fix(fetch): split charge windows at dawn again, unify low-power PV thresholds

[chalfontchubby]

Summary

Fixes #4699. self.pv_forecast_minute was reset to {} at the top of fetch_sensor_data(), but calc_pv_light_dark()/calc_dawn() still ran before the line that actually repopulates it - so the dawn split (#4557) always saw an empty forecast and never fired, for any combine_charge_slots user, since the feature first shipped. Fixed by moving the populating fetch_pv_forecast() call earlier.

Separately, the debug-replay test harness's own rate-rescan reimplementation (tests/test_single_debug.py) never passed pv_light_dark through to rate_scan_window() at all - an independent gap that meant no debug.yaml replay could exercise this path even after the production fix. Both are fixed here, and together they let the real fix be verified against the reporter's own debug.yaml: the window now splits at dawn with the pre-dawn slice going to low power instead of full rate.

Threshold unification

While investigating, found the two low-power PV thresholds were independently tuned and could disagree on genuine twilight PV:

• LOW_POWER_PV_LIGHT_FRACTION - a % of this forecast's own peak, used to decide where to split a window at dawn
• LOW_POWER_PV_THRESHOLD - a fixed kWh figure, used to decide whether to abandon low-power charging for a window

A %-of-peak threshold is also unreliable on a heavily overcast day, since that day's own peak is low too - the same % ends up being a much lower absolute wattage than on a clear day, and Predbat has no declared panel capacity to normalise against instead (only inverter_limit, which caps the inverter's own output, not the array's rating).

Replaced both constants with a single new config item, low_power_pv_threshold_w (absolute Watts, default 150W), used consistently by:

• calc_dawn (fetch.py) - the split boundary
• find_charge_rate (utils.py) - the abandon decision, now comparing average power over the remaining window against the threshold rather than accumulated energy against a fixed kWh figure, so a long window at a low constant trickle no longer creeps past it regardless of window length (the actual twilight-creep failure mode this was hiding)

Also added binary_sensor.predbat_dawn (on past dawn / off before or unclassified) and updated docs for both the sensor and the new config item.

Verification

• Root cause confirmed via git blame + reading (deterministic, not probabilistic): the reset and the read are in the same function, no path repopulates in between.
• The average-power fix verified as a genuine regression: temporarily reverted to the old fixed-kWh comparison and confirmed the new low_power_below_threshold_stays_low_power test fails without it (13.62kWh -> 14.2kWh final SoC), passes with it.
• One random regression scenario (seed 15) shifted as an expected consequence of the behaviour actually changing - baseline regenerated and confirmed it's the only one that moved.
• Full run_pre_commit clean.

Test plan

• ./run_all --quick passes
• ./run_all --test find_charge_window --test find_charge_rate --test model - new/updated scenarios pass
• ./run_all --test random - only seed 15 changed, as expected
• ./run_pre_commit passes (ruff, black, cspell, markdownlint, full test suite)
• Verified against reporter's own debug.yaml that the dawn split now fires and produces low-power pre-dawn charging

🤖 Generated with Claude Code

[Copilot]

🟡 Changes recommended

Threshold boundary handling, configuration visibility, dawn sensor state, and production-order test coverage need correction.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Fixes dawn-based charge-window splitting and unifies low-power PV thresholds.

Changes:

• Fetches PV forecasts before calculating dawn splits.
• Adds a configurable 150 W PV threshold and dawn sensor.
• Updates low-power behavior, tests, documentation, and regression baseline.

File summaries

File	Description
apps/predbat/config.py	Adds the PV threshold setting.
apps/predbat/const.py	Removes superseded constants.
apps/predbat/execute.py	Passes the threshold during execution.
apps/predbat/fetch.py	Reorders PV fetching, calculates dawn, and publishes its state.
apps/predbat/output.py	Applies the threshold when displaying charge rates.
apps/predbat/predbat.py	Initializes dawn classification state.
apps/predbat/prediction.py	Propagates the threshold into predictions.
apps/predbat/utils.py	Uses average PV power for low-power decisions.
apps/predbat/tests/test_find_charge_rate.py	Updates PV overlap tests.
apps/predbat/tests/test_find_charge_window.py	Tests absolute-threshold dawn detection.
apps/predbat/tests/test_model.py	Adds a low-PV regression scenario.
apps/predbat/tests/test_single_debug.py	Includes dawn splitting during debug replay.
coverage/cases/random_results.json	Updates the affected random baseline.
docs/customisation.md	Documents threshold behavior and configuration.
docs/output-data.md	Documents the dawn sensor.

Review details

Suppressed comments (1)

apps/predbat/fetch.py:1643

• A configured threshold of 0 W makes every all-zero bucket satisfy average >= 0, so an empty-production day is classified as light from midnight and the dawn sensor reports on. Since 0 is an allowed value, require some positive PV before latching light; this preserves the useful meaning “any non-zero PV” without treating darkness as dawn.

        light_threshold = self.low_power_pv_threshold_w / MINUTE_WATT

• Files reviewed: 14/15 changed files
• Comments generated: 4
• Review effort level: Balanced

---

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

[springfall2008]

No outstanding feedback or CI failures to address: all four review threads from the 2026-08-25 Copilot review were replied to and fixed in df1dd0f (2026-08-28) — the enable gate was removed from low_power_pv_threshold_w, the max-rate/abandon comparison was unified to >= with a pv_window_kwh > 0 guard, and the dawn sensor publish was moved into an unconditional fetch_pv_forecast_and_dawn() with ordering coverage in test_fetch_pv_forecast_and_dawn. Both CI checks (kernel-binaries, pre-commit) passed on that commit; no review activity since.