fix(inverter): stop re-committing a stable export window every cycle (Solis)

[chalfontchubby]

Fixes #4709

Problem

On Solis GS_fb00 (solax_modbus), pressing the charge/discharge update button zeroes the timed charge/discharge current registers. The reporter measured the current register being knocked to 0 A at :xx:22 on every 5-minute boundary throughout an active export window, stopping export until something restored it.

The cause is in adjust_force_export. For H M time format the discharge slot-time registers are rewritten unconditionally every cycle — a deliberate write-reliability workaround from #1529. #4000 then reused that same "we wrote something" flag (changed_start_end) to decide whether to press the update button, so the button fired every cycle even when the window had not changed.

#4000's stated premise was that adjust_charge_window "always re-writes time entities for H M format → button press fires every cycle". The write half is correct; the button-press half is not — the charge side's press is gated on a plain dirty check and always has been. So the every-cycle press was never symmetric with anything.

The same flag also gates the 30 second GivTCP settle sleep in adjust_inverter_mode and the export-slot notification, both of which were therefore firing every cycle too.

Fix

Separate "we rewrote the registers" from "the schedule needs committing", and gate the press, the sleep and the notify on the latter.

Tracking the last committed schedule keeps a stable window quiet while still committing once after a restart, when nothing has been committed yet — so #4000's own regression test passes unchanged. The commit-once safety net is scoped to the H M path; every other format already commits on a real change alone.

A commit is only recorded once it has actually succeeded. press_and_poll_button returns a success flag, and discarding it would mean a failed press was recorded as done and never retried — the schedule sitting uncommitted until it happened to change on its own. The behaviour being replaced retried implicitly, because changed_start_end fired every cycle, so gating on the real result preserves that retry without bringing back the every-cycle press. (Same failure shape as the known call_service_template issue, where a service call is marked done before its success is checked.)

The register writes themselves are untouched.

Also fixes idle-cycle presses on plain GS

Beyond the export window, this PR also stops a second symptom that only came to light during triage of #4712.

GS (pre-FB00 Solis) is charge_time_format: "H M", so changed_start_end is set unconditionally there too — which means on current main it presses the update button on every idle cycle as well, all day, not just during export. Dropping changed_start_end from the commit condition fixes that. Measured over four idle adjust_force_export(False) calls:

             GS      GS_fb00
main         4/4     4/4
+#4711       1       4/4     <- this PR
+#4713       1       1

GS_fb00 and the cloud types reach the same press by a second, independent route (a None time comparison) which this PR does not touch — that is #4712, fixed by #4713 stacked on top of this one.

Also included

press_and_poll_button now presses only the side that changed. Where separate charge_update_button / discharge_update_button entities are configured, both were pressed whenever either side had something to send — unchanged since the original split-button support. Halves the presses for those users; no effect on the combined-button setups the Solis template ships by default. This was written for #2328 and parked as too small to be worth landing alone.

Commits

9a3a0e23	press_and_poll_button only presses the button(s) that changed
be3d225f	commit a stable export window once, not every cycle
4499357c	only record an export commit that actually succeeded

Testing

• New regression test: a stable export window commits once, stays quiet on subsequent cycles, and commits again on a genuine change. Verified to fail without the fix.
• New regression test: a failed button press is retried on the next cycle rather than recorded as committed. Verified to fail without 4499357c.
• Existing Fix Solis GS_fb00: always write discharge time entities and press button during force export #4000 regression test passes unchanged.
• Existing adjust_force_export1 passes unchanged, which is what pins the GE path's behaviour.
• Side-scoping test from the parked work included.
• Full --quick suite green.

Not yet confirmed against a live install — logs requested on #4709.

🤖 Generated with Claude Code