ARCH-02: assert the invariant, because a fixture cannot enumerate arrangements

Review approved and then attacked part 5b, the control that stands between B1's defect and a green suite. It got through in three edits: A1 delete `END LEG` 5b CATCHES it 2 FAIL A2 ...and reorder the legs 5b BLIND 1 FAIL A3 ...and reword the closing summary 0 FAIL, "all checks passed" with a splitter that closed ZERO legs ⚠️ TWO OF THOSE THREE ARE ORDINARY REFACTORS - reordering report legs, rewording a summary line - and neither reads as touching a safety check. 5b caught the unclosed splitter only because the clean leg happened to be printed LAST, and the backstop that saved A2 was a hardcoded transcription of the tool's own sentence, so rewording that sentence blinded it too. FIXED BY ASSERTING THE STRUCTURE RATHER THAN ADDING A FIXTURE: a leg may not open while one is open, and every leg that opens must be explicitly closed. Driven - deleting `END LEG` now fails on the FIRST edit, whatever the order or the wording: AssertionError: leg 'SAME-STEP' opened while 'STAGED' was still open - the tool stopped closing its legs, and a splitter that never closes swallows the closing summary into the last leg. All three attacks include that edit, so one assertion kills the class. Clean tree unchanged: 28/28, all checks passed. ⭐ AND THE RULE IN MUTATION.md IS EXTENDED, because review's finding is the general case of it. It said "a per-thing check that has never seen the things DISAGREE is a total". Part 5b DOES make them disagree, every run, and was still blind. So: ...AND ONE THAT HAS ONLY SEEN THEM DISAGREE IN ONE ARRANGEMENT IS A TOTAL OVER THAT ARRANGEMENT. When a fixture cannot enumerate the arrangements, assert the invariant they share. Also recorded, NOT fixed, with the reason: the reference carries `entity_id unique_id device` and grades all three, but never the discovery TOPIC. Review measured `homeassistant/... -> ha/...` exiting 0 with no findings while every announced entity dies. `discovery_payloads()` already returns the topic and `announced_rows` discards it, so a fourth column closes it - but the prefix is HA's fixed contract rather than a name this project chooses, so it is not what a tidy-up rename reaches. Lower LIKELIHOOD than the N1 gap it resembles, not lower consequence, and it stops being a note the day the topic becomes configurable. Two stale figures corrected: the tool claimed five migration mutations where there are nine, and pointed at a DOCS.md heading that does not exist (the real one is "Before you commission inverter B", verified present at DOCS.md:1577). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BwRLw7eKYj2byVdwB6kbjU
ARCH-02: M7's evidence was filtered out by its own finding kind
2026-09-08 15:01:55 +02:00 · 2026-09-08 14:23:35 +02:00 · 2026-09-08 14:21:34 +02:00 · 2026-09-08 13:23:03 +02:00 · 2026-09-08 13:20:49 +02:00 · 2026-09-08 13:16:47 +02:00
26 changed files with 4734 additions and 230 deletions
@@ -46,8 +46,10 @@ threshold. Measured: a wedged inverter goes from named on 199 of 200 cycles to
 **0 of 200** at a load its sibling can carry alone. The alarm returns once the
 healthy pack saturates - swept, that is a 5000 W load on the fixture, **below**
 the proportional-regime formula's 6500 W and not above it. ⚠️ **N=1 IS
-UNAFFECTED**: `allocator.plan()` returns before weights are read at all. The
-closure is card SAFETY-43.
+UNAFFECTED**: `allocator.plan()` returns before weights are read at all.
+⚠️ **CLOSED BY SAFETY-43 ON THIS BRANCH** - see its section below; the regression
+above is history, and the record of it is kept because the measurement that
+found it is what the fix is graded against.

 **PKG-09** adds `tools/predeploy_check.py`. ⚠️ **RUN IT BEFORE DEPLOYING THIS
 RELEASE** - it is what caught the fact that 0.8.0 had five runtime commits
@@ -95,6 +97,107 @@ what happens when nobody does.
 | | |
 |---|---|
 | _this branch_ | UX-02 - a configuration page in the ingress panel, where the validator's refusals can be seen |
+| _this branch_ | UX-04 - UX-01's three unmet acceptance criteria: one withdrawn, one built, one re-attributed |
+| _this branch_ | TEST-21 - `delivery_hold_s`'s floor re-derived from a harness that is now committed, and the recorded 30.0 s did not reproduce |
+| _this branch_ | SAFETY-43 - delinquency measured against the ask as well as against a flat 500 W, closing OPT-09's disclosed detection regression |
+
+⚠️ **SAFETY-43 CLOSES THE DETECTION REGRESSION OPT-09 DISCLOSED, AND IT IS A
+TWO-INVERTER FIX. N=1 IS BYTE-IDENTICAL.** `delivery.delinquent()` named a unit
+whose shortfall exceeded an **absolute** `min_w = saturation_w = 500 W`. OPT-09's
+weights shrink a low-headroom unit's share, and shrinking the share shrinks the
+gap - so a unit standing at +125 W and delivering **0 W**, 100 % delinquent, was
+never named. It now names a unit whose gap exceeds `min_w` **OR** exceeds
+`DELIVERY_SHARE_FRACTION` (0.85) of that unit's **own share**. Measured on the
+same closed-loop fixture, same weights, same detector:
+
+    load 1500 W  OPT-09 weights  the old rule    named on   0/200 cycles
+    load 1500 W  OPT-09 weights  the shipped one named on 200/200 cycles
+
+⚠️ **`saturation_w` WAS NOT LOWERED, AND THAT IS THE POINT.** Lowering it would
+report normal saturation as delinquency on every unit at once - SAFETY-41's
+four-attempt false-trip regime, *"i'm getting really annoyed by the notifications
+in my HA"*. This ADDS a second way to be named: all seven of `test_delivery.py`
+part 1's existing fixtures answer identically under the old rule and the new one,
+and the absolute arm's suppression window is still exactly 1000 W / 6500 W.
+
+**0.85 is derived, not chosen.** `admin/goodwe:sim/delivery_window.py --campaign
+--fractions` grades it over the SAME 2640 no-fault runs that produced
+`delivery_hold_s`'s floor. The worst continuous legitimate episode is **35.0 s**
+under the absolute arm alone (SAFETY-42 §3.3, reproduced in the same run) and
+**40.0 s** at 0.85. ⚠️ **THE CURVE HAS NO CLIFF AND THAT IS THE FINDING**: swept
+at twenty fractions from 0.05 to 0.99, everything from **0.45 to 0.99** is 40.0
+or 45.0 s - within one control cycle of every other - and below 0.45 it rises
+gradually to 90.0 s at 0.05. ⚠️ **"Zero false namings" is not achievable at any
+fraction and the record says so**; what is zero is false ALARMS - the longest
+legitimate episode in the whole sweep is 90.0 s against a 185 s hold, so nothing
+comes within 95 s of notifying anybody at any fraction measured, and 145 s at the
+shipped one. The N=1 control arm is 0.0 s at every fraction in all 2640
+one-inverter runs, and the flapping family is 0.0 s in all 840.
+
+⚠️ **[[CFG-07]] GETS ONE CONTROL CYCLE WORSE AND IS NOT FIXED HERE.**
+`delivery_hold_s` is declared `int(30,900)` while the measured legitimate worst
+was 35.0 s; it is now 40.0 s, so the band of schema-legal false-tripping values
+widens from `[30, 35]` to `[30, 40]`. CFG-07 should be re-derived against 40.0 s.
+⚠️ **IT MUST CLOSE BEFORE INVERTER B IS COMMISSIONED (~2026-09-12)**, not before
+the next deploy: the band is UNREACHABLE at N=1 - `delinquent()` answers `{}` for
+every input at a roster of one, so no `delivery_hold_s` can false-trip on the
+install running today - and commissioning B is the moment it becomes reachable.
+⚠️ **AND CFG-07 HAS NO DECISION DOCUMENT IN THIS TREE** - it exists as a card
+reference here and in SAFETY-43's record and nowhere else, so whoever takes it
+writes `DECISIONS/CFG-07-*.md` first.
+**SAFETY-43 CHANGES NO OPTION VALUE, NO SCHEMA BOUND AND NO DEFAULT** - the
+fraction is a derived constant with a committed harness behind it, deliberately
+not a knob, for the reason there is no `delivery_w`. The one edit to
+`config.yaml` is a COMMENT: the `delivery_hold_s` block described the check as
+"short by more than `saturation_w`", which is now only half of it.
+
+**No behaviour change at one inverter, asserted rather than argued**: `plan()` is
+handed a `weights` mapping that raises if anything reads it and still answers
+identically at a roster of one (and the same mapping raises at a roster of two,
+so the tripwire is real), and `delinquent()` answers `{}` on all 84 single-unit
+(share, measured, rule) combinations including 0 of 125 W.
+`DECISIONS/SAFETY-43-delinquency-relative-to-the-ask.md` carries the curve, the
+positive controls and the CFG-07 assessment.
+
+⚠️ **TEST-21 CHANGES NO CODE AND NO OPTION. IT CHANGES A NUMBER THIS RECORD
+CLAIMED.** `delivery_hold_s: 185` shipped resting on *"30.0 s, the worst
+continuous legitimate episode at the shipped threshold across all 2640 runs"* -
+from a harness `DECISIONS/SAFETY-42-unit-delivery.md` §1 deliberately did **not
+commit**, because it imports `plantsim` and `frame_check` out of `admin/goodwe`.
+A shipped safety constant the next reviewer cannot re-run is PROC-01's shape.
+
+The harness is committed now, **in `admin/goodwe` as `sim/delivery_window.py`**,
+beside `sitesim.py`. It is there and not here on purpose: those two modules are
+that repo's, and restating them in this tree is the CI-08 coupling that has
+broken three times. The dependency runs one way instead - this repo's `control`,
+`arbiter` and the shipped `delivery.delinquent()` are imported live out of
+whichever checkout is under test. Re-run it with:
+
+```
+cd goodwe/sim && GOODWE_ADDON_REPO=/path/to/goodwe-addon \
+    python3 -B delivery_window.py --selftest --campaign --blag
+```
+
+⚠️ **AND 30.0 s DID NOT REPRODUCE - IT IS 35.0 s.** Reported as measured rather
+than adjusted to agree. What DID reproduce is the part the compatibility promise
+rests on: the **N=1 control arm is 0.0 s at every threshold in every run**, and
+the **flapping family is 0.0 s everywhere** in all 840 runs. What did not: the
+headline (one control cycle worse), the reversal family (20.0 → 30.0 s), the
+non-monotone curve (this rebuild's is monotone, so that finding is now an
+argument and not a measurement), and the low-threshold rows, badly. The B-lag
+rows moved the other way - a hypothetical unit four times slower than A gives
+**75.0 s**, not 90.0 s - and they are relabelled throughout as a **sensitivity
+study on a hypothesis about inverter B, which does not physically exist yet**,
+rather than as a measurement of it.
+
+**`delivery_hold_s` is NOT changed.** 185 s is 5.29× the re-measured worst
+(was claimed 6.2×), 2.47× the four-times-slower case, 6.2× the firmware's
+`LINK_STALE_S` veto and 4.9× inside the billing quarter: all four bounds still
+clear and none narrowly. ⚠️ One consequence is worth an operator's attention:
+at 35.0 s the rig's floor is now **above** the schema's own `int(30,900)`
+minimum, so `delivery_hold_s: 30` is below a measured legitimate episode and
+would report on ordinary saturation. The option is marked *derived, do not
+change*; raising the schema floor is left to its own card.

 ⚠️ **UX-02 ADDS A SECOND PLACE TO CONFIGURE THIS ADD-ON, AND NO NEW WAY TO
 COMMAND THE INVERTER.** The ingress panel gains a Configuration page beside
@@ -131,6 +234,47 @@ already declared, and the panel was already live.
 the Configuration tab **can** be removed - at the cost of all 71 options, which
 is why it stays. `DECISIONS/UX-02-two-config-surfaces.md` carries both.

+⚠️ **UX-04 CHANGES NO BEHAVIOUR, NO SCHEMA AND NO DEFAULT.** It adds one check
+file, three assertions to a probe, and a decisions record; `config.yaml`,
+`translations/` and every module under `app/` are byte-identical to `6e7b233`.
+It closes out [[UX-01]], which shipped in **0.7.0 and is live**, with three
+criteria that were never met:
+
+* **AC 3 - "the visible count drops to ~15" - is WITHDRAWN as unsatisfiable.**
+  `options:` and `schema:` declare the same 71 keys, so the set HA's *show
+  unused optional configuration options* toggle can hide is **empty**, and
+  Supervisor's own `UiOptions` marks **59 of the 71 required** - the frontend
+  shows those unconditionally, so 59 is the floor even with `options:` emptied
+  altogether. The target was ~4x below what any manifest can reach. ⚠️ And the
+  12 hides that *are* expressible would be bought with the deep-merge: a key
+  absent from `options:` has no default for the Supervisor to merge into a
+  0.7.0 box's stored blob, which is exactly the `delivery_hold_s` mechanism
+  that makes this release's upgrade safe rather than exit 78. The count was
+  **printed** by `tools/supervisor_shape_probe.py` §F from the day it was
+  written and never graded, which is how this passed review; it is now three
+  assertions, and removing one key from `options:` turns two of them red.
+* **AC 4's third clause - the AST check for stray in-code literal fallbacks -
+  is BUILT**, as `test_defaults_pin.py`. It discovers the six
+  `*_OPTION_DEFAULTS` mirrors from the tree rather than naming them, grades all
+  30 mirrored keys against `config.yaml`, and freezes the CFG-03 backlog at
+  exactly today's **27 (module, key) pairs** so it can only shrink. It found
+  one previously unreported divergence: `main.py` answers a missing
+  `target_grid_w` with **0** where `config.yaml` ships **-10** - unreachable on
+  a Supervisor install, recorded rather than corrected. Its red direction runs
+  on every invocation, and a literal planted in the real tree was shown red.
+* **AC 5 / the CFG-04 double-literal `max_w` hole is closed, by
+  [[UX-02]]'s `bcee235` and not by UX-01's `fef6ef2`.** Re-measured on this
+  branch: the `2000 -> 2500` triple mutation is **RED at `test_selfconfig.py`**,
+  where it survived 21/21 green on 0.7.0. The attribution is corrected in
+  `DECISIONS/UX-04-the-configuration-tab-and-the-defaults-pin.md`; nothing was
+  re-done.
+
+The 8 per-entry `inverters:` translations **stay**. They are shipped, validated
+and measured *not* to reach the Supervisor tab - `ha-selector-object` labels
+those fields `fields[name].label || name` and the converter sets no `label` -
+and UX-02's ingress panel is the surface that can render them. A written
+decision, not a change.
+
 ## 0.8.0

 ⚠️ **A CONFIG THAT STARTED ON 0.7.0 CAN BE REFUSED ON 0.8.0, AND THAT IS WHY
@@ -161,17 +161,18 @@ its floor is `account()`'s explained case, so SAFETY-42 stays silent - and after
 this card that ask does not arrive in the first place, which removes the benign
 arm from the detector's input rather than teaching it a new discriminator.

-⚠️ **BUT THE WEIGHT DOES MOVE, AND IT NARROWS SAFETY-42's DETECTION WINDOW. THIS
-IS A DETECTION REGRESSION AT N=2 AND IT IS NOT FIXED HERE.** An earlier draft of
+⚠️ **BUT THE WEIGHT DOES MOVE, AND IT NARROWED SAFETY-42's DETECTION WINDOW.
+THAT WAS A DETECTION REGRESSION AT N=2, IT WAS NOT FIXED HERE, AND SAFETY-43
+CLOSED IT - see the block at the end of this section.** An earlier draft of
 this section said "no envelope closes and no weight moves - OPT-09 cannot see it
 at all". The second clause was **false**, and it was load-bearing for the claim
 that the two cards do not overlap. A weight is a function of the pack's ENERGY
 HEADROOM, not of what the pack delivers, so a wedged unit's weight moves whenever
 its headroom differs from its sibling's - which is the ordinary case this card
-exists to exploit. And `delivery.delinquent()` (`app/delivery.py:484`)
-names a unit only while its `short_by()` gap exceeds `min_w =
-Tuning.saturation_w` - the comparison is `app/delivery.py:531`, and it is
-**500 W in absolute watts**. Shrinking a unit's share shrinks its gap. Shrink the
+exists to exploit. And `delivery.delinquent()`
+named a unit only while its `short_by()` gap exceeded `min_w =
+Tuning.saturation_w`, and that was **500 W in absolute watts** with nothing
+else beside it. Shrinking a unit's share shrinks its gap. Shrink the
 share far enough and the gap drops below 500 W while the unit is still delivering
 nothing, and the alarm stops firing.

@@ -200,20 +201,47 @@ a 9600 W site; above it - once the healthy pack is at its own rating and
 `_fill()`'s last-resort pass hands the wedged unit a big share again - the alarm
 returns. ⚠️ BOTH BOUNDS ARE ARITHMETIC ABOUT `_fill()`'s PROPORTIONAL REGIME AND NEITHER IS A REACHABLE STEADY STATE. Swept on this fixture, the healthy pack's own 5000 W rating binds FIRST: suppressed 0/200 up to a 3000 W load, 4/200 at 4000 W, 41/200 at 4900 W (site target 5320 W, the largest reached with the wedged unit still under 500 W), then 195/200 at 5000 W once `_fill()` pins A at its cap and hands B 3000 W. So the alarm returns BELOW the 6500 W formula bound, not above it. The equal-weights 1000 W bound is unreachable for the same reason - the equal arm names the wedged unit 199/200 at every load swept, 1000 W included.

