Sigenergy: stop Predbat and Axle fighting over the inverter, and make axle_control work

[mgazza]

Supersedes #4455 and #4456, which were opened as competing alternatives. A review showed they were not actually alternatives — see #4454 for the reasoning — so this is the two of them combined, with the review findings fixed.

Fixes #4454.

The problem

A Sigenergy accepts one controller at a time. VPP mode (Predbat) and the NorthBound Interface (Axle's dispatch channel) are mutually exclusive. Two things follow, and both are broken:

1. Where Predbat owns the inverter, it wins by accident. _manage_vpp_registration() ran on the 5-minute poll and silently switched the system back to VPP, so an Axle dispatch could hold it for up to 5 minutes before being displaced, logged only as "controls skipped until onboard is approved".

2. Where Axle should own it, axle_control does not work at all. Fetch.fetch_config_options() raises set_read_only_axle for exactly this purpose, but it was only consumed in execute.py and output.py. sigenergy.py read the user-facing switch.predbat_set_read_only instead, so it never learned Axle had the floor.

Observed live on 2026-08-06: Axle took the inverter into Northbound Integration at 19:35:18; it was back in VPP at 19:40:12 — exactly one SIGENERGY_POLL_INTERVAL.

Make axle_control work here

• _axle_has_control() evaluates axle_control and the Axle event sensor live, via fetch_axle_active(), rather than reading the cached set_read_only_axle. That flag is only refreshed by the 5-minute prediction loop and is still False from reset() when this phase-1 component makes its first=True run — precisely the case that matters, a restart during a live event, where the cached flag would have had Predbat reclaim VPP and kill the dispatch. It also removes an up-to-5-minute lag detecting event start and end.
• During an event the operating mode is left exactly as it is. If Axle has moved the system to NBI it stays there; if the event has started but Axle has not switched yet, do not drop to MSC either — that hands control to the owner's app rather than to Axle.
• Battery commands are suppressed for the same window.
• VPP is reclaimed when the event ends, and that is logged.

And where Predbat does keep ownership, make it deliberate

• Reclaim VPP every minute instead of every 5. set_operating_mode() is an MQTT publish that only fires when the mode is actually wrong, so this costs nothing against the REST rate limit.
• Log the reclaim naming the controller being displaced.
• Record last_contended_by on the status sensor, published on the same minute cadence and never cleared. Contention is usually shorter than one publish interval, so a marker reset on recovery would almost never be seen; "has this happened" is the useful signal, and in_vpp already answers "is it happening now".
• Stop reporting a system in NBI as pending_approval. Downstream UIs render that as an amber "approve this in the Sigenergy app" banner, so every Axle event told an already-onboarded user to approve something that needed no approval.

Compatibility

With axle_control unset — which is the current state of every system we operate — behaviour is unchanged except for the faster reclaim, the clearer logs, and the status fix. Nothing here silently changes who drives the battery.

Note that axle_control is a global option: it gates self.set_read_only, which covers the whole inverter-write path for any inverter type, not just Sigenergy. Enabling it is a separate, wider decision than merging this.

Tests

11 tests in tests/test_sigenergy.py, including two explicit regression tests for the review findings: axle_standoff_survives_restart (the stale-flag case) and contention_marker_published_before_recovery (the marker being cleared before it was ever published). Full Sigenergy suite 78 pass / 0 fail; unit_test.py --quick passes with 0 failures; run_pre_commit clean.

One test-infra note: MockSigenergyAPI.get_arg() had to accept **kwargs, because fetch_axle_active() passes indirect=False and the mock's two-argument signature rejected it. The mock had drifted from ComponentBase.get_arg; this widens it rather than fully realigning it.

What this does NOT do

Two honest limits, raised in review:

• The faster reclaim does not make Predbat and Axle coexist. With axle_control unset it converts a five-minute contest into a one-minute contest and makes Predbat win deterministically. That is the intended ownership policy, not a resolution of the conflict.
• The stand-down is best-effort mutual exclusion, not a dispatch protocol. During an event the mode is left untouched and commands are suppressed, but a battery command issued just beforehand can carry a duration of up to 720 minutes (apply_controls()), and nothing revokes it. If Axle moves the system to NBI that command stops applying; if Axle does not, it can remain active with no acknowledgement from either side that control changed hands. Enable axle_control selectively and watch the actual mode telemetry.

Worth stating plainly for context: load_axle_slot() adds pence_per_kwh to rate_export and adjusts load scaling. An Axle session is a price signal, not a required delivery — the optimiser may still decline to export on SOC, reserve or economics. Neither this PR nor axle_control changes that.

[Copilot]

Pull request overview

This PR updates the Sigenergy integration to handle single-controller contention explicitly: Predbat now either (a) deliberately re-asserts VPP ownership on a 60s cadence when it should own the inverter, or (b) stands down cleanly during live Axle events when axle_control is enabled, leaving operating mode untouched and suppressing battery commands.

Changes:

• Add live Axle-event evaluation (fetch_axle_active) to make axle_control effective immediately (including across restarts) and suppress controls during events.
• Reclaim VPP ownership every minute (instead of every poll interval) and log/record contention via a persistent last_contended_by marker; treat NBI contention as “active” (not “pending_approval”).
• Expand Sigenergy unit tests with regression coverage, and update the Sigenergy mock get_arg() signature to accept **kwargs to match ComponentBase.get_arg.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
apps/predbat/sigenergy.py	Implements minute-cadence VPP reclaim, Axle stand-down (axle_control) via live sensor evaluation, improved contention/status reporting, and control suppression during events.
apps/predbat/tests/test_sigenergy.py	Adds regression tests for contention/marker publication and Axle stand-down behavior; updates test mock get_arg() to accept broader kwargs.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.

Suppressed comments (2)

apps/predbat/sigenergy.py:2231

• When a system is offboarded, _manage_vpp_registration() returns before clearing _axle_standoff_logged. If the offboard switch is enabled mid-Axle event, the stand-down flag can remain set and later produce a misleading “Axle VPP event ended — resuming control…” log when offboard is turned off (even though the event may have ended earlier). Clearing the flag on the offboard path avoids stale state/logging.

        if is_offboard:
            return False

apps/predbat/sigenergy.py:2212

• The docstring says an active Axle event under axle_control “takes priority over every case above”, but the implementation returns early for is_offboard before checking Axle control. Either move the Axle check above the offboard return, or clarify in the docstring that offboard still short-circuits all mode management.

        An active Axle event under the ``axle_control`` option takes priority over every
        case above: Predbat stands down and leaves the operating mode untouched so Axle can
        drive the battery through the NorthBound Interface.