fix(plan): retain self-consumption buffer from cheapest export slot, not the peak

[springfall2008]

Problem

Reported in #4230 (and the same shape as the now-closed #4272): on Agile-style export tariffs Predbat stops exporting during the most profitable slot and resumes later on a cheaper one.

Root cause traced to optimise_detailed_pass:

• optimise_levels_pass switches every export window above the price threshold fully on. When the battery is power limited and a cheap charge is several hours away, this over-exports and forces expensive house import before the next charge, so the detailed pass must claw some export back.
• The single trim sub-pass walked price bands highest-first, so the first window it handed to optimise_export was the highest-priced peak — and shrinking it relieved the over-export. Result: the peak (e.g. a 19.67p slot) was clipped while cheaper 12–16p slots stayed fully exported. The buffer was carved out of the most valuable slot instead of the least valuable one.

Fix

Split trim into two directional, reduce-only passes:

• trim_export — runs first, price bands low→high, export windows only, reduce-only. The cheapest slots absorb the levels over-export, leaving the high-priced peak fully exported.
• trim_import — after it, high→low, charge windows only, reduce-only (the correct direction for charge: drop the priciest charging first; matches prior behaviour).

Both passes derive their iteration order locally rather than mutating price_set in place.

Result

On the #4230 debug data, the 19.67p peak is now exported for its full duration (previously clipped 5 min at the start), improving the plan metric by ~3.2p. #4272 is unchanged (that one was a metric_min_improvement_plan stickiness/config issue, not the optimiser).

Tests

• New test_trim_export — synthetic buffer scenario; verified it fails before the change (peak clipped) and passes after.
• optimise_windows_kernel (Python + C++ kernel), full --quick suite, and debug_cases all pass.
• The two debug_cases baselines that shifted (agile1, pre_saving1) are metric-identical — benign charge-window representation changes — and have been re-baselined.
• pre-commit clean.

Notes / caveats

• The trim_import half is symmetry/formalisation — the bug is entirely export-side. It is metric-neutral on every case tested but does introduce minor charge-window churn; happy to drop it and keep only trim_export if preferred.
• Still per-window greedy: this fixes the "levels over-selected, claw back from the cheap end" family (this bug). A case needing a simultaneous two-window swap with nothing over-selected is out of scope.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR adjusts the detailed optimisation phase so that when export must be clawed back to preserve a self-consumption buffer, the reduction is taken from the cheapest export slots first (keeping the most valuable peak fully exported). It introduces distinct export vs import trim passes and adds a regression test plus updated debug-case baselines.

Changes:

• Split the prior trim logic into trim_export (low→high, export-only, reduce-only) and trim_import (high→low, charge-only, reduce-only), each deriving its own iteration order without mutating price_set.
• Add and register test_trim_export to validate that the peak export slot is not clipped when trimming is needed.
• Update expected outputs for impacted debug cases (agile1, pre_saving1).

Reviewed changes

Copilot reviewed 3 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
apps/predbat/plan.py	Splits trim passes and changes price-band iteration order to trim export from cheapest first.
apps/predbat/tests/test_trim_export.py	Adds regression test scenario for export-trim ordering / buffer retention.
apps/predbat/unit_test.py	Registers the new trim_export test in the test runner.
coverage/cases/predbat_debug_agile1.yaml.expected.json	Updated debug-case expected output after optimiser behaviour change.
coverage/cases/predbat_debug_pre_saving1.yaml.expected.json	Updated debug-case expected output after optimiser behaviour change.

Comments suppressed due to low confidence (1)

apps/predbat/tests/test_trim_export.py:109

• The test only considers a cheaper window “reduced” if it is disabled (100) or start-clipped. The optimiser can also retain buffer by raising the export SoC limit (shallower discharge) or selecting freeze (99) without clipping the start, which would incorrectly fail this test.

        window = my_predbat.export_window_best[n]
        if my_predbat.export_limits_best[n] >= 100 or window["start"] > (mn + 30 * n):
            cheaper_reduced = True
            break