-`test_packshare.py` part 8b grades all of it and **pins the current,
-unflattering behaviour**, so that the day it changes a named check moves. The
-closure - `delinquent()` measured against the ask rather than a flat 500 W - is
-card **SAFETY-43**.
+`test_packshare.py` part 8b graded all of it and **pinned the unflattering
+behaviour**, so that the day it changed a named check would move. It has moved.

-So the seam, stated accurately: **OPT-09 stops the ask AND narrows SAFETY-42's
-detection window for a low-headroom unit.** The two cards do not merely meet from
-opposite sides; this one takes something away from the other one. Closing the gap
-means making `delinquent()` relative to the ask rather than absolute, which is
-SAFETY-42's code and a design change - it is filed as its own card, deliberately
-not done here, and this paragraph is its input. Until then this is a **disclosed
-residual**, in the same voice as SAFETY-13's "2 of 272" and SAFETY-42's own "the
-discriminator does not exist": at two inverters, a wedged unit with materially
-less energy headroom than its sibling can go unreported at ordinary house loads.
+### ⚠️ CLOSED BY SAFETY-43 - THIS IS NO LONGER A DISCLOSED RESIDUAL
+
+This section used to end *"until then this is a disclosed residual [...] at two
+inverters, a wedged unit with materially less energy headroom than its sibling
+can go unreported at ordinary house loads"*. That sentence is **retired, not
+softened**, and what replaces it is a measurement rather than an intention.
+
+`delivery.delinquent()` is now relative to the ask **as well as** absolute:
+
+    named if  short_by() > min_w   OR   short_by() > frac * share
+
+with `frac = delivery.DELIVERY_SHARE_FRACTION`. A unit delivering 0 of 125 W is
+100 % delinquent and is named; a unit delivering 0 of 1500 W was already named
+and still is. On the **same fixture, under the same weights**, part 8b now
+grades both rules in one run - `share_frac=None` is SAFETY-42's absolute arm,
+exactly:
+
+    load 1500 W  OPT-09 weights  absolute arm alone   named on   0/200 cycles
+    load 1500 W  OPT-09 weights  the shipped rule     named on 200/200 cycles
+
+⚠️ **`min_w` WAS NOT LOWERED, AND THAT IS THE HALF THAT MATTERS HERE.** The
+absolute arm's suppression window is still exactly the arithmetic above - 1000 W
+at equal weights, 6500 W at these - and part 8b still asserts both numbers.
+Lowering `saturation_w` would have traded a missed alarm for a false one on
+every unit at once, which is SAFETY-41's four-attempt regime. SAFETY-43 added a
+**second** way to be named and left the first where it was.
+
+⚠️ **AND WHAT OPT-09 STILL DOES NOT DO IS UNCHANGED.** This card stops the site
+ASKING a pack whose own SoC says it cannot deliver; it does not re-allocate a
+wedged unit's watts, and SAFETY-43 did not make it. The alarm got louder; the
+loop did not move. The derivation of `frac` - swept over the same 2640 no-fault
+runs that produced `delivery_hold_s`'s floor - is in
+`DECISIONS/SAFETY-43-delinquency-relative-to-the-ask.md`.
+
+So the seam, stated accurately **as of SAFETY-43**: OPT-09 stops the ask, and it
+still narrows the ABSOLUTE arm's window for a low-headroom unit exactly as
+described above - but it no longer narrows the DETECTOR, because the detector no
+longer reads only that arm.

 ⚠️ **And OPT-09 supplies the mechanism SAFETY-42 said was missing, without
 connecting it.** That module's rejection of re-allocation rests on *"the only
@@ -68,22 +68,102 @@ deliberately *not* in the model). **Every two-inverter figure in this document
 is two copies of inverter A.** §3.3 is the sensitivity sweep that sizes the
 margin that costs.

-### Reproducing it
+### Reproducing it — ⚠️ THE HARNESS IS COMMITTED NOW (TEST-21, 2026-09-07)
+
+⚠️ **THIS SECTION USED TO SAY THE HARNESS WAS NOT COMMITTED, AND THAT WAS THE
+DEFECT TEST-21 FOUND.** A shipped safety constant whose floor the next reviewer
+cannot re-run is PROC-01's shape, and PROC-01 had eleven instances in three
+days. The paragraph this replaces was true — what the harness does *is* small
+enough to restate — and it was not enough: a restatement in prose leaves every
+axis it does not name to the next person's judgement, and §3.3 below now records
+how far that moves the number. **It moved. See §3.3's second table.**
+
+It is committed in the **other** repository, `admin/goodwe`, as
+`sim/delivery_window.py`, beside `sitesim.py` — the container-free two-master
+harness it is built on. From clean checkouts of both repos:
+
+```
+cd goodwe/sim
+GOODWE_ADDON_REPO=/path/to/goodwe-addon python3 -B delivery_window.py --selftest
+GOODWE_ADDON_REPO=/path/to/goodwe-addon python3 -B delivery_window.py --campaign --blag
+```
+
+stdlib only. No container, no HA, no esphome, **no rig and no rig lock** — it is
+a `plantsim` campaign, not a rig job. About fifteen minutes for the whole thing.
+
+⚠️ **AND IT IS IN `admin/goodwe` RATHER THAN HERE, WHICH IS CI-08's RULE
+APPLIED RATHER THAN IGNORED.** It imports `plantsim` (the plant, the per-unit
+wire, `Site`) and `frame_check` (the firmware's own encoder), both **owned by
+that repo** and neither present in this one's CI image. Committed here it had
+two possible fates and both are the coupling CI-08 has been broken by three
+times: skipped for want of its imports — a green that establishes nothing — or
+the plant and the encoder **restated in this tree**, which is a second copy of
+something the other repo owns, free to drift, and drift in a fixture is
+invisible until the number it defends is wrong. There the dependency runs one
+way and stays one-directional: **this** repo's `control`, `arbiter` and
+`delivery` are imported live out of whichever checkout is under test
+(`$GOODWE_ADDON_REPO`), never copied — so a change to `delinquent()` moves the
+number on the next run instead of silently disagreeing with a fixture. That is
+the seam `sitesim.py` already stands on, which is the other reason the campaign
+is its neighbour and not a third arrangement.
+
+⚠️ **THE SELFTEST IS CI-GATED; THE CAMPAIGN IS NOT, AND THE SPLIT IS
+DELIBERATE.** `admin/goodwe`'s `.gitea/workflows/checks.yml` gained one step —
+`python3 -B delivery_window.py --selftest`, seconds, beside the existing
+`norig, no lock` TEST-19 and TEST-20 steps, in the workflow that **already**
+checks out this repo for `two_writers` and accepts that boundary in writing. So
+a change **here** that moved `saturation_w`, `max_w`, `delivery_hold_s` or the
+shape of `delinquent()` turns that step red rather than leaving a stale number
+in this record. The **campaign** (`--campaign --blag`, ~15 minutes) is not
+gated: it is a campaign, not a gate, and `failure-injection.yml`'s jobs run on
+the **rig** runner, which a container-free simulation has no business holding.
+What TEST-21 asked for is a command a reviewer can run from a clean checkout —
+that is the block above.
+
+⚠️ **`--selftest` IS A FALSIFIABILITY ARM, NOT A SMOKE TEST, AND IT RUNS
+FIRST.** A campaign whose N=1 column and whose flapping column are 0.0 s
+everywhere is indistinguishable from a campaign whose detector was never wired
+up — `sitesim.broadcast_onset` exists for the same reason. So the selftest
+drives a genuinely **wedged** unit through the same loop and refuses to proceed
+unless the shipped `delinquent()` names it for longer than the shipped hold
+(it names it for 275.0 s). Every 0.0 s in §3.3 is only worth reading because
+that arm passes.
+
+What it does, unchanged from the restatement it replaces: build two
+`plantsim.Inverter`s from `INVERTER_A` / `INVERTER_B`; close the loop with the
+add-on's own `compute()` on the SITE sum (`Controller.site_batt_w`) and
+`arbiter.allocate_record()` over per-unit envelopes of `±min(max_w, rating)`;
+put each share on its own wire through `Inverter.command()`, honouring a
+configurable **write stagger** so B's number can land up to four cycles after
+A's (the serialised write path); advance `Site.settle()` one control cycle; and
+record, per unit, the command standing *before* the write and that unit's 0x0518
+power. Then grade each cycle with the **shipped** `delivery.delinquent()` and
+take the **longest run of consecutive cycles** on which it names anybody — not a
+count of naming samples, because the window is cleared by any cycle that does
+not.
+
+⚠️ **SINCE [[SAFETY-43]] THE ROWS OF §3.3's TABLE ARE GRADED WITH
+`share_frac=None`** — `delinquent()`'s ABSOLUTE arm alone, which is this card's
+rule exactly. Said here because it is the difference between §3.3 still meaning
+what it is labelled as and quietly becoming a table about a different predicate:
+SAFETY-43 added a SECOND way to be named, the harness grades it in its own rows
+(`--fractions`), and the numbers below are unchanged and were re-measured to
+confirm it. `DECISIONS/SAFETY-43-delinquency-relative-to-the-ask.md` §3 carries
+those rows and what they cost — the worst legitimate episode goes from 35.0 s to
+40.0 s at the shipped fraction, which does not move `delivery_hold_s`.
+
+⚠️ **ONE THING THE RESTATEMENT DID NOT SAY, AND IT IS LOAD-BEARING.** "The
+command standing before the write" is the **controller's** standing command —
+`Inverter.standing_w()`, i.e. `last_write` — and **not** the far end of the
+wire. A write that has been issued but has not landed yet is standing as far as
+this add-on is concerned, which is the whole of the write-stagger axis. Grading
+against the plant's latch instead reads a number no add-on has, and it makes
+both the stagger *and a wedged unit* invisible: a unit whose writes never arrive
+shows a standing command of 0 W, is never `asked`, and is never named. That is
+not a conservative harness, it is a harness that cannot fail — and the rebuild
+did exactly this on its first run, which is how the falsifiability arm earned
+its place above.

-The harness is **not committed here**, for the reason SAFETY-30's and
-SAFETY-31's were not: it imports `plantsim` and `frame_check`, which live in the
-other repository and are not present in this one's CI image. What it does is
-small enough to restate: build two `plantsim.Inverter`s from `INVERTER_A` /
-`INVERTER_B`; close the loop with the add-on's own `compute()` on the SITE sum
-(`Controller.site_batt_w`) and `arbiter.allocate_record()` over per-unit
-envelopes of `±min(max_w, rating)`; put each share on its own wire through
-`Inverter.command()`, honouring a configurable **write stagger** so B's number
-can land up to four cycles after A's (the serialised write path); advance
-`Site.settle()` one control cycle; and record, per unit, the command standing
-*before* the write and that unit's 0x0518 power. Then grade each cycle with the
-**shipped** `delivery.delinquent()` and take the **longest run of consecutive
-cycles** on which it names anybody — not a count of naming samples, because the
-window is cleared by any cycle that does not.
 ⚠️ MUTATION.md's note applies: this harness writes `__pycache__` into **both**
 trees. Run it with `python3 -B` and purge.

@@ -213,7 +293,9 @@ Every run is executed a second time at a **roster of one** on the same stimulus.
 check"* from *"a property of the second inverter"*.

 Worst **CONTINUOUS** legitimate episode, as a function of the shortfall
-threshold:
+threshold. **ORIGINAL CAMPAIGN, 2026-09-06, from the harness that was not
+committed — SUPERSEDED, and kept here because a record that quietly replaced its
+own numbers would be worth less than one that shows both:**

 | threshold | reversals (of 1050) | flapping (of 840) | weighted (of 750) | **N=1 control (1890 runs)** |
 |---:|---:|---:|---:|---:|
@@ -225,7 +307,100 @@ threshold:
 | 1500 W | 10.0 s (51) | 0.0 s (0) | 10.0 s (50) | **0.0 s (0)** |
 | 2000 W | 10.0 s (24) | 0.0 s (0) | 5.0 s (23) | **0.0 s (0)** |

-Three findings worth stating rather than leaving in the table:
+#### ⚠️ TEST-21, 2026-09-07 — THE RE-RUN, AND 30.0 s DID NOT REPRODUCE
+
+`admin/goodwe:sim/delivery_window.py --campaign`, same three families, same
+counts, same thresholds, the shipped `delinquent()`:
+
+| threshold | reversals (of 1050) | flapping (of 840) | weighted (of 750) | **N=1 control (2640 runs)** |
+|---:|---:|---:|---:|---:|
+| 100 W | 120.0 s (770) | 0.0 s (0) | 165.0 s (685) | **0.0 s (0)** |
+| 250 W | 65.0 s (687) | 0.0 s (0) | 100.0 s (590) | **0.0 s (0)** |
+| **500 W** | **30.0 s (460)** | **0.0 s (0)** | **35.0 s (403)** | **0.0 s (0)** |
+| 700 W | 25.0 s (342) | 0.0 s (0) | 30.0 s (318) | **0.0 s (0)** |
+| 1000 W | 15.0 s (4) | 0.0 s (0) | 25.0 s (131) | **0.0 s (0)** |
+| 1500 W | 0.0 s (0) | 0.0 s (0) | 20.0 s (43) | **0.0 s (0)** |
+| 2000 W | 0.0 s (0) | 0.0 s (0) | 0.0 s (0) | **0.0 s (0)** |
+
+**The floor is 35.0 s, not 30.0 s** — weighted family, and the worst single run
+is `gain=0.2 slew_w=1000 stagger=4 split=0.10/0.90` on the ±2200 W reversal. The
+argmax is printed with the table on purpose (see below): a worst case sitting on
+an edge of the grid would be a number about the grid.
+
+⚠️ **AND BY THAT STANDARD, TWO OF ITS FOUR AXES ARE ON AN EDGE. Measured, at the
+argmax's other axes held fixed:**
+
+    stagger  0 -> 1 -> 2 -> 3 -> 4      0.0  15.0  25.0  30.0  35.0 s
+    weights (A) at stagger 4            0.5 -> 20.0   0.7/0.9/0.99 -> 0.0
+                                        0.10 -> 35.0 s
+
+Strictly increasing in stagger and maximal at 4, the largest modelled stagger;
+and maximal at the most asymmetric split. It is interior only in `gain` (2nd of
+6) and `slew_w` (3rd of 5).
+
+⚠️ **SO THE HEADLINE IS 0.0 s AT stagger=0, AND 35.0 s IS ENTIRELY A PRODUCT OF
+THE MODELLED SERIALISED-WRITE STAGGER.** That stagger is declared MODELLED at
+`sim/delivery_window.py:60-64` — it is a term of this figure, not a detail of
+it. A site whose two writes land in the same cycle does not produce this episode
+at all.
+
+This does not move `delivery_hold_s`: 185 s must cover the worst case the
+modelled site can produce, and a stagger of four cycles is inside what a real
+two-inverter write path can do. But a reader entitled to ask "is 35.0 s about
+the plant or about the grid?" deserves the answer, and the answer is: **partly
+about the grid, and the grid's shape is the record's own, fixed by §3.3 before
+this rebuild ran.** Found in review of TEST-21; the PR body had claimed the
+argmax was interior, and it is not.
+
+**WHAT REPRODUCED, EXACTLY:**
+
+* the **N=1 control column is 0.0 s at every threshold**, in every run. That is
+  §3.2's algebra measured, and it is the claim the compatibility promise rests
+  on. (This rebuild ran the control arm over all three families — 2640 runs —
+  where the original ran 1890, i.e. the two families whose stimulus is not
+  weighted. The weighted stimulus at a roster of one is the same stimulus with
+  inert weights, and it too reports nothing at any threshold.)
+* the **flapping family is 0.0 s at every threshold**, in all 840 runs. The
+  proper-subset rule really does defeat the family that defeated SAFETY-30's
+  site window.
+* **45.0 s at twice A's lag**, to the cycle. See the B-lag table below.
+
+**WHAT DID NOT:**
+
+* the headline. **35.0 s against a recorded 30.0 s** — one control cycle worse,
+  and 20.0 s → 30.0 s on the reversal family, two cycles worse.
+* **the non-monotonicity.** §3.3 finding 3 read the original 55.0 s at 250 W
+  against 40.0 s at 100 W as a property of the reduction. **This rebuild's
+  curve is monotone decreasing on both N=2 families.** The *mechanism* finding 3
+  describes is real and is still the reason not to read either table as a
+  sensitivity curve — lowering the threshold really can turn a proper subset
+  into the whole roster and break the run — but **this grid does not exhibit
+  it**, so the claim now rests on the mechanism alone and no longer on a
+  measurement. That is a weaker claim than the one this record made, and it is
+  written down as weaker.
+* the low-threshold rows, badly: 120.0 s / 165.0 s at 100 W against 40.0 s /
+  75.0 s. Nothing ships at 100 W, but the size of that gap is the honest measure
+  of how much of the original table was a property of axes nobody wrote down.
+
+⚠️ **WHY THE NUMBERS MOVED, AS FAR AS IT CAN BE ESTABLISHED.** §3.3 fixes the
+*counts* of every axis (6 gains × 5 slews; 7 reversal sizes; 7 flap periods × 4
+amplitudes; 5 staggers; and the 5 splits, which it lists) and the *ranges*
+(`config.yaml`'s `gain` `float(0.05,1.0)` and `slew_w` `int(0,5000)`). It does
+**not** fix the individual values, the run length, or the per-unit cap. The
+rebuild spread them over the admissible ranges and **prints every one of them
+above its own table**, so the caveat travels with the number instead of living
+in somebody's memory. The residual disagreement — one to two control cycles at
+the shipped threshold — is inside the sensitivity to that choice; the
+low-threshold rows are not, and are reported as not.
+
+⚠️ **AND THE HARNESS WAS NOT ADJUSTED UNTIL IT AGREED.** The one substantive
+correction made during the rebuild was made *against* the doc's number, not
+towards it, and it is recorded in §1: grading on the plant's latch rather than
+the controller's standing command produced a harness that could not name even a
+wedged unit. Fixing that made the numbers **larger**. Nothing else was tuned.
+
+Findings 1-3 below are the original campaign's, kept with finding 3 now
+qualified by the re-run above:

 1. **The N=1 column is zero everywhere, in all 1890 runs and at every
   threshold.** That is the compatibility promise measured, not argued.
@@ -268,37 +443,77 @@ would be a preference, and all the margin this check needs is in the window.

