fix(plan): split low-power charge windows at dawn, not window end

[chalfontchubby]

Summary

Fixes #4557: a charge window built from a single long cheap-rate period spanning sunrise (e.g. 03:00-10:00) was treated as one window by find_charge_rate's PV-overlap check (#4373), which abandons low power charging for the whole window the moment any PV is forecast in it - including the still-dark hours before sunrise.

• Adds calc_dawn (fetch.py): classifies each plan_interval_minutes bucket as PV-light or PV-dark, and find_charge_window now splits charge windows at that boundary (new pv_light_dark param), reusing the same mechanism already used for export windows' alternate_rate_boundary. The dark portion now stays low power throughout; the light portion keeps today's existing behaviour.
• Dawn is the first bucket that crosses LOW_POWER_PV_LIGHT_FRACTION (10%) of the peak PV forecast anywhere in the horizon - a fraction of the forecast's own peak rather than a fixed Watts figure, so the same threshold is meaningful whether the system is 2kWp or 15kWp.
• Within a calendar day it's a one-way latch: once crossed, later buckets stay light even if PV genuinely dips back under the threshold (a cloud passing), so an intermittently cloudy morning can't chop the window into several flip-flopping pieces. The latch resets at each day boundary, which also makes it behave correctly at extreme latitudes - a polar-night day never crosses (stays all-dark) and a polar-day day crosses from the first bucket (stays all-light).
• No dusk-side equivalent needed: the PV-overlap check is a forward sum from "now" to window end, recomputed every loop, so it already self-corrects once a light patch is behind it - the false-trigger-then-never-corrects failure mode this fix addresses only exists in the dawn direction.

Verified throughout against the #4557 reporter's own debug.yaml - the real 03:00-10:00 window splits cleanly at the actual crossing point, using that household's own forecast peak.

Test plan

• New test_calc_dawn covering: no forecast, all-zero forecast, within-bucket noise stability, first-crossing confirmation, cloud-dip latching, day-boundary reset, and threshold scaling with the forecast's own peak
• New find_charge_window regression test (Path J) covering the light/dark boundary split
• Full quick suite (./run_all --quick) passes
• ./run_pre_commit clean
• Replayed against the Low power charge toggle does nothing #4557 reporter's actual debug.yaml before and after - confirms the real 03:00-10:00 window now splits cleanly at the dawn crossing point

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR addresses #4557 by preventing a single long cheap-rate charge window that spans sunrise from being treated as one window for the low-power PV-overlap logic (which previously caused low-power charging to be abandoned for the entire window as soon as any PV was forecast later in it). It introduces a dawn-based PV light/dark classification so charge windows can be split at the dawn boundary, keeping the pre-dawn portion eligible for low-power charging.

Changes:

• Added calc_dawn() to derive a per-minute PV “light/dark” map based on a fraction of the forecast peak, latched per day.
• Extended find_charge_window() / rate_scan_window() to accept pv_light_dark and split charge windows at the light/dark transition.
• Added regression/unit tests covering both the dawn classifier and the window-splitting behaviour.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
apps/predbat/fetch.py	Adds calc_dawn() and uses pv_light_dark to split charge windows at the dawn boundary during rate scanning.
apps/predbat/const.py	Introduces LOW_POWER_PV_LIGHT_FRACTION constant used by the dawn classifier.
apps/predbat/tests/test_find_charge_window.py	Adds Path J regression test for PV boundary splitting and a test_calc_dawn suite.

Suppressed comments (1)

apps/predbat/fetch.py:1978

• pv_light_dark defaults to {} in the function signature. Using a mutable object as a default argument is error-prone (shared across calls if it’s ever mutated later). It would be safer to default to None and normalise inside the function.

    def rate_scan_window(self, rates, rate_low_min_window, threshold_rate, find_high, return_raw=False, alt_rates={}, pv_light_dark={}):

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[chalfontchubby]

Fixed the mutable default argument on both find_charge_window() and rate_scan_window() (the second location Copilot flagged) - alt_rates/pv_light_dark now default to None, normalised to {} inside. No call site passes None explicitly, so no behaviour change.

On the PV-boundary split timing point: leaving that as-is. The pv_boundary flag is sticky once set, so worst case the split lands up to plan_interval_minutes - 5 late when the window's own start isn't aligned to a plan-interval boundary - but that's exactly the window right at dawn, where PV output is still ramping up from near-zero anyway. A precise fix would just relocate an already-small approximation error, not remove it.