0001 — Wiring solaredge2mqtt to the pvlearn forecast core

Status: Accepted Context: Phase 1a of pvlearn's Umsetzungsplan — Forecaster, PFISelector, ForecasterType and all encoders moved out of this repo into the standalone pvlearn package.

Decision 1: regression test compares predictive quality, not exact predictions

Decision: tests/test_pvlearn_wiring_regression.py asserts the wired-up ForecastService reproduces the frozen Phase 0 baseline's MAE (within 10%) and R² (within 0.05), not bit-identical prediction values.

Context: The test originally asserted exact equality. It passed on every local run and failed in CI. Root cause, isolated by ruling out alternatives one at a time (thread count via n_jobs/OMP_NUM_THREADS, Python 3.12 vs 3.13 — none of these changed the result on the same machine): HistGradientBoostingRegressor's greedy split search is sensitive to CPU-microarchitecture-dependent floating point rounding (SIMD reduction order). A razor-edge split threshold flips on different hardware, and the tree below it looks completely different, even with random_state=42 fixed and identical library versions everywhere.

Consequence: Bit-identical reproducibility across arbitrary hardware isn't achievable with this pipeline. Any future regression test built on this dataset/model needs a tolerance-based comparison, not exact equality. The same finding and fix apply on the pvlearn side; see chapter 6 of its Umsetzungsplan and pvlearn PR #10.

Decision 2: pvlearn's exceptions are translated back to InvalidDataException at the service boundary

Decision: ForecastService.training() and .predict() catch pvlearn.exceptions.PVLearnError and re-raise as InvalidDataException(str(error)).

Context: Forecaster.train()/.predict() used to raise this repo's own InvalidDataException directly, since Forecaster used to live here. Now it raises pvlearn's own InsufficientDataError/ModelNotTrainedError. The EventBus's listener error handling (core/events/__init__.py::_notify_listener) only catches InvalidDataException to log a warning and continue; anything else is logged as an unhandled listener error, which is a real (if non-fatal) behavior change.

Consequence: Translating at the two call sites keeps the EventBus's existing error handling untouched and behavior-identical, without teaching the EventBus about a third-party library's exception hierarchy.

Decision 3: Forecast inherits pvlearn.ForecastResult, re-decorating each aggregation property

Decision: Forecast(Component, ForecastResult). pvlearn.ForecastResult carries the energy aggregation logic (energy_today, energy_today_remaining, energy_current_hour, energy_next_hour, energy_tomorrow) as plain @property. Forecast redeclares each one as @computed_field @property, delegating to super().

Context: pvlearn's version is intentionally decoration-free — it doesn't know about Home Assistant or @computed_field. Without redeclaring the properties in Forecast, the values would still be reachable via plain attribute access but would silently drop out of model_dump()/model_dump_json(), and therefore out of the MQTT/HA discovery payload.

Consequence: The aggregation logic exists exactly once (in pvlearn); the HA-specific decoration lives only in Forecast. Verified via Forecast.model_dump() including the fields and Forecast.model_json_schema() still excluding energy_period/power_period as before.