 Four bounds, and the value clears all four.

-* **lower, from the rig — 30.0 s.** The worst continuous legitimate episode at
-  the shipped threshold across all 2640 runs (weighted family; 20.0 s on
-  reversals, 0.0 s on flapping).
+* **lower, from the rig — 35.0 s** (TEST-21, 2026-09-07; **the original record
+  said 30.0 s and it did not reproduce**). The worst continuous legitimate
+  episode at the shipped threshold across all 2640 runs — weighted family, worst
+  single run `gain=0.2 slew_w=1000 stagger=4 split=0.10/0.90` on the ±2200 W
+  reversal; 30.0 s on reversals, 0.0 s on flapping. Re-run it with
+  `admin/goodwe:sim/delivery_window.py --campaign` (§1).
+  ⚠️ **THIS IS NOW STRICTLY THE BINDING LOWER BOUND, WHICH IT WAS NOT BEFORE.**
+  At 30.0 s it coincided with the firmware's veto below and a reader could not
+  tell which of the two was doing the work. At 35.0 s the rig's number is on its
+  own, above the firmware's — so **the rig is what stops `delivery_hold_s` going
+  low**, and the schema's `int(30,900)` floor no longer sits above the worst
+  legitimate episode. An operator who sets it to the schema minimum of 30 s is
+  now **below** a measured legitimate episode and will get false reports; the
+  option is marked *derived, do not change* and this is what that marking is
+  for. Raising the schema floor to 35 is a change to a shipped option's
+  contract and is left to its own card rather than smuggled in here.
 * **lower, from the firmware — 30 s (`inverter.LINK_STALE_S`).** A unit not
  executing its share for less than the firmware's own veto window is a *link*
  fault the add-on already answers by dropping it from the split. Reporting
  inside that window would be a second, louder alarm about a state already
  being handled — the SAFETY-15/SAFETY-19 failure exactly.
-* **margin, from inverter B not existing — and it is MEASURED, not felt.** The
-  campaign's B is A copied, so 30.0 s is two identical machines. Sweeping a
-  *different* second unit through the weighted family (750 runs each):
+* **margin, from inverter B not existing.** ⚠️ **THIS ROW IS A SENSITIVITY
+  STUDY ON A HYPOTHESIS ABOUT B, NOT A MEASUREMENT OF B, AND THE ORIGINAL
+  WORDING ("MEASURED, not felt") OVERSTATED IT.** Inverter B is not installed
+  (~2026-09-12; NOTES.md §1.1b). The campaign's B is **A copied**, `measured=
+  False`, so the 35.0 s above is **two identical machines** and every row below
+  is a *what-if*: what a slower second unit would cost this window. **No row
+  here is B's lag.** What it is genuinely evidence for is the SHAPE of the
+  margin — which term the window has to be sized against — and that is what it
+  is used for. Sweeping a *different* second unit through the weighted family
+  (750 runs each). Second column is the original record, third is TEST-21's
+  re-run with the committed harness (`--blag`):

-  | B lag | B ramp | worst legitimate | runs naming |
-  |---:|---:|---:|---:|
-  | 10.0 s (A) | 800 W/s (A) | **30.0 s** | 368/750 |
-  | 20.0 s | 800 W/s | 45.0 s | 469/750 |
-  | 10.0 s | 400 W/s | **30.0 s** | 370/750 |
-  | 20.0 s | 400 W/s | 45.0 s | 469/750 |
-  | 40.0 s | 200 W/s | **90.0 s** | 500/750 |
+  | B lag | B ramp | 2026-09-06 (uncommitted) | **TEST-21 re-run** | runs naming |
+  |---:|---:|---:|---:|---:|
+  | 10.0 s (A) | 800 W/s (A) | 30.0 s | **35.0 s** | 403/750 |
+  | 20.0 s | 800 W/s | 45.0 s | **45.0 s** | 454/750 |
+  | 10.0 s | 400 W/s | 30.0 s | **35.0 s** | 411/750 |
+  | 20.0 s | 400 W/s | 45.0 s | **50.0 s** | 454/750 |
+  | 40.0 s | 200 W/s | 90.0 s | **75.0 s** | 478/750 |

-  **The RAMP does not move it at all and the LAG moves it roughly linearly.**
-  The shortfall lasts as long as the unit's first-order lag holds it more than
-  `saturation_w` below its share; at these step sizes the ramp is never the
-  binding term. So the number to carry margin against is B's lag, and a unit
-  **four times slower than A** produces 90.0 s.
+  **The LAG moves it roughly linearly and the RAMP barely moves it**, which is
+  the original finding and it survives — with one correction: *"the RAMP does
+  not move it at all"* is too strong. Halving the ramp is worth 0 s at A's lag
+  and **+5 s at twice it** (45.0 → 50.0), so the ramp is a second-order term
+  rather than no term. The shortfall lasts as long as the unit's first-order lag
+  holds it more than `saturation_w` below its share; at these step sizes the
+  ramp is rarely the binding one. So the number to carry margin against is still
+  B's lag — and a hypothetical unit **four times slower than A** produces
+  **75.0 s**, where the original record said 90.0 s. ⚠️ Note the direction: this
+  row moved **in the margin's favour**, while the floor above moved against it.
+  Neither was adjusted to suit the other.
 * **upper, from the bill — 900 s**, the quarter that sets the capacity tariff.

-**185 s** is thirty-seven control cycles: **6.2×** the measured worst, **2.06×**
-the four-times-slower case, **6.2×** the firmware veto and **4.9×** inside the
-quarter. ⚠️ Deliberately **not 180** — that is SAFETY-02's site hold, and a
-reader must not be able to think this was carried across — and not 150 or 200.
+**185 s** is thirty-seven control cycles. Against the TEST-21 re-run it is
+**5.29×** the measured worst legitimate episode (35.0 s), **2.47×** the
+hypothetical four-times-slower second unit (75.0 s), **6.2×** the firmware veto
+(30 s) and **4.9×** inside the quarter (900 s). ⚠️ Deliberately **not 180** —
+that is SAFETY-02's site hold, and a reader must not be able to think this was
+carried across — and not 150 or 200.
+
+⚠️ **185 STANDS, AND HERE IS THE ARITHMETIC RATHER THAN THE ASSURANCE.** The
+record claimed 6.2× the measured worst; the re-run makes that **5.29×**. All
+four bounds still clear, none of them narrowly, and the *ordering* of the
+argument is unchanged: the rig's floor is the binding one, the quarter is the
+ceiling, and 185 sits between them with margin at both ends. So the value is
+**not changed** — a shipped safety constant is not moved because its margin
+went from 6.2× to 5.3× while every bound still holds, and moving it would
+invalidate SAFETY-41's four-attempt false-trip regime for nothing. What IS
+changed is the claim: this record no longer says 6.2×, and no longer says 30.0
+s, because neither is what the harness produces.

