is unavailable.
• backend/api.py:1543 — same pattern in get_inverter_status.

Both endpoints return an opaque 500 with a TypeError message instead of a
clear "sensor unavailable" error.

Fix

1. write_schedule_to_hardware: collect per-segment failures and raise once
   after attempting all segments, so the existing retry plumbing in
   _apply_schedule actually engages.
2. Raise a clear error before the division in both call sites when
   battery_soc is None, instead of an opaque TypeError.

[Frank-Leysen]

Upgraded to 9.9.0b23 — live test (Growatt MOD 5000TL3-XH, solax_modbus, Luminus Dynamic/Belpex H)

From @Frank-Leysen.

Installed 9.9.0b23 this morning (~06:18–06:20 CEST). System looks normal after the update (SOC ~24%, continuing overnight Load Support → solar day ahead).

Debug after upgrade (06:20, b23): https://gist.github.com/Frank-Leysen/41682b31b6ea4434bd302132c8f66ff5

Debug just before upgrade (06:12, still b22): https://gist.github.com/Frank-Leysen/98dfdbfd33cd192d0dd30e952c6e74a3

We grabbed both so you can compare if useful for #365 (TOU write failure / retry / dashboard vs inverter drift) or general regression checks.

Context for yesterday (grey day, as expected):

• Meter: import 0.8 kWh / 0.30 EUR, export 1 kWh / −0.10 EUR → net energy −0.2 kWh but net cost +0.20 EUR (buy >> sell on the small volumes)
• Overnight: again 2× short grid-import spikes during hot-water boiler reheat while LOAD_SUPPORT discharge was throttled (same bug: low-rate BATTERY_EXPORT writes grid_first at a forced sub-load discharge rate, causing avoidable grid import on any load spike #352 pattern; separate from this release)
• Today forecast Net Grid Cost ~−1.80 EUR if PV cooperates

Note: #337 (SolaxController Min SOC write) doesn’t apply to our Growatt path — we upgraded mainly for the #365 reliability fixes. Continuing live test on b23; evening export / TOU apply will be the next interesting check.