fix: prevent false-positive gap detection on zero-consumption periods

[mgazza]

Summary

• Fixes false-positive gap detection in previous_days_modal_filter() that injected phantom load (~6 kWh/night) during legitimate zero-consumption overnight periods
• Tracks sensor data point provenance during minute_data() processing via a new data_point_minutes set
• In the gap detector, checks whether the sensor was actively reporting during each gap. If online (≥1 data point/hour), skips filling. If offline, fills as before.
• Fully backward-compatible: existing callers unaffected, no attribute = old behavior

Root Cause

The gap detector checks data[m] == data[m+5] to find missing data. After clean_incrementing_reverse(), zero-consumption periods have equal consecutive cumulative values. The detector falsely flags these as "gaps" and fills them with average_day / 1440 kWh/minute of synthetic consumption.

Test 10 proves this: 8 false gaps per 8 days, 3360 minutes (29.2% of data) incorrectly filled.

Approach: Data Point Provenance

Rather than changing the gap detection heuristic (which serves a legitimate purpose for sensor-offline periods), we record which minutes had actual sensor readings during minute_data(). The gap filter then distinguishes:

• Sensor online, zero consumption → skip filling (correct)
• Sensor offline, no data → fill with average (correct, existing behavior)

Files Changed

File	Change
utils.py	Add data_point_minutes param to minute_data()
fetch.py	Init set, pass through minute_data_load() + download_ge_data(), filter in previous_days_modal_filter()
tests/test_previous_days_modal.py	Tests 6-8: sensor active → skip, sensor offline → fill, backward compat
tests/test_fill_load_from_power.py	Test 11: e2e verification all 8 false-positive gaps filtered out

Test plan

• All existing tests pass (previous_days_modal, fill_load_from_power, minute_data, minute_data_load, get_now_cumulative)
• Test 6: Overnight zeros with sensor data points → gaps skipped (12 kWh stays 12 kWh)
• Test 7: Gap with no sensor data points → gap correctly filled (12 → 18 kWh)
• Test 8: No load_data_point_minutes attribute → backward compat (gaps filled as before)
• Test 11: End-to-end proof all 8 false-positive overnight gaps filtered out

Supersedes #3546

🤖 Generated with Claude Code