 **What it costs, stated rather than hidden:** up to 185 s of a site running
 short by one unit's share before anybody is told. Against a fault that ran
@@ -0,0 +1,381 @@
+# SAFETY-43 — delinquency measured against the ask, not against a flat 500 W
+
+**Status: closed.** `delivery.delinquent()` names a unit whose `short_by()` gap
+exceeds `min_w` **or** exceeds a fraction of that unit's own share. The fraction
+is `delivery.DELIVERY_SHARE_FRACTION = 0.85`, derived on the campaign in §3.
+
+⚠️ **INVERTER B DOES NOT PHYSICALLY EXIST** (expected ~2026-09-12; NOTES.md
+§1.1b). `plantsim.INVERTER_B` is inverter A **copied**, `measured=False`, and
+`test_packshare.py`'s plant is modelled in that file. **Every two-inverter
+figure in this document is two copies of A.** Nothing here is a measurement of
+inverter B.
+
+## 0. The defect, and why it is the quiet regime that matters
+
+[[OPT-09]] passes SoC- and capacity-aware weights to the allocator. A weight is
+a function of a pack's **energy headroom**, not of what the pack delivers, so a
+**wedged** unit — one that accepts its share and produces nothing — is weighted
+down exactly like a legitimately low-headroom one. `delinquent()` read an
+**absolute** gap, `min_w = Tuning.saturation_w = 500 W`. Shrinking the share
+shrinks the gap. Measured end to end on the real loop (`test_packshare.py`
+part 8b: real `compute()`, real `plan()`, real `packshare.weights()`, real
+`delivery.account()`, real `delinquent()`; B wedged at a healthy 15 % against its
+own 10 % cut-out, so no envelope closes and `account()` explains nothing — a
+SAFETY-42 **true positive** by that card's own definition):
+
+    load 1500 W  equal weights   shares {a:1500, b:1500}   named on 199/200
+    load 1500 W  OPT-09 weights  shares {a:1505, b: 125}   named on   0/200
+
+A unit delivering **0 of 125 W** is 100 % delinquent and 125 W short. Only the
+absolute threshold disagreed.
+
+⚠️ **THE SUPPRESSED REGIME IS THE DANGEROUS ONE.** At a load the healthy sibling
+covers, the meter is satisfied (grid within 5 W of zero on that fixture),
+`unmet_w` is 0 because both units *accepted* their shares, no envelope closes and
+no action-log line fires. The delinquency alarm is the **only** instrument that
+would have told the operator the site was quietly running one inverter.
+
+## 1. What was NOT done, and it is the larger half of this card
+
+⚠️ **`min_w` WAS NOT LOWERED.** It is `Tuning.saturation_w` — the gap at which
+the control law itself concludes the inverter is not following — and it exists to
+stop **normal saturation** being reported as delinquency. Lowering it trades a
+missed alarm for a false one on every unit at once. [[SAFETY-41]] spent **four
+attempts** on notification cadence in this system and the owner's words were
+*"i'm getting really annoyed by the notifications in my HA"*.
+
+So the change is an **`or`**, never a looser first arm:
+
+    named if  short_by() > min_w   OR   short_by() > frac * share
+
+Everything `min_w` named before is still named. Nothing `min_w` refused becomes
+noisy **unless the unit is delivering almost none of its ask**, which is not
+saturation-shaped at all. `test_delivery.py` part 1 runs all seven of its
+existing `delinquent()` fixtures under **both** rules — `share_frac=None` is
+SAFETY-42's rule exactly — and requires identical answers on every one.
+
+The absolute arm's suppression window is therefore **unchanged**: still
+`site target < 2·min_w` = 1000 W at equal weights, still
+`site target < min_w·(w_A+w_B)/w_B` = 6500 W under OPT-09's weights on part 8b's
+fixture. Part 8b still asserts both numbers, now as evidence that the first arm
+did not move.
+
+## 2. The evidence is a diff between two rules on ONE run
+
+`delinquent()` takes `share_frac`, and **`None` means the absolute arm alone**.
+That parameter exists for one reason: so the red and the green come out of one
+run over one tree instead of out of two commits that a reader has to trust.
+`test_packshare.py` part 8b grades every cycle twice:
+
+    load 1500 W  OPT-09 weights  absolute arm alone   named on   0/200 cycles
+    load 1500 W  OPT-09 weights  the shipped rule     named on 200/200 cycles
+
+    load 1500 W  equal weights   absolute arm alone   named on 199/200 cycles
+    load 1500 W  equal weights   the shipped rule     named on 200/200 cycles
+
+The equal-weights arm gains **exactly one cycle** — the first, where the slew
+limit holds the split at 500/500, a 500 W gap is not *over* a 500 W threshold
+(the absolute bound is strict) and the relative arm names a box delivering none
+of its 500 W. That single cycle is the whole of what this card costs an arm that
+was already firing.
+
+⚠️ **THE PIN THAT INVERTED.** Part 8b was written by OPT-09 to pin the
+**suppressed** behaviour, and its comment said: *"if it ever goes red because the
+alarm fired, read the card that made it fire before 'fixing' this line."* This is
+that card. The pin now asserts the opposite answer and keeps the old one beside
+it, measured rather than remembered.
+
+## 3. ⭐ Deriving the fraction
+
+⚠️ **DERIVED ON THE SAME CAMPAIGN AS `delivery_hold_s`'s FLOOR**, not chosen.
+`admin/goodwe:sim/delivery_window.py` — the harness [[TEST-21]] committed —
+gained a second grading axis: the same **2640 no-fault runs at N=2** (three
+families × 6 gains × 5 slews × write staggers; every unit executes the share the
+allocator gave it) and the same **2640 N=1 control runs**, graded at the shipped
+`min_w = 500 W` for each candidate fraction, on every cycle of every run — so a
+pass grades every fraction it is given, and the absolute arm beside them, over
+one tree and one set of runs. Every row in a pass is comparable to every other
+row in it, run for run.
+
+    cd goodwe/sim && GOODWE_ADDON_REPO=/path/to/goodwe-addon \
+        python3 -B delivery_window.py --selftest --campaign --fractions
+
+⚠️ **THE TABLE BELOW WAS PRODUCED BY TWO PASSES; `FRACTIONS` NOW HOLDS ALL
+TWENTY SO IT TAKES ONE.** The derivation ran the shipped ten
+`(0.25 … 0.99)` and then the ten that resolve the low end
+`(0.05 … 0.65)` with only that tuple edited between them. **The absolute-arm
+control row came out identical in both** (30.0 / 0.0 / 35.0 / 0.0, same argmax),
+which is what says the two passes are one campaign. A table a reviewer cannot
+reproduce without knowing to edit a line nobody wrote down is PROC-01's shape in
+miniature, so the committed `FRACTIONS` was widened to all twenty in review: the
+command above now reproduces every row below.
+
+### 3.1 The curve
+
+Worst **continuous** legitimate episode (the measure `DeliveryWatchdog` actually
+sees — its window is cleared by any cycle that does not name), runs naming in
+brackets:
+
+| `share_frac` | reversals (of 1050) | flapping (of 840) | weighted (of 750) | **N=1 control (2640)** |
+|---:|---:|---:|---:|---:|
+| absolute only | 30.0 s (460) | 0.0 s (0) | 35.0 s (403) | **0.0 s (0)** |
+| 0.05 | 85.0 s (826) | 0.0 s (0) | 90.0 s (616) | **0.0 s (0)** |
+| 0.10 | 75.0 s (824) | 0.0 s (0) | 90.0 s (607) | **0.0 s (0)** |
+| 0.15 | 70.0 s (836) | 0.0 s (0) | 70.0 s (609) | **0.0 s (0)** |
+| 0.20 | 70.0 s (839) | 0.0 s (0) | 60.0 s (629) | **0.0 s (0)** |
+| 0.25 | 55.0 s (832) | 0.0 s (0) | 60.0 s (611) | **0.0 s (0)** |
+| 0.30 | 55.0 s (840) | 0.0 s (0) | 50.0 s (614) | **0.0 s (0)** |
+| 0.35 | 55.0 s (840) | 0.0 s (0) | 55.0 s (618) | **0.0 s (0)** |
+| 0.40 | 50.0 s (840) | 0.0 s (0) | 60.0 s (658) | **0.0 s (0)** |
+| 0.45 | 45.0 s (840) | 0.0 s (0) | 45.0 s (627) | **0.0 s (0)** |
+| 0.50 | 45.0 s (840) | 0.0 s (0) | 45.0 s (636) | **0.0 s (0)** |
+| 0.55 | 40.0 s (840) | 0.0 s (0) | 40.0 s (636) | **0.0 s (0)** |
+| 0.60 | 40.0 s (840) | 0.0 s (0) | 40.0 s (645) | **0.0 s (0)** |
+| 0.65 | 35.0 s (840) | 0.0 s (0) | 45.0 s (635) | **0.0 s (0)** |
+| 0.70 | 35.0 s (840) | 0.0 s (0) | 45.0 s (635) | **0.0 s (0)** |
+| 0.75 | 35.0 s (840) | 0.0 s (0) | 45.0 s (635) | **0.0 s (0)** |
+| 0.80 | 35.0 s (840) | 0.0 s (0) | 45.0 s (635) | **0.0 s (0)** |
+| **0.85** | **35.0 s (840)** | **0.0 s (0)** | **40.0 s (635)** | **0.0 s (0)** |
+| 0.90 | 35.0 s (840) | 0.0 s (0) | 40.0 s (635) | **0.0 s (0)** |
+| 0.95 | 35.0 s (840) | 0.0 s (0) | 40.0 s (635) | **0.0 s (0)** |
+| 0.99 | 35.0 s (840) | 0.0 s (0) | 40.0 s (635) | **0.0 s (0)** |
+
+Worst of the three families, which is what `delivery_hold_s` has to cover:
+
+    frac   0.05  0.10  0.15  0.20  0.25  0.30  0.35  0.40  0.45  0.50
+    worst  90.0  90.0  70.0  70.0  60.0  55.0  55.0  60.0  45.0  45.0  s
+
+    frac   0.55  0.60  0.65  0.70  0.75  0.80 *0.85  0.90  0.95  0.99
+    worst  40.0  40.0  45.0  45.0  45.0  45.0  40.0  40.0  40.0  40.0  s
+
+⚠️ **THE FIRST ROW IS THE CONTROL AND IT REPRODUCED SAFETY-42 §3.3 EXACTLY** —
+30.0 s / 0.0 s / 35.0 s, and the same argmax (`gain=0.2 slew_w=1000 stagger=4
+split=0.10/0.90`, ±2200 W reversal). It is measured **in the same run** as every
+row below it, which is what makes those rows a comparison rather than a claim.
+
+### 3.2 What the curve says
+
+* ⚠️ **THERE IS NO CLIFF ANYWHERE IN THE CURVE, AND THAT IS THE FINDING.** From
+  **0.45 to 0.99** every figure is 40.0 or 45.0 s — within **one control cycle
+  (5.0 s)** of every other — and below 0.45 it rises **gradually**: 60.0 s at
+  0.25, 70.0 s at 0.15, 90.0 s at 0.05. A reviewer asking "why 0.85 and not
+  0.90?" is asking about a difference this campaign cannot resolve, and the
+  honest answer is that it cannot.
+* ⚠️ **IT IS NOT MONOTONE** — 0.40 reports 60.0 s where 0.35 reports 55.0, and
+  0.65–0.80 report 45.0 s where 0.55/0.60 and 0.85–0.99 report 40.0 — and that is
+  SAFETY-42 §3.3's own mechanism (c) rather than noise in the plant: **a more
+  sensitive rule puts MORE units into the short set, which can turn a proper
+  subset into the WHOLE roster and END the episode.** The reduction is not
+  monotone in its own threshold. Do not read this column as a sensitivity curve
+  in the ordinary sense.
+* **The shoulder is at about 0.45**, and it is a shoulder rather than an edge:
+  the reversal family's naming count saturates at 840 of 1050 runs from 0.30
+  upward, so below the shoulder what grows is the LENGTH of an episode, not the
+  number of runs that have one.
+* **The N=1 column is 0.0 s at every fraction, in all 2640 one-inverter runs.**
+  The compatibility promise is measured for the new arm, not argued.
+* **The flapping family is 0.0 s at every fraction, in all 840 runs.** A common
+  stimulus makes both units short at once and the subset stops being proper —
+  the property that defeated SAFETY-30's site window still holds.
+
+### 3.3 ⚠️ "Zero false namings" is not available at ANY fraction
+
+The card asked for *the largest fraction that yields zero false namings across
+the whole no-fault campaign*. **There is none.** Every fraction adds naming
+cycles: the reversal family goes from 460 runs naming to 840, the weighted family
+from 403 to 635. A relative arm that never fired on a legitimate transient would
+be a relative arm that never fired at all — the write stagger lands B's number up
+to four cycles late, and a unit whose new command has not arrived is, for those
+cycles, delivering a small fraction of what is standing on its wire.
+
+What **is** zero at every fraction is **false alarms**:
+
+> the longest continuous legitimate episode anywhere in the 20-point sweep is
+> **90.0 s** — at 0.05, the most sensitive fraction measured — against a
+> `delivery_hold_s` of **185 s**. Nothing in 2640 no-fault runs comes within
+> **95 s** of notifying an operator, **at any fraction measured**. At the
+> shipped 0.85 it is 40.0 s, i.e. 145 s of slack.
+
+That is the quantity SAFETY-41 cares about and it is the one this card had to
+protect. Stated as margin, at the shipped fraction: **185 s is 4.62×** the worst
+legitimate episode, where the absolute arm alone gave 5.29×.
+
+### 3.4 The bounds 0.85 answers to
+
+* **LOWER, from the campaign** — the shoulder at about **0.45**, below which the
+  added episode starts to grow (60.0 s at 0.25, 90.0 s at 0.05). 0.85 is 0.40
+  above it.
+* **UPPER, from the card's own case** — strictly **below 1.0**. At 1.0 the arm
+  can only fire on a unit that is *importing* while asked to discharge, which is
+  SAFETY-30's fault and not this one, so a unit delivering **none** of its share
+  (gap/share exactly 1.0) would be missed. The 0.15 below 1.0 is what covers a
+  wedged box reporting a small standby figure instead of a clean zero.
+* **UPPER, from the hold** — every fraction in the flat range leaves
+  `delivery_hold_s` between **4.11×** (45.0 s) and **4.62×** (40.0 s) the worst
+  legitimate episode, well inside the four bounds SAFETY-42 §3.3 requires of it.
+  **`delivery_hold_s` IS NOT CHANGED BY THIS CARD.**
+* ⚠️ **NOT DERIVED, AND SAID SO: the exact value inside [0.45, 1.0).** The
+  campaign cannot separate them. 0.85 is the midpoint of the argument above —
+  0.40 of margin above the shoulder, 0.15 below the ceiling — and a
+  re-derivation that moved it to 0.90 would be moving a number this campaign
+  never resolved. It is not moved for that reason.
+
+## 4. The positive control — what the relative arm must still NOT say
+
+A card that adds a way to raise an alarm has to show what it did **not** add.
+`test_packshare.py` part 8b runs the same fixture, the same weights and the same
+125 W share with the one fact that makes the shortfall legitimate changed:
+
+| fixture | B's share | named | why it is silent |
+|---|---|---|---|
+| B wedged at a healthy 15 % (the fault) | 77–154 W | **200 / 200** | — it is the fault |
+| B at its own 10 % cut-out | **0 W, all 200 cycles** | **0 / 200** | it is never **asked** |
+| B delivering **half** its share | 120 W | **0 / 200** | short by 50 %, inside both arms |
+
+### ⚠️ The cut-out row does NOT grade `account()`, and this record said it did
+
+**Corrected in review.** The first draft of this section claimed the wedged and
+cut-out arms had *"identical power traces — 0 W against a positive share, 100 %
+of the ask missing"* and that *"what separates them is `account()`"*. **Every
+part of that was false**, and the check that carried it was **vacuous**:
+
+* at its own cut-out the pack's usable energy **above that cut-out is zero**, so
+  `packshare.weights()` answers **0.0** for it, `_fill()`'s preferred pass gives
+  it nothing, the healthy sibling covers the load alone so the last-resort pass
+  never runs, and its share is **0 W on all 200 cycles**;
+* a unit with a 0 W share is not in `asked`, so `len(asked) < 2` returns `{}`
+  **before either arm is evaluated** and `account()` is **never consulted**;
+* the traces are not identical — the cut-out arm's share is **0 W on every
+  cycle** (asserted in part 8b and printed on its line) while the wedged arm's
+  is **125 W** at the last cycle (asserted in the GREEN check) and takes four
+  distinct values 77/125/133/154 W across the 200, measured by re-running the
+  same loop with per-cycle recording during review — and "100 % of the ask" is
+  0/0.
+
+Measured, by deleting the gate the claim rested on:
+
+    cut-out arm, account() live         named 0/200, distinct B shares: [0]
+    cut-out arm, account() DELETED      named 0/200, distinct B shares: [0]
+    wedged  arm, account() live         named 200/200, shares [77,125,133,154]
+
+⚠️ **The fact that replaces it is better, not weaker.** At its own cut-out
+**OPT-09 stops asking the pack at all**, so SAFETY-43's relative arm is never
+reached — a **second, independent protection sitting upstream of `account()`**:
+the allocator declines to ask before the detector has to decline to alarm. Part
+8b now asserts the 0 W share and prints it, and runs the arm a second time with
+`account()` deleted to show which mechanism is doing the work.
+
+⚠️ **It was invisible for one reason worth keeping**: that line was the only one
+in part 8b that did not print its shares. `inverter-b+0` would have been on
+screen from the first run. Every line in that part prints them now, and the
+derated control's vacuity guard now asserts **its own** unit is asked, rather
+than re-asserting a different run's 200/200 — which is exactly why the old guard
+could not see this.
+
+**The account gate is graded where it is real**, on a positive 125 W share, in
+`test_delivery.py` part 1 (*"THE ACCOUNT GATE IS UNTOUCHED BY IT"* — the same
+0 of 125 W with the shortfall explained names nobody). That part also carries
+that the relative bound is **strict** (a gap of exactly `frac · share` is not
+over it) and that the proper-subset rule still refuses a site-wide cause (both
+units delivering none of a 125 W share names nobody).
+
+## 5. N=1 is byte-identical, and it is asserted rather than argued
+
+The live install is **N=1**. Two independent statements, each graded against the
+thing that would break it (`test_packshare.py` part 8b):
+
+1. **`allocator.plan()` returns before `weights` is read at all.** Graded by
+   handing it a `weights` mapping that **raises on any read**: at a roster of one
+   it answers identically to the same call with no weights; at a roster of two
+   the same mapping is read and raises, so the tripwire is a real one.
+2. **`delinquent()` answers `{}` at a roster of one for every input** — 84
+   combinations of (share, measured, rule), including **0 of 125 W**, the exact
+   input this card added an arm for. `len(asked) < 2` returns before either
+   threshold is compared, so the fraction is never even multiplied.
+
+Nothing in this card is reachable at one inverter. It must land before inverter B
+is commissioned (~2026-09-12), not before the next deploy.
+
+## 5b. Mutation — the fraction is pinned from BOTH directions
+
+⚠️ **RUN, NOT REASONED ABOUT.** Each mutation was applied to a tree **copied
+fresh** from this branch and mounted `:ro` inside the checks image, so no `.pyc`
+could go stale — MUTATION.md's defect, and all three of these are **reds**, the
+direction that defect cannot fabricate. Recorded in `test_packshare.py` part 11
+as M12–M14.
+
+| mutation | named checks | where |
+|---|---:|---|
+| **M12** drop the relative arm, keep `gap > min_w` — *the card itself reverted* | **7** | `test_delivery.py` part 1 (3: the 0-of-125 W naming, the proportion in the sentence, the strict relative bound) + `test_packshare.py` 8b (4: GREEN back to 0/200, the equal arm's 200 back to 199, the 6000 W arm back to 195/199, the derated control's guard). `test_alloc.py` / `test_control.py` stay green — nothing outside this check moves. |
+| **M13** `DELIVERY_SHARE_FRACTION` → **1.0** — *too loose* | **7** | part 1 (2) + 8b (5), including the PRECONDITION that the fraction is strictly below 1.0. At 1.0 the card is a no-op: gap/share of exactly 1.0 is not *over* 1.0. |
+| **M14** `DELIVERY_SHARE_FRACTION` → **0.05** — *too sensitive* | **6** | part 1 (2: the strict bound, and *"IT TOOK NOTHING AWAY"* stops being true) + part **4b** (3: the `saturation_w` pin goes **vacuous**, the relative arm naming B at every threshold) + 8b's derated control, where a unit delivering half its share starts being named. |
+
+M13 and M14 are the pair that matters: the constant cannot be moved in **either**
+direction without a named check moving with it.
+
+### ⚠️ Two of these counts were wrong, and the reason is worth keeping
+
+**Corrected in review.**
+
+* **M12 and M13 read 1 and 1 in `test_delivery.py`, not 3 and 2.** Part 1 printed
+  the alarm sentence with a bare subscript, `delinquent(_tiny, 500)["b"]`. On any
+  tree where the relative arm is broken or absent that raises `KeyError` and
+  **aborts the file at part 1**, leaving parts 2–8 ungraded — so the recorded
+  coverage of a mutation run read *smaller* than the truth. The direction is safe
+  (an abort is a red, and MUTATION.md's defect can only fabricate a green), but a
+  figure that does not reproduce is a figure that does not reproduce. It is a
+  `.get` now, with the reason written beside it, and the counts above are from
+  re-runs with the guard in place.
+* **M14 was recorded as 4 named checks in `test_delivery.py` and it is 5** — part
+  1's pre-existing *"a shortfall AT the threshold is not one"* also goes red. That
+  one under-claimed, so it never inflated the evidence; it is still a recorded
+  number that did not reproduce.
+
+## 6. ⚠️ [[CFG-07]] is adjacent and is NOT absorbed here
+
+CFG-07 is that `delivery_hold_s` is declared `int(30,900)` in `config.yaml`'s
+schema while the measured worst legitimate episode at the shipped threshold is
+**35.0 s** — so the schema offers an operator a value (30) that is **below a
+measured legitimate episode** and would report on ordinary saturation.
+
+**This card makes it worse, and by exactly one control cycle.** The relative arm
+takes the worst legitimate episode from **35.0 s to 40.0 s** at the shipped
+fraction, so the band of schema-legal but false-tripping values widens from
+`[30, 35]` to `[30, 40]`. The floor the schema should have is now **40 s**, not
+35 s. It is not fixed here: raising a schema floor is an option-surface change
+with its own compatibility argument (an existing config carrying `35` would start
+being refused), and folding it into a safety fix would hide it. **CFG-07 should
+be re-derived against 40.0 s, not 35.0 s.**
