SAFETY-73: a battery that refuses, per inverter, with two or more

SAFETY-52's hold applied per unit at N>1 (owner, 2026-09-27, option R).
A unit that gives < 100 W for 12 cycles in one direction while asked
> 200 W, or trips its stale-reading gate on a ~0 W reading, gets a
P_SAFETY `refused` unit limit closing that direction, so the allocator
hands its share to the others within their own limits. Probe: a +-200 W
pin every 15 min, one unit at a time, only while the site asks 200 W that
way. All units refusing: the site law holds 0 W. Exits reseed the site
target from the wire. INFO plus one WARNING (REFUSES_ABOVE_CUT_OUT) above
the configured cut-out; SAFETY-42/43 no longer report a refusing unit.
N=1 unchanged.

test_refusal_units.py: cases a/b/c on flat and jitter registers against
the old behaviour, probes, recovery, exits; mutations U1-U13 all red.
TEST-19 no-fault campaign: 0 engagements in 5280 runs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

goodwe_controller/DOCS.md

@@ -295,7 +295,7 @@ When the add-on refuses to start, each reason names the setting by its label and
 - **Stuck meter alarm after** (`meter_freeze_s`): measured, not chosen (SAFETY-18). Over 2399 s of this house at its operating point, the longest run without a new reading is 15.4 s, and 300 s is 19.5 times that. It raises an alarm only and does not set 0 W.
 - **Opposing inverters threshold** (`circulation_w`): measured on 924 closed-loop test runs (SAFETY-31). The worst legitimate opposition drops from 505 s at 100 W to 5 s at 400 W, the lowest value accepted. 700 W is 1.75 times that knee, with margin because the second inverter is still a copy of the first.
 - **Opposing inverters time** (`circulation_hold_s`): derived (SAFETY-31). 75 s is fifteen control cycles, 15 times the worst legitimate episode, and 2.5 times the control board's 30 s link limit.
-- **Not delivering alarm after** (`delivery_hold_s`): two or more inverters only. It reports only (SAFETY-42): nothing is commanded differently. Derived from 2640 closed-loop test runs with nothing faulty. The worst legitimate episode is 35.0 s and 185 s is 5.29 times that. The schema allows 30 s, below that, so expect false reports there. A battery at or below its own cut-out level is never reported. Re-derive with admin/goodwe `sim/delivery_window.py --campaign --blag`.
+- **Not delivering alarm after** (`delivery_hold_s`): two or more inverters only. It reports only (SAFETY-42): nothing is commanded differently. Derived from 2640 closed-loop test runs with nothing faulty. The worst legitimate episode is 35.0 s and 185 s is 5.29 times that. The schema allows 30 s, below that, so expect false reports there. A battery at or below its own cut-out level is never reported, and neither is one that gives under 100 W (SAFETY-73: it is a battery that refuses). Re-derive with admin/goodwe `sim/delivery_window.py --campaign --blag`.
 - **Shadow mode** (`shadow_mode`): the whole decision path runs and nothing reaches the inverter. The status page and action log show what would have been sent. A safety stop that is already up stays up. See [Shadow mode](#shadow-mode-shadow_mode).
 - **Control gain** (`gain`): measured on real hardware and already at the limit.
 - **Largest change per cycle** (`slew_w`): 0 means follow the rating: about a ninth of it, derived (SAFETY-39) from a measured ~18 s step response against a 2 s meter tick. A fixed value faster than the rating allows is refused. At 1000 W against a 5000 W rating control swung the meter from 3.4 kW export to 2.8 kW import on 2026-09-05.
@@ -550,15 +550,23 @@ control is stopped or started. 200 W is below the 250 W the stale-reading check
 needs, so a test never trips it. A battery that gives some power, even 100 W,
 is not refusing and keeps its request.
 
-⚠️ **One inverter only.** With several inverters this hold is off. There, the
-battery level limit moves an empty battery's share to the others, but only if
-its “Battery cut-out level” matches where the battery really stops. A battery
-that refuses at another level, or for another reason (cold, a battery fault),
-can still hold its last request until the grid starts to export.
+**With two or more inverters** the same rule works per inverter. An inverter
+whose battery gives less than 100 W for 12 cycles while asked for more than
+200 W is asked 0 W, and the other inverters take over its share, each up to
+its own limits. The status page names it: *inverter-a: the battery refuses to
+give power, the others take over*. It is tested at 200 W every 15 minutes, one
+inverter at a time, and only while the site asks at least 200 W that way.
+It still accepts charge, and a minute of charging ends the refusal. When every
+inverter refuses the same way, the whole site asks 0 W as with one inverter.
+If the battery refuses while its level reads above its “Battery cut-out level”,
+the log adds one warning ending in `REFUSES_ABOVE_CUT_OUT`: check that setting.
+A battery that refuses is not reported as *not delivering*.
 
-*For installers:* SAFETY-52. The event log reads `BATTERY REFUSES …` and
-`BATTERY REFUSES ENDED: …`, and the INFO line ends in `BATTERY_REFUSES`.
-`DECISIONS/SAFETY-52-a-battery-that-refuses.md` has the replay figures.
+*For installers:* SAFETY-52 and SAFETY-73. The event log reads
+`BATTERY REFUSES …` and `BATTERY REFUSES ENDED …` (with the inverter's name at
+N>1), and the INFO line ends in `BATTERY_REFUSES`.
+`DECISIONS/SAFETY-52-a-battery-that-refuses.md` has the replay figures and
+`DECISIONS/SAFETY-73-refusal-per-inverter.md` the two-inverter ones.
 
 ⚠️ **`Loop reason` is not the `Why` row on the status page, and the two will
 often disagree.** `Why` is the *arbiter's* answer in words — *Keeping the grid