[frahlg]

Built on #1852 (surplus on a charger with no step table), now merged. This PR is one commit on master. It touches pickSurplusSteps and surplus3PhaseSteps, which #1852 also changed.

Fixes the site rig's known failure s5s-25: "auto stays on one phase at 9 kW surplus". s5-25 (no step table) is #1852's.

Cause

The rig's loadpoint is surplus-only, phase mode auto, with Easee's steps. It has 9.3 kW of surplus, and its plan shows no PV.

1. Tick 1: the live surplus is above the three-phase minimum and there is no choice on file, so pickSurplusSteps picks three phases.
2. Tick 2: the choice is on file, so the live check is skipped. The near-term dwell holds three phases. The day check then reads the plan's peak, 0 W, and sets the day's one-phase lock. Nothing ends that lock before midnight.
3. After 30 min the near-term dwell would have gone to one phase on the plan's 0 W as well.

A second fault sits under it. surplus3PhaseSteps counted 3,680 W, which is 16 A on one phase, as a three-phase step. So the three-phase minimum in the phase choice was 3,680 W, not 4,140 W.

Change

• The surplus measured now is a floor under both forecast peaks. Each forecast window starts now, so its peak cannot be below what the meter shows now. The reading counts only up to the average the pause gate already keeps. One high reading then cannot switch to three phases, and a real drop still shows at once.
• The day's one-phase lock gives way to measured surplus above the three-phase minimum once it has held for 30 min (phaseSwitchMinHold). Three phases then hold for 30 min too.
• The lock waits for a three-phase choice to hold. On master a cloud right after a three-phase choice set the lock at once.
• Three phases start at 6 A each, as PlanningSteps counts them for planned charging. surplus3PhaseSteps keeps fix(loadpoint): surplus orders any power on a charger with no step table #1852's branch for no table and 1p. PhaseSplitW still tells an auto driver where to split. It no longer sorts the table, and its doc comment says so.
• No step table: fix(loadpoint): surplus orders any power on a charger with no step table #1852 returns first, clears an old one-phase lock and records three phases, so none of the above reaches such a charger. An earlier commit here set a no-table threshold; it went when fix(loadpoint): surplus orders any power on a charger with no step table #1852 added that return. fix(loadpoint): surplus orders any power on a charger with no step table #1852's selectPhaseLocked3P records three phases once; this PR's selectPhaseLocked records either choice with its time, which the near-term dwell needs.

Smaller effects: a one-phase charger left on an empty mode, and a table on a one-phase site, now pause below their smallest step (1,380 W), not below 3,680 W. A bat-SoC unlock (SurplusOnly false) still never sets the day lock.

Before and after

Check	#1852 (base)	this branch
TestSurplusAboveTheThreePhaseMinimumRunsThreePhases: 9.3 kW, plan without PV, 45 min	order 1 of 540 is 3,680 W on 1p	every order 3p, 8,970 W, past the 30 min hold
TestSurplusBelowTheThreePhaseMinimumRunsOnePhase: 3.8 kW (rig s5t)	first order 4,140 W on 3p, which imports	1p throughout
TestSurplusOnePhaseYieldsAfterTheHold: 2 kW, then 9 kW, then 2 kW; dwell and day lock	1p all day	1p, 3p after 30 min, 1p after 30 more
TestSurplusOneHighReadingKeepsOnePhase: 1.5 kW and one 4,300 W reading	1p	1p (a guard for the floor)
TestSurplusOnePhaseOfferStaysWithinTheCharger, three-phase case	asserted 1p at 3,680 W: the s5s behaviour	3p at 8,970 W; the 16 A per phase check stays
Field matrix (342 cases)	known rows	same known rows; no change to sitesim_field_known.txt

With the window average alone as the floor, timing/25A/auto/mixed/fast2s/car3p/L1/surplus failed idle. Export fell from 9 kW to 3 kW, the average held three phases for 8 s, the pause gate paused, and its 35 s hold kept the car idle. Taking the lower of the reading and the average keeps that row green.

go/test/site/scenarios_test.go: deleted s5s-25's known line, since the rig fails a known criterion that passes. #1852 already deleted s5-25's. The rig passed s5s-25 on 4ad0798, this commit on an earlier #1852 base, with both planner engines. Later rebases changed only the lines next to #1852's no-table return.

go vet and go test for ./internal/..., ./cmd/ftw/... and ./test/site/ pass, the field matrix among them (342 cases, the known rows unchanged). go test -race passes for loadpoint and control. Since #1852 merged, this PR targets master.

An independent review (Sonnet) found no blocker. It found that one reading chose the phase, fixed with the reading-and-average floor. A second look after the rebase found nothing in this diff.

Open

• With no PV in the plan, one low reading at the end of a hold still starts a 30 min one-phase hold. Master stayed on one phase all day there.
• Not run on real hardware. Hardware evidence is still needed: surplus on auto with an Easee above and below 4.1 kW.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-25T14:24:13.401672Z	8d6d03b	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.