+
+⚠️ **WHEN IT HAS TO CLOSE: BEFORE INVERTER B IS COMMISSIONED (~2026-09-12).**
+Not before the next deploy. The band is **unreachable at N=1** — `delinquent()`
+returns `{}` for every input at a roster of one, so no value of
+`delivery_hold_s` can false-trip on the install that is running today — and the
+live install is N=1 until B exists. Commissioning B is the moment the band
+becomes reachable, and that is the deadline this card hands it.
+
+⚠️ **AND CFG-07 HAS NO DECISION DOCUMENT IN THIS TREE.** It exists as a
+`[[CFG-07]]` mention in `CHANGELOG.md` and in this section, and nowhere else —
+so there is no file for the 40.0 s to be recorded in, and a reader who goes
+looking will find a card reference with no record behind it. Whoever takes
+CFG-07 writes `DECISIONS/CFG-07-*.md` first; the number it has to answer to is
+**40.0 s**, measured here, and the harness that produces it is
+`admin/goodwe:sim/delivery_window.py --campaign --fractions`.
+
+## 7. Provenance of every number in this document
+
+| number | where it came from |
+|---|---|
+| 199/200, 0/200, 200/200, shares 1505/125 | `test_packshare.py` part 8b, run |
+| 35.0 s absolute-arm floor, and its argmax | `delivery_window.py --campaign`, re-run here, reproduces SAFETY-42 §3.3 |
+| every row of §3.1 | `delivery_window.py --campaign --fractions`, two passes over the same campaign with `FRACTIONS` changed between them |
+| 1000 W / 6500 W suppression windows | part 8b's arithmetic, asserted there |
+| 185 s, 5.29× | SAFETY-42 §3.3 / TEST-21, unchanged by this card |
+| 4.62× | 185 / 40.0, printed by the campaign |
+| 0/200 on both positive controls, and B's 0 W share at the cut-out | part 8b, run and asserted |
+| 0 W cut-out share / 125 W wedged share | part 8b, asserted and printed |
+| 77/125/133/154 W (the wedged arm's four distinct shares) | the same loop re-run in review with per-cycle recording — NOT part 8b's printed line, which shows the last cycle only |
+| M12/M13/M14 named-check counts (7 / 7 / 6) | re-run in review on fresh `:ro` trees, after the `KeyError` abort in part 1 was guarded |
+| 84 single-unit rosters | part 8b, run |
+
+⚠️ Every two-inverter figure above is **two copies of inverter A**.
@@ -0,0 +1,379 @@
+# UX-04 — what the Configuration tab renders, and the pin that was never built
+
+**Status: DECIDED 2026-09-07.** Three decisions and one confirmation, in the
+order [[UX-01]]'s unmet acceptance criteria are numbered:
+
+1. **UX-01 AC 3 (hide the optional fields behind HA's toggle) is WITHDRAWN as
+   unsatisfiable**, not deferred. The target was arithmetically unreachable, and
+   the part of it that *was* reachable would be bought with the upgrade path.
+   The measurement is now **graded** in `tools/supervisor_shape_probe.py` §F
+   rather than printed, so the withdrawal cannot be quietly reversed.
+2. **UX-01 AC 4's third clause is now BUILT**: `test_defaults_pin.py`, an AST
+   check, with its red direction executed on every run.
+3. **UX-01 AC 5 / the CFG-04 double-literal hole is CONFIRMED CLOSED on
+   `release/1.0`, by [[UX-02]]'s `bcee235`, not by UX-01's `fef6ef2`.** The
+   attribution is corrected here; no work was re-done.
+4. **The 8 per-entry `inverters:` translations STAY.** They do not reach the
+   Supervisor tab and cannot be made to; [[UX-02]]'s ingress panel is where they
+   become visible.
+
+Filed with UX-04. Card `6a9ec94617e9c48ed74ca5dd`. Every number below was
+measured on `release/1.0` at `6e7b233` on 2026-09-07, against
+`ghcr.io/home-assistant/amd64-hassio-supervisor:2026.08.0` — the version
+`ha info` reports on the live install — and against the 26-file check suite in
+`goodwe-addon-checks:1`. Nothing here contacted the live Supervisor at
+192.168.2.158, posted options, or restarted anything.
+
+---
+
+## 1. Hiding: withdrawn, and why that is the *safe* answer
+
+### 1.1 What AC 3 claimed
+
+> "The visible count drops to ~15, with the rest behind HA's *show unused
+> optional configuration options* toggle."
+
+### 1.2 What the tab actually renders
+
+`config.yaml` declares **71 keys under `options:` and the same 71 under
+`schema:`**. The set that HA's toggle is *able* to hide — schema-only keys — is
+therefore **empty**:
+
+```
+    hideable = set(schema) - set(options)  ->  []        (0 of 71)
+```
+
+On the deployed 0.7.0 tree the same measurement is 70/70. Nothing was hidden
+before UX-01, nothing is hidden after it, and no line of `fef6ef2` moved a key
+either way. AC 3 is not partially met; it is unstarted.
+
+### 1.3 The rule that decides it, replayed rather than paraphrased
+
+From the frontend bundle shipped **in this image**,
+`api/panel/frontend_latest/5680.*.js` (frontend 20250925.1):
+
+```js
+_filteredSchema = (opts, schema) => schema.filter(e => e.name in opts || e.required)
+```
+
+`opts` is the add-on's options **as Supervisor serves them** — the operator's
+stored blob after `options:` has been deep-merged into it. So a field is hidden
+only when it is **both** absent from `options:` **and** not `required`. The
+probe now runs that filter on this manifest rather than quoting it.
+
+### 1.4 Reason one: the target was never reachable
+
+Supervisor's own `UiOptions` marks `required` on **59** of the 71 fields; 12 are
+optional. The filter above shows a required field unconditionally. So the
+**floor** — what the tab renders with `options:` emptied *altogether*, which is
+the most hiding that is even expressible — is 59 fields.
+
+| `options:` block | fields rendered |
+|---|---|
+| as shipped (71 keys) | **71** |
+| emptied entirely (0 keys) | **59** |
+| AC 3's target | ~15 |
+
+AC 3 asked for a number roughly four times below the floor. No arrangement of
+`options:` reaches it. The criterion could not have been met by any
+implementation; it was mis-specified, and the honest disposal of a
+mis-specified criterion is withdrawal, not a partial gesture toward it.
+
+### 1.5 Reason two: the 12 hides that *are* expressible cost the deep-merge
+
+Moving a key out of `options:` is not free. `App.options` is
+`deepmerge(config.yaml defaults, the stored user options)`, so **a key absent
+from `options:` has no default for the Supervisor to merge**.
+
+That is exactly the mechanism the 0.9.0 upgrade rests on. From `CHANGELOG.md`,
+the pre-deploy check run against this release on 2026-09-07:
+
+> The reference install's stored blob has **70 keys, not 71**.
+> `delivery_hold_s` arrived with SAFETY-42 in 0.8.0, so an install running
+> 0.7.0 has never stored it — and it is REQUIRED in the schema, no trailing
+> `?`, which is the shape that becomes **exit 78**. It resolves to its
+> `options:` default of **185** because the Supervisor deep-merges.
+
+Hiding a key is done by taking it *out* of `options:` and marking it `?`. That
+is, precisely, the edit that removes a key's merge default — the thing standing
+between a 0.7.0 box and exit 78 on upgrade. It would be traded for at most 12
+fewer rows on a form, on a site where the add-on is holding a real house's grid
+exchange at zero.
+
+### 1.6 Reason three: it would promote 27 unpinned literals to load-bearing
+
+And the cost compounds with §2 of this document. `test_defaults_pin.py` part 4
+ledgers **27 (module, key) pairs** in `app/` that still answer a missing option
+with an in-code literal. Today those literals are unreachable on a Supervisor
+install, because the deep-merge guarantees every `options:` key is present in
+the served blob. Remove keys from `options:` to hide them and those literals
+stop being dead code: they become **the actual source of the defaults on a live
+install** — the CFG-03 defect the same card's AC 4 exists to police, made
+load-bearing by the card's own AC 3.
+
+### 1.7 What now grades this
+
+`tools/supervisor_shape_probe.py` §F. The count was printed there from the day
+it was written — *"the Configuration tab renders 71 fields"* — and the probe
+exited 0 through UX-01's whole review while displaying the number that
+contradicted the criterion it was evidence for. **A measurement that is
+displayed but not asserted is how this passed review.** Three assertions
+replace the bare print:
+
+```
+PASS  every one of the 71 `schema:` keys is ALSO in `options:`, so schema-only
+      - the only set HA's toggle can hide - is EMPTY. Hideable: []
+PASS  ...so replaying the frontend's own _filteredSchema on the served options
+      hides NOTHING: 71 of 71 fields render
+PASS  ...and with `options:` EMPTIED ENTIRELY - the maximum hiding that is
+      expressible - the tab still renders 59 required fields. THAT is the
+      floor, and AC 3 asked for ~15.
+```
+
+Every number in them is computed from the manifest and from Supervisor's own
+`required` flag; none is typed. **The red direction was run**: deleting one key
+(`price_feed_url`) from `options:` and nothing else turns the first two to FAIL
+— `Hideable: ['price_feed_url']`, `70 of 71 fields render` — and the probe
+exits 1. The counterfactual is isolated: one key varied, nothing else.
+
+So if a future editor empties `options:` to shorten the form, this goes red and
+somebody has to read §1.5 before merging.
+
+---
+
+## 2. The defaults pin: `test_defaults_pin.py`
+
+### 2.1 What AC 4 asked for and what shipped
+
+AC 4 has three clauses.
+
+| clause | state before UX-04 |
+|---|---|
+| "every `schema:` key absent from `options:` has exactly one defaults-table entry" | **vacuous** — that set is empty (§1.2), so the clause is true of any tree, including one with no defaults at all |
+| "the table changed → red" | **works**, six times over, once per mirror in its own suite |
+| "no module carries its own literal fallback for a schema default — an AST check, not a grep" | **never built** |
+
+The only literal scan in the repository before this card is
+`test_validate.py:1025`: a `re.findall` restricted to eight `maintenance_*`
+keys across three named files. It is a grep, it predates UX-01, and it was not
+written for this.
+
+### 2.2 What an "in-code literal fallback" is, structurally
+
+A call in `app/` taking a string literal that names a `schema:` key, with a
+constant in the next positional slot:
+
+```python
+opts.get("max_w", 2000)            _number(opts, "max_w", 2000.0)
+```
+
+Found by walking `ast.Call` nodes — so a multi-line call and a renamed `opts`
+are found, and the same text inside a comment or docstring is not. Two
+exclusions, both structural and both argued at the site:
+
+* **`getattr(obj, "max_w", 0.0)`** reads an attribute off an object, not an
+  option off a blob. Excluded by callee name.
+* **the empty string.** `opts.get("static_profile", "")` is Python's "not
+  stated", and all six such sites in `validate.py` normalise it immediately —
+  `str(opts.get(k, "") or "").strip()`. It cannot be a stale copy of a *number*,
+  which is the whole of the CFG-04 defect class. Excluded by value. **`0` is
+  not excluded**: `slew_w: 0` and `integrator_max_w: 0` are real shipped
+  defaults with real meanings.
+
+### 2.3 What it grades, measured on `release/1.0` @ `6e7b233`
+
+```
+2. the mirrors, DISCOVERED from the tree rather than listed here
+    6 table(s): circulation.py:CIRC_OPTION_DEFAULTS, delivery.py:DELIVERY_OPTION_DEFAULTS,
+                maintenance.py:MAINT_OPTION_DEFAULTS, peakguard.py:PEAK_OPTION_DEFAULTS,
+                price.py:PRICE_OPTION_DEFAULTS, watchdog.py:SIGN_OPTION_DEFAULTS
+    30 mirrored key(s) of 71 in `schema:`
+3. the AST scan
+    scanned 29 modules, 34 literal fallback site(s): 0 over mirrored keys,
+    34 over unmirrored ones
+4. the CFG-03 ledger
+    27 (module, key) pair(s) carry a literal for a key no mirror answers for
+```
+
+Four properties, none of them vacuous:
+
+1. **The mirrors are discovered, not listed.** Six suites already pin their own
+   table to `config.yaml` — and every one of those pins names its table by
+   hand, so a *seventh* table added tomorrow would be pinned by nobody and
+   nothing would say so. This walks `app/` for module-level `*_OPTION_DEFAULTS`
+   assignments and grades whatever it finds against `config.yaml`. That is the
+   property the six hand-written pins cannot have.
+2. **No literal survives for a key a mirror already answers for.** 30 keys,
+   0 violations.
+3. **The CFG-03 ledger is FROZEN, by (module, key) pair, not by count.** A
+   count goes green again the moment somebody deletes one site and adds
+   another. Line numbers are deliberately *not* in the key: a ledger that goes
+   red on an unrelated reformat is a ledger people learn to re-bless without
+   reading. The 27 pairs are the open CFG-03 backlog; they are **not** removed
+   here, because that is behaviour change across `main.py`, `p1.py`,
+   `inverter.py` and `validate.py` that no single card should make in passing.
+   The set can now only shrink.
+4. **Every remaining literal is graded against `config.yaml`** — the half that
+   catches CFG-04's actual failure mode, two copies of one number and one moves.
+
+### 2.4 One divergence found, and it is new
+
+```
+      main.py  target_grid_w: code 0 vs config.yaml -10
+```
+
+`main.py:290` reads `target_grid_w=float(opts.get("target_grid_w", 0))` while
+`config.yaml` ships **-10**. A dict handed to `Controller` without the key holds
+net exchange at 0 W instead of the shipped -10 W bias into import. **It cannot
+reach an operator**: on a Supervisor install the deep-merge guarantees the key
+is present, and the only paths that hand over a partial dict are
+`sim/inject.py` and this repository's own fixtures. It is recorded in the check
+as the single permitted divergence, with the reason, rather than corrected —
+changing it alters the control loop's target and belongs to a card that grades
+the loop. A *second* divergence goes red.
+
+### 2.5 ⭐ The red direction, shown rather than claimed
+
+A check that has never been shown failing is not a check — which is exactly
+what clause one of this AC was. Two things follow from that:
+
+**(a) The red direction runs on every invocation.** Part 5 feeds the scanner a
+synthetic module containing a deliberate violation of each rule and *requires*
+it to be flagged, in-process, through the same `literal_fallbacks()` the graded
+parts call. A scanner that goes blind — a renamed helper, an `ast` change, a
+`SCHEMA` that comes back empty — stops working here too, and the total site
+count is additionally graded against a floor.
+
+**(b) A literal was planted in the real tree and the check went red.** Planted
+at `main.py:290`, `float(opts.get("maintenance_charge_w", 2000))` — a mirrored
+key:
+
+```
+3. the AST scan: no module carries its own literal fallback for a MIRRORED schema default
+    scanned 29 modules, 35 literal fallback site(s): 1 over mirrored keys, 34 over unmirrored ones
+  ok   the scanner is not blind: it found 35 sites (floor 20)
+      main.py:290  get('maintenance_charge_w', 2000)
+  FAIL no literal fallback survives for any of the 30 keys a mirror already
+       answers for - those sites must index the mirror
+
+1 FAILED     (exit 1)
+```
+
+And part 4's two directions, planted the same way with
+`opts.get("p1_host", "x")` and `opts.get("slew_w", 999)`:
+
+```
+4. the CFG-03 ledger
+    28 (module, key) pair(s) carry a literal for a key no mirror answers for
+  FAIL no NEW module/key pair has started carrying a literal fallback [('main.py', 'p1_host')]
+      main.py  p1_host: code 'x' vs config.yaml ''
+      main.py  slew_w: code 999 vs config.yaml 0
+      main.py  target_grid_w: code 0 vs config.yaml -10
+  FAIL every remaining literal still AGREES with config.yaml, except the 1
+       divergence(s) recorded above it
+
+     (exit 1)
+```
+
+The second of those is the CFG-04 defect class itself: a literal that has
+drifted away from `config.yaml` while the file still says something else. Both
+plants were removed; the tree is byte-identical to `6e7b233` apart from this
+card's own files.
+
+---
+
+## 3. CFG-04: closed on `release/1.0`, by UX-02, not by UX-01
+
+UX-01's AC 5 claimed the double-literal `max_w` hole. It did not close it.
+`fef6ef2` touched `config.yaml`, `translations/`, `test_translations.py` and the
+probe — **no defaults code at all** — and on the 0.7.0 tree the triple mutation
+survives: 21 files, 0 RED.
+
+**Confirmed on current `release/1.0` @ `6e7b233`, 2026-09-07.** The same triple
+mutation, `2000 -> 2500` at the three descendants of 0.7.0's
+`inverter.py:438` / `validate.py:383` / `validate.py:684`:
+
+| 0.7.0 | `release/1.0` @ `6e7b233` |
+|---|---|
+| `inverter.py:438` `max_w=float(opts.get("max_w", 2000))` | `inverter.py:640` `raw = opts.get("max_w", 2000)` |
+| `validate.py:383` `cap = _number(opts, "max_w", 2000.0)` | `validate.py:466`, same line |
+| `validate.py:684` `max_w = _number(opts, "max_w", 2000.0)` | `validate.py:836`, same line |
+
+Result — 26 files, **RED at `test_selfconfig.py`**, exit 1:
+
+```
+1 FAILED: ⚠️ CLEARING max_w ON THE LIVE SHAPE IS REFUSED - rule 10 catches the
+rail collapsing to the 2000 default under an explicit 400 W slew, so the box
+cannot be emptied into a quietly halved rail
+```
+
+`git log -L 310,325:goodwe_controller/test_selfconfig.py` attributes that check
+to **`bcee235`, "UX-02: pin both halves of the max_w fix, because neither was
+graded", glenn schrooyen, 2026-09-06** — not to UX-01. The hole ships closed
+with the pending 0.9.0 deploy. **The AC is met by UX-02.** No work was re-done
+here; the mutation was run once, on a throwaway copy, to confirm it.
+
+---
+
+## 4. The 8 per-entry translations: they stay
+
+### 4.1 The measurement, and it is negative
+
+`translations/{en,nl}.yaml` carry a name and a description for all 8 per-entry
+keys of the repeatable `inverters:` section. They are loaded by Supervisor's own
+loader, they pass `SCHEMA_APP_TRANSLATIONS`, and `test_translations.py` grades
+their coverage. **They do not reach the screen**, and this is not a defect in
+them. `UiOptions` gives `inverters` `type: "schema"` with `multiple: True`, and
+the frontend's `_convertSchemaElement` sends exactly that case to a selector
+rather than to a nested `ha-form`:
+
+```js
+if ("schema" === t.type && !t.multiple) return {...type:"expandable"...}   // not us
+if ("schema" === t.type) { const o = {}; for (const a of t.schema)
+    o[a.name] = {required: a.required, selector: ...}; return {object:{multiple:t.multiple, fields:o}} }
+```
+
+and `ha-selector-object` labels those fields `fields[name].label || name`, while
+`_convertSchemaElementToSelector` sets **neither `label` nor `translation_key`**.
+So an operator sees raw `name`, `setpoint_entity`, `max_w`, `bms_soc_floor`. The
+probe already asserts this, negatively and by name.
+
+### 4.2 Decision: keep them. Four reasons.
+
+1. **The Supervisor tab cannot be made to render them.** The gap is in
+   somebody else's converter, not in this repository. There is no change here
+   that fixes it; deleting the translations does not fix it either.
+2. **[[UX-02]]'s ingress panel is where they become visible**, and it renders
+   the roster itself rather than handing it to `ha-selector-object`. The
+   `fields:` block is the shape Supervisor's own schema defines for exactly
+   this, and it is what a surface that wires `label` through consumes.
+3. **They cost nothing.** No runtime path reads them; they are two YAML blocks
+   in files the Supervisor already parses, and the parse is graded green.
+4. **Deleting them would remove the only thing that catches a per-entry key
+   added without a translation.** `test_translations.py` grades top-level and
+   per-entry coverage symmetrically; drop the per-entry half and SAFETY-32's
+   `bms_soc_floor` — a *required* per-entry key added after UX-01 — would have
+   joined the roster with no label and nothing to say so.
+
+This is a written decision, not a change. The tree is unmodified in
+`translations/`.
+
+### 4.3 Not re-filed
+
+`tools/supervisor_selfapi_probe.py` has been known red since OPT-09 — it
+hardcodes `len(_sub) == 7` against 8 per-entry keys. That is **TOOL-04**,
+already filed, and it is not re-filed or fixed here. It is named only so a
+reader of §4.1 knows why the *other* probe is the one quoted.
+
+---
+
+## 5. What this card did not do
+
+* It did not remove a key from `options:`. See §1.5 — the deep-merge default is
+  what makes the pending 0.9.0 upgrade safe on a 0.7.0 box.
+* It did not pay down the 27-pair CFG-03 ledger. See §2.3.
+* It did not correct `main.py`'s `target_grid_w` fallback. See §2.4.
+* It did not touch `translations/`. See §4.2.
+* It did not re-do CFG-04. See §3.
+* It did not touch hardware, the live Home Assistant at 192.168.2.158, or
+  `sim/`. Both probe runs were throwaway containers over a read-only mount.
@@ -110,7 +110,8 @@ so Home Assistant's *Configuration* tab renders human labels and a description
 for every option instead of the raw option keys. What that does and does not buy
 you is worth stating plainly, because it was **measured against the Supervisor
 image this site runs** (`ghcr.io/home-assistant/amd64-hassio-supervisor:2026.08.0`)
-rather than assumed — `tools/supervisor_shape_probe.py` section F prints the
+rather than assumed — `tools/supervisor_shape_probe.py` section F GRADES (it
+used only to print) the
 whole rendered list, in both languages.

 | | |
@@ -1573,6 +1574,185 @@ battery power, circulating power, state of charge, maintenance phase and
 controller status by MQTT discovery. This is observability only — the controller works fine without a
 broker, and MQTT problems can never affect control.

+## Before you commission inverter B: `tools/rehearse_roster.py` (ARCH-02)
+
+The day the `inverters:` roster is declared, two things change that are
+**one-way on a live box and invisible until after the restart that applies
+them**: the store migration renames the keys that hold the SAFETY-02 latch and
+the maintenance history, and the entity ids Home Assistant composes change
+shape. This tool runs the **real `Controller`** through that transition against
+a **copy** of the state file and reports the difference.
+
+    docker run --rm -e PYTHONDONTWRITEBYTECODE=1 -v <repo>:/w:ro \
+      goodwe-addon-checks:1 sh -c 'cd /w/goodwe_controller && \
+      python3 tools/rehearse_roster.py \
+        --state /w/state-copy.json \
+        --before /w/live-options.json \
+        --after  /w/roster-options.json \
+        --live-entities /w/live-entities.txt \
+        --reverse --round-trip'
+
+Exit `0` clean · `1` at least one finding · `2` could not run. ⚠️ **`2` is not a
+pass.**
+
+### The owner steps, because this session could not take them
+
+⚠️ **The rehearsal is only as good as the state it is pointed at, and nothing in
+this repository can fetch that.** SSH to the live box is blocked by policy and
+that is correct — `/data/state.json` is the file the running add-on is writing.
+So three things have to be exported by hand, onto the machine that runs the
+check, **before** the roster is declared:
+
+1. **The state file.** From the Supervisor host:
+
+       ha addons info local_goodwe_controller     # confirm the slug first
+       docker cp addon_local_goodwe_controller:/data/state.json ./state-copy.json
+
+   or take it out of a full add-on backup, which is the same bytes. ⚠️ **Copy
+   it, never move it, and never write to the original.** The tool opens the path
+   it is given exactly once and does all of its writing on a scratch copy — but
+   the safe habit is to hand it a file the add-on has never heard of.
+
+2. **The options blob**, as `--before`. It is not reachable through the Home
+   Assistant REST API and no entity carries it; export it from the Supervisor,
+   the same way `tools/predeploy_check.py` needs it. Write the roster you intend
+   to deploy into a second file and pass it as `--after`.
+
+3. **The entity list**, as `--live-entities` — one entity id per line:
+
+       Developer tools -> Template:
+       {{ states | map(attribute='entity_id') | select('search','goodwe') | join('\n') }}
+
+   ⚠️ **Omitting it is a REFUSAL (exit 2), not a quieter run.** The orphan check
+   is graded against the ids the install actually has, and without that file it
+   is not graded against reality at all — so the tool will not run rather than
+   hand back a clean bill of health for a half of the transition it did not
+   check. If you genuinely have to run without it, pass `--no-live-entities`:
+   the run then carries a standing banner in the report header **and** on the
+   exit line saying the orphan check was not graded. The asymmetry used to be
+   the other way round — an unreadable file refused, a missing flag passed with
+   exit 0 and no disclosure at all — and the owner runs this once, at
+   commissioning, under time pressure.
+
+   The install carries roughly 49 `goodwe*` entities and this add-on announces
+   eight of them. The tool separates the ids it owns from the ids it does not,
+   rather than claiming orphans it cannot cause.
+
+### The reference, which is loaded on every run
+
+`--reference` defaults to `tools/rehearsal/announced-reference.txt` and is read
+**whether or not** you pass an entity list. It records what the build is
+supposed to announce, three columns per row:
+
+    <entity_id>   <unique_id>   <device identifier>
+
+⚠️ **All three, and the second and third are the ones an entity_id-only list
+cannot see.** The tool's own before/after diff is composed twice by the same
+build, so a change to the composition moves both sides equally and the diff
+reports nothing. Worse, two changes orphan an install *without moving a single
+composed entity_id*: renaming `mqtt.BASE` (every `unique_id` changes) and
+changing `DEVICE["identifiers"]` (HA sees a new device). Home Assistant keys its
+entity registry on `unique_id` and its device registry on the identifier, so
+either one creates a **new** entity beside the old — and because discovery is
+published `retain=True` with no removal payload, the old config stays on the
+broker and the old entity stays in the registry, still named by every automation
+that referenced it, never updating again. Both were measured coming back
+**exit 0** before the reference carried three columns.
+
+`object_id` is deliberately **not** a column: `app/mqtt.py` records that HA
+2026.8 does not honour it in a discovery payload, so renaming one is a true
+no-op, and `test_rehearse.py` requires that mutation to **survive** — otherwise
+"the reference notices any edit to `mqtt.py`" would be indistinguishable from
+"the reference notices what HA keys on".
+
+⚠️ **No result in this repository was produced against real state.** Everything
+recorded here and in `test_rehearse.py` was measured on the synthetic fixtures
+in `tools/rehearsal/`.
+
+### What it confirms on the synthetic fixtures
+
+⚠️ **This is a SECOND INSTRUMENT reaching a conclusion this repository already
+holds, not a discovery.** `test_maintenance.py` section 13 — *"RECORDED, NOT
+FIXED: SAFETY-02's migration has the same wrong-namespace hole, and it is a
+LATCH"* — already states it, already names `namespaced_latched`, and already
+carries the sentence *"the guard is defeated by the very key the defect wrote"*.
+That file is on `release/1.0` and this card does not touch it. What is new here
+is that the finding is now reachable **from a state file the owner supplies**
+rather than only from a fixture built in a test, and that it is reached by a
+different route: a full `Controller` start diffed against its predecessor,
+rather than assertions over a `MemStore`. Two instruments, one answer.
+
+**The finding.** A pre-DEV-02 flat store that declares a multi-inverter roster
+comes up with its SAFETY-02 latch **DISARMED**, unless the first roster entry
+happens to be named `inverter-a`. `Controller.__init__` calls
+`migrate_unit_keys(store, UNIT_A)` with a hardcoded constant, so the latch lands
+in `unit/inverter-a/sign_tripped`; a roster of `["garage", "shed"]` reads
+`unit/garage/sign_tripped`, which is absent. And `_grade_store_migration()` —
+the guard written for exactly this — sees the migration's own misdirected key
+satisfy its `namespaced_latched` test and returns without re-arming. The file
+says `sign_tripped: true`, the banner is green, and nothing is held at 0 W.
+
+**Both routes reach it, and they are NOT the same defect.** The rehearsal
+reports them as separate legs because they need separate fixes:
+
+| leg | how it gets there |
+|---|---|
+| `SAME-STEP` | the roster is declared in the same step as the upgrade. The migration runs, and is handed `UNIT_A`. |
+| `STAGED` | the flat config runs once under 0.7.0+ first. That start stamps `schema_version = 1`, so the roster start's migration takes its `from_version >= SCHEMA_VERSION` early return and **the latch never moves at all**. |
+
+⚠️ **So "hand `migrate_unit_keys()` the roster's own first name" is NOT the
+whole fix, and the rehearsal measures that rather than asserting it.**
+`test_rehearse.py` part 5b builds exactly that change and reports:
+
+    LEG STAGED: DISARMED
+    LEG SAME-STEP: clean
+
+The staged path never reaches the line that was changed. A fix that stops there
+closes the half an operator is least likely to take and leaves the half that
+"upgrade first, declare the roster later" produces. The guard side —
+`namespaced_latched` counting a namespace that is not on the roster — is what
+covers both, and a change there turns every SAFETY-44 check in part 5 red, which
+is what a fix should do to a pin that specifies a defect.
+
+The fix belongs to its own card, shown red first. Part 5 pins the current
+behaviour as a requirement so it cannot drift green by accident.
+
+### The reverse direction: the retreat is supported, the round trip is not
+
+If commissioning B goes wrong, **backing out to the flat single-inverter config
+is supported and does not corrupt state.** Measured on the fixtures:
+
+* the SAFETY-02 latch and the first pack's maintenance history both survive,
+  because `store.py` rule 2 keeps the flat spellings **live** rather than merely
+  present — `maintenance.UnitHistory.mirror_flat` writes them on every change,
+  so a single-FSM add-on finds a current history rather than one dated the day
+  the roster was declared;
+* **it orphans eight Home Assistant entities** — the four per-unit rows for each
+  of the two units. **Measured:** the add-on stops announcing them and stops
+  publishing their state, so their history gaps and any automation whose
+  trigger names them stops firing, silently and without an error.
+
+  ⚠️ **What is NOT measured, and was previously overstated here:** whether
+  advancing again restores them. An earlier draft of this section said the
+  orphaning was "not recoverable by advancing again: HA keeps the id it composed
+  at first discovery" — that reasoning cuts the other way, since keeping the
+  composed id is precisely what would let a re-advertised entity land back on
+  it, and discovery is published `retain=True` with no removal payload. Nothing
+  in this repository has tested a re-advance against a real Home Assistant, so
+  the claim is withdrawn rather than restated in the other direction. Plan the
+  retreat around the measured part: **the eight rows stop updating for as long
+  as you are retreated.**
+
+⚠️ **What is NOT supported is the round trip.** A maintenance cycle that
+completes *while the install is rolled back* is **discarded** when it
+re-advances: `maint_schema_version` is already stamped, so `migrate_maint_keys()`
+takes its early return and the pack reads the namespaced history from before the
+retreat. The pack is then recalibrated on a schedule that does not know it
+already was. **So: retreat freely, but treat a retreat as ending the current
+maintenance bookkeeping — and if a cycle ran during it, correct
+`unit/<first>/last_completed` by hand before advancing again.** This is a stated
+decision, not a discovery made at the inverter.
+
 ## Before you deploy: `tools/predeploy_check.py` (PKG-09)

 Two ways a deploy has gone wrong on this system, both of which the Supervisor
@@ -225,6 +225,78 @@ same place twice, the check is decoration.
  forever, *including after the call site is fixed*. The check was true and
  passed for the wrong reason. Habit **9** below is what falls out of it.

+* ⚠️ **ARCH-02's OWN REHEARSAL HARNESS, CAUGHT BY ITS OWN MUTATION BATTERY —
+  the THIRTEENTH instance, and the third in an INSTRUMENT rather than in a
+  check.** `tools/rehearse_roster.py` reports orphaned Home Assistant entities
+  by composing the ids the add-on announces **before** the roster switch and
+  **after** it, and diffing. Both sides are composed by the same build. So the
+  two mutations that change the id composition itself — renaming the MQTT device
+  (`"GoodWe"` -> `"GoodWe Controller"`), renaming one site sensor (`"Grid power"`
+  -> `"Grid power W"`) — moved **every id on both sides equally**, and the
+  harness reported *"no orphans"* about a change that orphans the entire live
+  entity list. It read exactly like a clean transition. The first draft's
+  M5/M6 came back `exit 0` and `exit 2`, and neither was the failure they name.
+
+  The fix is habit 2 with a filename: **the reference has to be data that does
+  not move with the code.** `--live-entities` takes a list of the ids the
+  install actually has, `tools/rehearsal/live-entities-recorded.txt` is the
+  recorded stand-in that keeps it live in CI, and a build that announces an id
+  the list does not carry is now a `COMPOSITION DRIFT` finding rather than a
+  silence. ⚠️ **The anchor moved for the same reason**: it had been
+  `compose(mqtt.DEVICE["name"], p1["name"]) == "sensor.goodwe_p1_sample_age"`,
+  which any renaming of the device satisfies by moving both operands, and it is
+  now composed from two recorded **literals**.
+
+  Ask of an instrument what habit 4 asks of a check: *where did each side of
+  this comparison come from* — and if a mutation of the subject moves both
+  sides, the instrument is decoration.
+
+  ⚠️ **AND REVIEW FOUND TWO MORE IN THE SAME CARD, ONE OF THEM INSIDE THE FIX
+  FOR THE OTHER — the FOURTEENTH and FIFTEENTH.** Both are the pin, not the
+  tool, and both are the same sentence in a new costume: *a check whose data
+  cannot reach the failure it names.*
+
+  * **Fourteenth — a COUNT cannot tell one red leg from two.** The rehearsal
+    reports two independent routes to the disarmed latch (`SAME-STEP` and
+    `STAGED`) and prints every finding **twice per red leg**: once in that
+    leg's own block and once in the closing summary. The check named *"NOT THE
+    SAME-STEP PATH ALONE"* was written `out.count("LATCH CAME UP DISARMED")
+    >= 2`, so **one** red leg always satisfied it. Measured on three builds:
+    shipped code (both legs red) → 4; `migrate_unit_keys()` handed the roster
+    name (SAME-STEP fixed) → 2, **passes**; a guard-side fix (both fixed) → 0.
+    The check reported `ok` on the build whose only remaining failure *is* the
+    path it says is not the only one. The fix is to parse the legs and require
+    the finding in **each**, which is habit 2 again: grade the thing, not a
+    total over the thing.
+
+  * **Fifteenth — and it is INSIDE THAT FIX, which is the third time this file
+    has had to write that sentence.** The first leg-splitter opened on a
+    `LEG <id>` marker and never closed, so every line after the *last* marker
+    was attributed to that leg — including the closing summary, which repeats
+    every finding from every leg. The last leg therefore read red whenever
+    **any** leg was red, and the new per-leg check was the old count with extra
+    steps. It was caught by building the one fixture that separates them: a
+    tree in which `SAME-STEP` is genuinely clean and `STAGED` is not. The tool
+    now prints `END LEG <id>` and the splitter closes on it.
+
+    The rule that falls out, and it is the cheap one: **a per-thing check that
+    has never seen the things DISAGREE is a total.** Part 5b exists to make
+    them disagree on every run.
+
+    ⚠️ **AND ONE THAT HAS ONLY SEEN THEM DISAGREE IN ONE ARRANGEMENT IS A TOTAL
+    OVER THAT ARRANGEMENT.** Review found this by attacking the fix itself. Part
+    5b does make the legs disagree, on every run - and was still blind to a build
+    that merely REORDERS them, because it had only ever seen the clean leg
+    printed LAST. Three edits, two of them ordinary refactors (reorder the report
+    legs, reword a summary line), produced `all checks passed` from a splitter
+    that closed ZERO legs.
+
+    A fixture cannot enumerate the arrangements, so the STRUCTURE is asserted
+    instead: a leg may not open while one is open, and every leg that opens must
+    be explicitly closed. That fires on the FIRST edit of all three attacks,
+    whatever the order or the wording. **When a fixture cannot enumerate the
+    arrangements, assert the invariant they share.**
+
 * **CFG-04's `tools/supervisor_shape_probe.py` §B** — and this one is the rule
  reaching outside the suite. The probe is the committed evidence that
  `config.yaml`'s `inverters: []` default is load-bearing, and four documents
@@ -264,36 +264,66 @@ DELIVERY_OPTION_DEFAULTS = {
    # such family, so reading its numbers across would have understated this.
    #
    # Every run was executed a second time at a roster of ONE, on the same
-    # stimulus, as the control. Worst CONTINUOUS legitimate episode:
+    # stimulus, as the control.
+    #
+    # ⚠️ RE-RUN BY TEST-21 (2026-09-07) AND THE HEADLINE DID NOT REPRODUCE. The
+    # harness that produced the original table was never committed, which is
+    # PROC-01's shape; it is committed now, in the OTHER repository, as
+    # `admin/goodwe:sim/delivery_window.py` (it imports `plantsim` and
+    # `frame_check`, which that repo owns - restating them here would be CI-08's
+    # coupling, and CI-08 has broken three times). Re-derive it with:
+    #
+    #     cd goodwe/sim && GOODWE_ADDON_REPO=/path/to/goodwe-addon \
+    #         python3 -B delivery_window.py --selftest --campaign --blag
+    #
+    # Worst CONTINUOUS legitimate episode. Second block is the re-run:
    #
    #     threshold   reversals   flapping   weighted   N=1 control (1890 runs)
-    #        100 W       40.0 s     0.0 s      75.0 s        0.0 s
-    #        250 W       55.0 s     0.0 s      60.0 s        0.0 s
-    #      * 500 W       20.0 s     0.0 s      30.0 s        0.0 s
-    #        700 W       20.0 s     0.0 s      20.0 s        0.0 s
-    #       1000 W       15.0 s     0.0 s      15.0 s        0.0 s
-    #       2000 W       10.0 s     0.0 s       5.0 s        0.0 s
+    #        100 W       40.0 s     0.0 s      75.0 s        0.0 s   } 2026-09-06
+    #        250 W       55.0 s     0.0 s      60.0 s        0.0 s   } harness
+    #      * 500 W       20.0 s     0.0 s      30.0 s        0.0 s   } NOT
+    #        700 W       20.0 s     0.0 s      20.0 s        0.0 s   } COMMITTED
+    #       1000 W       15.0 s     0.0 s      15.0 s        0.0 s   } - see
+    #       2000 W       10.0 s     0.0 s       5.0 s        0.0 s   } DECISIONS
+    #
+    #     threshold   reversals   flapping   weighted   N=1 control (2640 runs)
+    #        100 W      120.0 s     0.0 s     165.0 s        0.0 s   } TEST-21
+    #        250 W       65.0 s     0.0 s     100.0 s        0.0 s   } 2026-09-07
+    #      * 500 W       30.0 s     0.0 s      35.0 s        0.0 s   } committed
+    #        700 W       25.0 s     0.0 s      30.0 s        0.0 s   } harness,
+    #       1000 W       15.0 s     0.0 s      25.0 s        0.0 s   } re-runnable
+    #       1500 W        0.0 s     0.0 s      20.0 s        0.0 s   }
+    #       2000 W        0.0 s     0.0 s       0.0 s        0.0 s   }
    #
    #   * is the shipped threshold, `Tuning.saturation_w`.
    #
    # ⚠️ THREE THINGS IN THAT TABLE ARE WORTH MORE THAN THE ROW THEY SIT ON.
-    # (a) The N=1 column is ZERO EVERYWHERE, in all 1890 runs at every
-    #     threshold. The compatibility promise is measured, not argued.
-    # (b) The FLAPPING column is zero everywhere too - and that is the family
-    #     that defeated SAFETY-30's site window (980 s at N=2, 1000 s at N=1, no
-    #     window inside the billing quarter surviving it). It cannot touch this
-    #     check, because a common stimulus makes BOTH units short at once and
-    #     the subset stops being proper.
-    # (c) THE CURVE IS NOT MONOTONE - 55.0 s at 250 W against 40.0 s at 100 W -
-    #     and that is a property of the reduction, not noise: a lower threshold
-    #     puts MORE units into the short set, which can turn a proper subset
-    #     into the whole roster and BREAK the run. Do not read it as a
-    #     sensitivity curve.
+    # (a) The N=1 column is ZERO EVERYWHERE, at every threshold, IN BOTH
+    #     CAMPAIGNS. The compatibility promise is measured, not argued, and it
+    #     is the part that reproduced most cleanly.
+    # (b) The FLAPPING column is zero everywhere too, in both - and that is the
+    #     family that defeated SAFETY-30's site window (980 s at N=2, 1000 s at
+    #     N=1, no window inside the billing quarter surviving it). It cannot
+    #     touch this check, because a common stimulus makes BOTH units short at
+    #     once and the subset stops being proper.
+    # (c) THE ORIGINAL CURVE WAS NOT MONOTONE - 55.0 s at 250 W against 40.0 s
+    #     at 100 W - and that was read as a property of the reduction: a lower
+    #     threshold puts MORE units into the short set, which can turn a proper
+    #     subset into the whole roster and BREAK the run. ⚠️ THE RE-RUN'S CURVE
+    #     IS MONOTONE. The mechanism is real and is still the reason not to read
+    #     either table as a sensitivity curve, but it is no longer something
+    #     these numbers demonstrate - it is an argument now, not a measurement.
    #
    # THE BOUNDS THE VALUE ANSWERS TO:
    #
-    #   LOWER, from the rig - 30.0 s, the worst continuous legitimate episode at
-    #   the shipped threshold across all 2640 runs.
+    #   LOWER, from the rig - 35.0 s (TEST-21; the original record said 30.0 s
+    #   and it did not reproduce), the worst continuous legitimate episode at
+    #   the shipped threshold across all 2640 runs. ⚠️ THIS IS NOW STRICTLY THE
+    #   BINDING LOWER BOUND: at 30.0 s it coincided with the firmware's veto
+    #   below and a reader could not tell which was doing the work; at 35.0 s it
+    #   is ABOVE it, and it is above the schema's own `int(30,900)` floor too -
+    #   so an operator who sets `delivery_hold_s` to 30 is below a MEASURED
+    #   legitimate episode. That is what "derived, do not change" is for.
    #
    #   LOWER, from the firmware - `inverter.LINK_STALE_S`, 30 s. A unit not
    #   executing its share for less than the firmware's own veto window is a
@@ -302,26 +332,47 @@ DELIVERY_OPTION_DEFAULTS = {
    #   second, louder alarm about a state already being handled - the
    #   SAFETY-15/SAFETY-19 failure exactly.
    #
-    #   MARGIN, from INVERTER B NOT EXISTING, AND IT IS MEASURED RATHER THAN
-    #   GUESSED. The campaign's B is A COPIED (`plantsim.INVERTER_B`,
-    #   `measured=False`), so the 30.0 s above is two identical machines.
-    #   Sweeping a DIFFERENT second unit through the weighted family: the worst
-    #   legitimate episode is insensitive to B's RAMP (30.0 s at half A's) and
-    #   scales with its LAG - 45.0 s at twice A's measured 10 s lag, 90.0 s at
-    #   four times it. That is the sensitivity the margin has to carry until
-    #   SPIKE-04 measures the real unit.
+    #   MARGIN, from INVERTER B NOT EXISTING. ⚠️ THIS IS A SENSITIVITY STUDY ON
+    #   A HYPOTHESIS ABOUT B, NOT A MEASUREMENT OF B, and the wording here used
+    #   to say "MEASURED rather than GUESSED", which overstated it. B is not
+    #   installed (~2026-09-12). The campaign's B is A COPIED
+    #   (`plantsim.INVERTER_B`, `measured=False`), so the 35.0 s above is TWO
+    #   IDENTICAL MACHINES and every row below is a what-if. NO ROW HERE IS B's
+    #   LAG. Sweeping a DIFFERENT second unit through the weighted family
+    #   (2026-09-06 -> TEST-21 re-run):
+    #
+    #     B lag / ramp        original      re-run
+    #     10.0 s / 800 W/s      30.0 s      35.0 s     (= A, both copies)
+    #     20.0 s / 800 W/s      45.0 s      45.0 s
+    #     10.0 s / 400 W/s      30.0 s      35.0 s
+    #     20.0 s / 400 W/s      45.0 s      50.0 s
+    #     40.0 s / 200 W/s      90.0 s      75.0 s     (4x A's lag)
+    #
+    #   The LAG moves it roughly linearly; the RAMP barely does. "The ramp does
+    #   not move it AT ALL" was too strong - halving it is worth 0 s at A's lag
+    #   and +5 s at twice it - so the ramp is second-order, not absent. That is
+    #   the sensitivity the margin has to carry until SPIKE-04 measures the real
+    #   unit.
    #
    #   UPPER, from the bill - 900 s, the quarter that sets the capacity tariff.
    #   A site running short of one unit's share must be reported well inside
    #   one.
    #
-    # 185 s: thirty-seven control cycles. 6.2x the worst legitimate episode this
-    # rig produces, 2.06x the worst a second unit FOUR TIMES slower than A
-    # produces, 6.2x the firmware's own veto window, and 4.9x inside the
-    # quarter. ⚠️ DELIBERATELY NOT 180 - that is SAFETY-02's site hold and this
+    # 185 s: thirty-seven control cycles. Against the TEST-21 re-run: 5.29x the
+    # worst legitimate episode this rig produces (35.0 s), 2.47x the worst a
+    # hypothetical second unit FOUR TIMES slower than A produces (75.0 s), 6.2x
+    # the firmware's own veto window (30 s), and 4.9x inside the quarter
+    # (900 s). ⚠️ DELIBERATELY NOT 180 - that is SAFETY-02's site hold and this
    # is a different check on a different quantity, so a reader must not be
    # able to think it was carried across - and not 150 or 200.
    #
+    # ⚠️ 185 STANDS AND IS NOT CHANGED. The record claimed 6.2x the measured
+    # worst; the re-run makes it 5.29x. All four bounds still clear and none of
+    # them narrowly, and the ordering of the argument is unchanged. A shipped
+    # safety constant is not moved because a margin went from 6.2x to 5.3x while
+    # every bound holds - moving it would re-open SAFETY-41's four-attempt
+    # false-trip regime for nothing. What changed is the CLAIM, not the value.
+    #
    # WHAT IT COSTS, stated rather than hidden: up to 185 s of a site running
    # short by one unit's share before anybody is told. Against a fault that ran
    # 6 h 25 m unnoticed, and against a guard that TAKES NO ACTION, this window
@@ -363,6 +414,94 @@ DELIVERY_OPTION_DEFAULTS = {
 MIN_W_IS_SATURATION_W = "control.Tuning.saturation_w"


+# ---------------------------------------------------------------------------
+# SAFETY-43. THE SECOND WAY TO BE NAMED, AND WHY IT IS A FRACTION OF THE ASK.
+#
+# `min_w` above is an ABSOLUTE gap, and OPT-09 made that a hole. `weights()`
+# shrinks a low-headroom unit's share, and shrinking the share shrinks the gap:
+# a unit standing at +125 W and delivering 0 W is 100 % delinquent and 125 W
+# short, so an absolute 500 W never names it. Measured end to end
+# (`test_packshare.py` part 8b, real `compute()`, real `plan()`, real
+# `delinquent()`, B wedged at a healthy 15 % against its own 10 % cut-out):
+#
+#     load 1500 W, equal weights    the wedged unit named on 199/200 cycles
+#     load 1500 W, OPT-09 weights   named on   0/200 cycles
+#
+# ⚠️ AND THE SUPPRESSED REGIME IS THE QUIET ONE. At a load the healthy sibling
+# covers, the meter is satisfied, `unmet_w` is 0 and no action-log line fires,
+# so this alarm is the ONLY instrument that would have told the operator.
+#
+# ⚠️ THE FIX IS NOT A SMALLER `min_w`, AND THAT IS THE WHOLE SHAPE OF THIS
+# CARD. `min_w` is `Tuning.saturation_w` and it exists to stop NORMAL
+# SATURATION being reported as delinquency - the crying-wolf failure the whole
+# module is written against, and the one SAFETY-41 spent four attempts on
+# ("i'm getting really annoyed by the notifications in my HA"). Lowering it
+# trades a missed alarm for a false one on every unit at once. A fraction of
+# the unit's OWN SHARE is a SECOND way to be named that leaves the first
+# exactly where it was: `gap > min_w OR gap > frac * share`. Nothing that was
+# silent before this card becomes noisy unless the unit is delivering almost
+# none of what it was asked for, which is not saturation-shaped at all.
+#
+# WHERE 0.85 COMES FROM - MEASURED ON THE SAME CAMPAIGN AS `delivery_hold_s`.
+# `admin/goodwe:sim/delivery_window.py --campaign --fractions` grades this
+# constant over the SAME 2640 no-fault runs (three families x 6 gains x 5 slews
+# x the write staggers, every unit executing the share it was given) that
+# produced the 35.0 s worst legitimate episode, at the shipped `min_w`, and
+# reports what the relative arm ADDS. Worst continuous legitimate episode,
+# worst of the three families:
+#
+#   frac   0.05  0.10  0.15  0.20  0.25  0.30  0.35  0.40  0.45  0.50
+#   worst  90.0  90.0  70.0  70.0  60.0  55.0  55.0  60.0  45.0  45.0  s
+#
+#   frac   0.55  0.60  0.65  0.70  0.75  0.80 *0.85  0.90  0.95  0.99
+#   worst  40.0  40.0  45.0  45.0  45.0  45.0  40.0  40.0  40.0  40.0  s
+#
+#   absolute arm alone: 35.0 s - SAFETY-42 §3.3's figure, re-measured in the
+#   SAME runs, which is what makes every row above a comparison and not a claim.
+#   N=1 control: 0.0 s at EVERY fraction, in all 2640 one-inverter runs.
+#   Flapping family: 0.0 s at EVERY fraction, in all 840 runs.
+#
+# ⚠️ READ THE CURVE, NOT THE ROW. THERE IS NO CLIFF ANYWHERE IN IT. From 0.45
+# to 0.99 every figure is 40.0 or 45.0 s - within ONE CONTROL CYCLE (5.0 s) of
+# every other - and below 0.45 it rises gradually to 90.0 s at 0.05. It is not
+# monotone either (0.40 reports 60.0 s where 0.35 reports 55.0), which is
+# §3.3's own mechanism (c) rather than noise in the plant: a more sensitive rule
+# puts MORE units into the short set, which can turn a proper subset into the
+# WHOLE roster and END the episode. So this constant is INSENSITIVE over the
+# whole usable range, and that is the finding rather than a caveat on it. 0.85
+# sits in the flat part, 0.40 above the shoulder at ~0.45 and 0.15 below 1.0.
+#
+# ⚠️ AND "ZERO FALSE NAMINGS" IS NOT AVAILABLE AT ANY FRACTION - said here
+# because the card asked for exactly that. Every fraction adds naming CYCLES to
+# the no-fault campaign; a relative arm that never fired on a legitimate
+# transient would be one that never fired at all, because the modelled write
+# stagger lands B's number up to four cycles late and a unit whose new command
+# has not arrived IS delivering a small fraction of what is standing on its
+# wire. What is zero at every fraction is false ALARMS: the longest legitimate
+# episode anywhere in the 20-point sweep is 90.0 s - at 0.05, the most sensitive
+# fraction measured and far below anything this card would ship - against a
+# `delivery_hold_s` of 185 s, so NOTHING in 2640 no-fault runs comes within 95 s
+# of notifying an operator at any fraction measured. At 0.85 the hold is 4.62x
+# the worst legitimate episode where the absolute arm alone gave 5.29x. That is
+# the price, in full.
+#
+# ⚠️ THE DIRECTION IS THE OPPOSITE OF `min_w`'s AND IT IS EASY TO GET BACKWARDS.
+# A SMALLER fraction names MORE units - it is the more sensitive setting and the
+# more dangerous one - so the campaign puts a LOWER bound on this number, not an
+# upper one. It is bounded ABOVE by the card's own case: at 1.0 the arm can only
+# fire on a unit that is IMPORTING while asked to discharge, which is
+# SAFETY-30's fault and not this one, so a unit delivering NONE of its share
+# (gap/share exactly 1.0) would be missed. The 0.15 below 1.0 is what covers a
+# wedged box reporting a small standby figure rather than a clean zero.
+#
+# ⚠️ AND IT IS NOT AN OPTION, FOR `delivery_w`'s REASON ONE PARAGRAPH UP. An
+# operator has no instrument to derive it with, a knob here is a knob on how
+# loud this add-on is, and SAFETY-41 is the record of what that costs. It is a
+# derived constant with a committed harness behind it, and the way to move it
+# is to re-run the campaign.
+DELIVERY_SHARE_FRACTION = 0.85
+
+
@dataclass(frozen=True)
 class DeliveryConfig:
    """Defaults, not constants - config.yaml carries both. See CFG-03 above."""
@@ -481,12 +620,30 @@ def account(*, live: bool, parked: bool, soc, bms_floor) -> str | None:
    return None


-def delinquent(units, min_w: float) -> dict:
+def delinquent(units, min_w: float,
+               share_frac: float = DELIVERY_SHARE_FRACTION) -> dict:
    """Which units are short with NO ACCOUNT, when the roster disagrees with them.

-    `units`  one `UnitDelivery` per inverter, in roster order.
-    `min_w`  the shortfall that counts - `control.Tuning.saturation_w`. See
-             `MIN_W_IS_SATURATION_W`.
+    `units`       one `UnitDelivery` per inverter, in roster order.
+    `min_w`       the ABSOLUTE shortfall that counts - `control.Tuning.
+                  saturation_w`. See `MIN_W_IS_SATURATION_W`.
+    `share_frac`  SAFETY-43. The RELATIVE shortfall that counts, as a fraction
+                  of that unit's own share. See `DELIVERY_SHARE_FRACTION`; it
+                  is a parameter rather than a read of the constant so the
+                  campaign that derived it can sweep it over the shipped code.
+                  ⚠️ `None` MEANS THE ABSOLUTE ARM ALONE - SAFETY-42's rule,
+                  exactly - and it exists for ONE caller: the campaign in
+                  `admin/goodwe:sim/delivery_window.py`, which has to be able
+                  to measure what this card ADDS against what it replaced, run
+                  for run, over ONE tree. Nothing in the add-on passes it, and
+                  `main.py` passes nothing at all.
+
+    ⚠️ THE TWO ARMS ARE AN `or` AND THE ORDER OF THAT SENTENCE MATTERS. The
+    absolute arm is untouched by SAFETY-43: everything `min_w` named before it
+    is still named, and nothing `min_w` refused becomes noisy unless the unit
+    is delivering almost NONE of its ask. Lowering `min_w` would have been the
+    other change - the one that reports normal saturation on every unit at once
+    - and `DELIVERY_SHARE_FRACTION` says at length why it was refused.

    Returns `{name: sentence}`, empty when there is nothing to say. The sentence
    is the operator's first line and it carries the peer clause, because "B is
@@ -528,7 +685,13 @@ def delinquent(units, min_w: float) -> dict:
    gaps = {}
    for u in asked:
        gap = short_by(u.share_w, u.measured_w)
-        if gap is not None and gap > float(min_w):
+        if gap is None:
+            continue
+        # SAFETY-43: absolute OR relative. `share_w` is > 0 for every member
+        # of `asked`, so the second comparison is always well formed.
+        if gap > float(min_w) or (share_frac is not None
+                                  and gap > float(share_frac)
+                                  * float(u.share_w)):
            gaps[u.name] = gap

    unaccounted = {u.name: gaps[u.name] for u in asked
@@ -544,7 +707,13 @@ def delinquent(units, min_w: float) -> dict:
        out[u.name] = (
            f"{REASON_CODE}: {u.name} is standing at {float(u.share_w):+.0f} W "
            f"and delivering {float(u.measured_w):+.0f} W - "
-            f"{unaccounted[u.name]:.0f} W of its share is not arriving - while "
+            f"{unaccounted[u.name]:.0f} W of its share "
+            # SAFETY-43. The PROPORTION is in the sentence because the watts
+            # alone stopped being the whole story the moment a unit could be
+            # named on the relative arm: "125 W is not arriving" reads like a
+            # rounding error, and "125 W (100 % of its share)" is the fault.
+            f"({100.0 * unaccounted[u.name] / float(u.share_w):.0f} % of it) "
+            f"is not arriving - while "
            f"{', '.join(peers)} {'is' if len(peers) == 1 else 'are'} not in "
            f"that state. Nothing this add-on reads explains it")
    return out
@@ -57,8 +57,8 @@ deliver DOES NOT SHIFT TO ITS SIBLING by any mechanism in this add-on.
    never rises. That module DETECTS it; this one cannot fix it, and saying so
    is worth more than a weighting scheme that looks like it might.

-⚠️ AND IT IS WORSE THAN "CANNOT FIX IT": THESE WEIGHTS NARROW SAFETY-42's
-DETECTION WINDOW, WHICH IS A DETECTION REGRESSION AT N=2
+⚠️ AND IT WAS WORSE THAN "CANNOT FIX IT": THESE WEIGHTS NARROWED SAFETY-42's
+DETECTION WINDOW - A DETECTION REGRESSION AT N=2, NOW CLOSED BY SAFETY-43
 =======================================================================
 This module used to say - and `DECISIONS/OPT-09-soc-aware-weights.md` §5 said
 with it - that a wedged unit's weight does not move, so OPT-09 cannot see the
@@ -68,27 +68,40 @@ pack's ENERGY HEADROOM, not of what the pack delivers, so a wedged unit's weight
 moves whenever its headroom differs from its sibling's - the ordinary case this
 module exists to exploit.

-`delivery.delinquent()` names a unit only while its `short_by()` gap exceeds
+`delivery.delinquent()` named a unit only while its `short_by()` gap exceeded
 `min_w` (`Tuning.saturation_w`, 500 W) in ABSOLUTE WATTS. Shrinking a unit's
 share shrinks its gap. Measured end to end (`test_packshare.py` part 8b: real
 `compute()`, real `plan()`, real `delinquent()`, B wedged at 15 % against its
-own 10 % floor, so `envelope()` does not close and `account()` explains
-nothing): at a 1500 W load the healthy pack can carry alone, equal weights name
-the wedged unit on 199 of 200 cycles and these weights name it on NONE. The
-suppression window widens from `site target < 2 * min_w` = 1000 W to
-`site target < min_w * (w_A + w_B) / w_B` = 6500 W on that fixture - and it has
+own 10 % cut-out, so `envelope()` does not close and `account()` explains
+nothing): at a 1500 W load the healthy pack can carry alone, equal weights named
+the wedged unit on 199 of 200 cycles and these weights named it on NONE. The
+suppression window widened from `site target < 2 * min_w` = 1000 W to
+`site target < min_w * (w_A + w_B) / w_B` = 6500 W on that fixture - and it had
 an upper edge: past the healthy pack's own rating, `_fill()`'s last-resort pass
-hands the wedged unit a big share again and the alarm returns.
+hands the wedged unit a big share again and the absolute arm names it.

-    ⚠️ BOTH BOUNDS ARE ARITHMETIC ABOUT `_fill()`'s PROPORTIONAL REGIME AND NEITHER IS A REACHABLE STEADY STATE. Swept on this fixture, the healthy pack's own 5000 W rating binds FIRST: suppressed 0/200 up to a 3000 W load, 4/200 at 4000 W, 41/200 at 4900 W (site target 5320 W, the largest reached with the wedged unit still under 500 W), then 195/200 at 5000 W once `_fill()` pins A at its cap and hands B 3000 W. So the alarm returns BELOW the 6500 W formula bound, not above it. The equal-weights 1000 W bound is unreachable for the same reason - the equal arm names the wedged unit 199/200 at every load swept, 1000 W included.
+    ⚠️ BOTH BOUNDS ARE ARITHMETIC ABOUT `_fill()`'s PROPORTIONAL REGIME AND NEITHER IS A REACHABLE STEADY STATE - KEPT HERE BECAUSE IT IS A MEASUREMENT, EVEN THOUGH SAFETY-43 HAS SINCE MADE THE WINDOW IRRELEVANT TO THE DETECTOR. Swept on this fixture with the ABSOLUTE arm alone, the healthy pack's own 5000 W rating binds FIRST: suppressed 0/200 up to a 3000 W load, 4/200 at 4000 W, 41/200 at 4900 W (site target 5320 W, the largest reached with the wedged unit still under 500 W), then 195/200 at 5000 W once `_fill()` pins A at its cap and hands B 3000 W. So the alarm returned BELOW the 6500 W formula bound, not above it. The equal-weights 1000 W bound is unreachable for the same reason - the equal arm named the wedged unit 199/200 at every load swept, 1000 W included.

-    ⚠️ SO THE SEAM IS NOT "OPT-09 CANNOT SEE IT". IT IS: OPT-09 STOPS THE ASK
-    **AND** NARROWS SAFETY-42's DETECTION WINDOW FOR A LOW-HEADROOM UNIT. That
-    is a disclosed residual, not a fixed thing. Closing it means making
-    `delinquent()` relative to the ask rather than absolute - SAFETY-42's code
-    and a design change, filed as its own card and deliberately NOT done here.
-    Do not "fix" part 8b by making it assert the alarm still fires; it does not,
-    and a green check that says otherwise would be worse than the gap.
+    ⚠️ THAT IS NO LONGER A RESIDUAL. SAFETY-43 made `delinquent()` relative to
+    the ask AS WELL AS absolute - `gap > min_w OR gap > frac * share`, and a
+    unit delivering 0 of 125 W is as delinquent as one delivering 0 of 1500 W.
+    On the SAME fixture, under the SAME weights, the wedged unit is now named
+    on 200 of 200 cycles. Part 8b grades both rules on one run:
+    `share_frac=None` is SAFETY-42's absolute arm alone and still says nothing,
+    which is what keeps the paragraph above honest rather than nostalgic.
+
+    ⚠️ AND `min_w` WAS NOT LOWERED. The absolute arm's suppression window is
+    still exactly the arithmetic above - 1000 W at equal weights, 6500 W at
+    these - because a lower `saturation_w` would report NORMAL SATURATION as
+    delinquency on every unit at once, which is the false-trip regime SAFETY-41
+    spent four attempts in. What changed is that there is a SECOND way to be
+    named; nothing that was quiet before became noisy.
+
+    ⚠️ WHAT THIS MODULE STILL CANNOT DO IS UNCHANGED. OPT-09 stops the site
+    ASKING a pack whose own SoC says it cannot deliver; it does not, and will
+    not, re-allocate a wedged unit's watts to its sibling - `delivery.py`'s
+    "IT REPORTS. IT DOES NOT ACT" block is the measurement behind that, and
+    SAFETY-43 did not touch it. The alarm is louder; the loop is identical.

 WHY ENERGY AND NOT POWER
 ========================
@@ -548,8 +548,10 @@ options:
  # only ever ran at 500 could not tell this wiring from a hard-coded number.
  #
  # How long ONE inverter may stand at a discharge share it is not producing -
-  # short by more than `saturation_w` - while at least one sibling is NOT in
-  # that state, before the add-on names it. It REPORTS: nothing is commanded
+  # short by more than `saturation_w`, OR short by more than a fixed FRACTION
+  # of its own share (SAFETY-43; the fraction is derived and is not an option,
+  # see app/delivery.py) - while at least one sibling is NOT in that state,
+  # before the add-on names it. It REPORTS: nothing is commanded
  # differently, the loop is untouched, and the site goes on running short by
  # that unit's share until somebody looks. app/delivery.py carries why, with
  # the rig measurements on both sides of the decision.
@@ -560,17 +562,31 @@ options:
  # ⚠️ MEASURED ON THE TEST-19 TWO-MASTER RIG, NOT CHOSEN. 2640 closed-loop runs
  # at a roster of two in which nothing was faulty, over 6 gains x 5 slews - the
  # whole range this schema accepts - across load reversals, a flapping load and
-  # ASYMMETRIC SPLITS, plus 1890 runs at a roster of ONE as the control. At the
-  # shipped threshold the worst CONTINUOUS legitimate episode is 30.0 s (the
-  # weighted family; 20.0 s on reversals, 0.0 s on flapping, and 0.0 s at every
-  # threshold in all 1890 one-inverter runs). 185 s is 6.2x that.
+  # ASYMMETRIC SPLITS, plus a roster-of-ONE arm as the control. At the shipped
+  # threshold the worst CONTINUOUS legitimate episode is 35.0 s (the weighted
+  # family; 30.0 s on reversals, 0.0 s on flapping, and 0.0 s at every threshold
+  # in every one-inverter run). 185 s is 5.29x that.
  #
-  # ⚠️ AND THE MARGIN IS SIZED ON A MEASUREMENT, NOT ON A FEELING, BECAUSE
-  # INVERTER B DOES NOT PHYSICALLY EXIST. The rig's B is A copied. Sweeping a
-  # DIFFERENT second unit through that campaign: the worst legitimate episode is
-  # insensitive to B's ramp (30.0 s at half A's ramp) and scales with its LAG -
-  # 45.0 s at twice A's measured 10 s lag, 90.0 s at four times it. 185 s is
-  # 2.06x even the four-times-slower case.
+  # ⚠️ AND THE CAMPAIGN IS RE-RUNNABLE NOW, WHICH IT WAS NOT WHEN THIS NUMBER
+  # SHIPPED. TEST-21 (2026-09-07) committed the harness to the OTHER repository,
+  # `admin/goodwe:sim/delivery_window.py` - it imports plantsim and frame_check,
+  # which that repo owns. From clean checkouts of both, side by side:
+  #
+  #     cd goodwe/sim && GOODWE_ADDON_REPO=/path/to/goodwe-addon \
+  #         python3 -B delivery_window.py --selftest --campaign --blag
+  #
+  # ⚠️ THE ORIGINAL RECORD SAID 30.0 s AND IT DID NOT REPRODUCE - the re-run
+  # says 35.0 s. 185 s still clears all four bounds and is unchanged; what
+  # changed is the claim. DECISIONS/SAFETY-42-unit-delivery.md §3.3 carries both
+  # tables and what did and did not reproduce.
+  #
+  # ⚠️ AND THE MARGIN IS A SENSITIVITY STUDY ON A HYPOTHESIS, NOT A MEASUREMENT,
+  # BECAUSE INVERTER B DOES NOT PHYSICALLY EXIST (~2026-09-12). The rig's B is A
+  # COPIED. Sweeping a DIFFERENT second unit through that campaign: the worst
+  # legitimate episode barely moves with B's ramp (35.0 s at half A's ramp) and
+  # scales with its LAG - 45.0 s at twice A's measured 10 s lag, 75.0 s at four
+  # times it. 185 s is 2.47x even the four-times-slower case. NO NUMBER HERE IS
+  # A MEASUREMENT OF INVERTER B.
  #
  # DELIBERATELY NOT 180 - that is SAFETY-02's site hold and this is not it -
  # and not 150 or 200. 37 control cycles.
@@ -0,0 +1,372 @@
+"""UX-04 - the defaults pin UX-01 AC 4 promised and never built.
+`python3 -B test_defaults_pin.py`
+
+WHAT UX-01 AC 4 ASKED FOR, IN ITS OWN WORDS: "every `schema:` key absent from
+`options:` has exactly one defaults-table entry ... the table changed goes red
+... and NO MODULE CARRIES ITS OWN LITERAL FALLBACK FOR A SCHEMA DEFAULT - an AST
+check, not a grep."
+
+WHAT SHIPPED. Clause one is VACUOUS: `options:` and `schema:` declare the same
+71 keys, so "every schema key absent from options" quantifies over the empty set
+and is true of any tree, including one with no defaults at all. Clause two -
+"the table changed goes red" - does work, six times over, once per mirror in its
+own suite. Clause three, THE ONE WITH THE CONTENT, was never written. The only
+literal scan in this repository before this file is
+`test_validate.py`'s `re.findall(r'"(maintenance_[a-z_]+)"...')`, which is a
+grep, is restricted to eight `maintenance_*` keys and three named files, and
+predates UX-01.
+
+⚠️ AND A CHECK THAT CANNOT GO RED IS NOT A CHECK, so this file does not repeat
+clause one's mistake. Every part below states the size of the set it quantifies
+over and grades that size against a floor derived from the tree, and part 5
+plants a literal in a synthetic module and REQUIRES the scanner to catch it, so
+the red direction is exercised on every run rather than attested in a commit
+message. MUTATION.md habit 4.
+
+--------------------------------------------------------------------------- #
+WHAT AN "IN-CODE LITERAL FALLBACK" IS, STRUCTURALLY, so a reader can argue with
+the definition rather than guess it. It is a call in `app/` that takes a STRING
+LITERAL naming a `schema:` key and, in the next positional slot, a CONSTANT:
+
+    opts.get("max_w", 2000)          _number(opts, "max_w", 2000.0)
+
+Two shapes, one rule, found by walking `ast.Call` nodes - not by matching text,
+so `opts.get(\n    "max_w",\n    2000,\n)` and a renamed `opts` are found too and
+a `max_w` inside a comment or a docstring is not.
+
+TWO EXCLUSIONS, both structural and both narrow:
+
+  * `getattr(obj, "max_w", 0.0)` reads an ATTRIBUTE off an object, not an option
+    off the operator's blob. Its third argument is an absent-attribute guard and
+    has no relationship to config.yaml. Excluded by function name.
+
+  * THE EMPTY STRING. `opts.get("static_profile", "")` is Python's "this key is
+    not stated", and every one of the six sites in `validate.py` immediately
+    normalises it - `str(opts.get(k, "") or "").strip()`. It cannot be a stale
+    copy of a NUMBER, which is the whole of the CFG-04 defect class, and
+    rewriting those six to index a mirror would replace a sentence that reads
+    "absent" with one that reads "the shipped default" at sites where absent is
+    what is meant. Excluded by value. `0` is NOT excluded: `slew_w: 0` and
+    `integrator_max_w: 0` are real shipped defaults with real meanings.
+
+--------------------------------------------------------------------------- #
+WHAT THIS FILE DOES *NOT* DO, said before the parts rather than after them.
+
+It does not remove the 27 (module, key) pairs in part 4's ledger. That is
+CFG-03's open ticket - option defaults living in four places - and it is a body
+of behaviour change across `main.py`, `p1.py`, `inverter.py` and `validate.py`
+that no single card should make in passing. What this file does instead is
+FREEZE the ledger: the set is graded EXACTLY, so the twenty-eighth pair is red
+on the commit that introduces it and the ledger can only ever shrink. A backlog
+that cannot grow is a backlog being paid down; one that is merely written in a
+comment is the state UX-01 left.
+
+MUTATION.md is the method for the red-direction evidence in part 5, and
+`tools/mutation_run.sh` is the harness.
+"""
+
+import ast
+import io
+import re
+import sys
+from pathlib import Path
+
+HERE = Path(__file__).resolve().parent
+APP = HERE / "app"
+
+FAILED = 0
+
+
+def check(what, ok):
+    global FAILED
+    print(("  ok   " if ok else "  FAIL ") + what)
+    if not ok:
+        FAILED += 1
+
+
+# --------------------------------------------------------------------------- #
+# config.yaml, parsed by hand.
+#
+# ⚠️ NO PyYAML, AND NOT AS A PREFERENCE. The add-on image installs py3-aiohttp
+# and py3-paho-mqtt only; `import yaml` raises in the checks container and a
+# check that skips on ImportError is not a check. `test_maintenance.py` and
+# `test_price.py` each carry their own two-line reader for the same reason, and
+# this is the third - which is itself an instance of the defect this file
+# grades, one level up. It is written down rather than tidied away because
+# tidying it into a shared helper is a change to two suites that are green.
+CFG_TEXT = io.open(HERE / "config.yaml", encoding="utf-8").read()
+_SCHEMA_AT = CFG_TEXT.index("\nschema:")
+_OPTS_BLOCK, _SCHEMA_BLOCK = CFG_TEXT[:_SCHEMA_AT], CFG_TEXT[_SCHEMA_AT:]
+
+
+def _scalar(raw):
+    """One `options:` value, as the yaml module this image lacks would read it."""
+    raw = raw.split("#", 1)[0].strip().strip('"').strip("'")
+    if raw in ("true", "false"):
+        return raw == "true"
+    if raw == "[]":
+        return []
+    for cast in (int, float):
+        try:
+            return cast(raw)
+        except ValueError:
+            pass
+    return raw
+
+
+def _block_keys(block):
+    return [m.group(1) for m in re.finditer(r"^  ([a-z0-9_]+):", block, re.M)]
+
+
+SHIPPED = {m.group(1): _scalar(m.group(2))
+           for m in re.finditer(r"^  ([a-z0-9_]+): *(.*)$", _OPTS_BLOCK, re.M)}
+SCHEMA = set(_block_keys(_SCHEMA_BLOCK))
+
+print("UX-04 - the defaults pin\n")
+print("1. the surface this file quantifies over")
+print(f"    config.yaml: {len(SHIPPED)} keys under `options:`, {len(SCHEMA)} "
+      f"under `schema:`")
+# The floor, for part 1's own reason: a reader that returned {} would make every
+# check below vacuously true. 60 is the count of REQUIRED schema keys and is the
+# number `tools/supervisor_shape_probe.py` section F now grades; any floor well
+# inside the real size does the job, and this one cannot be reached by accident.
+check(f"the hand reader found a real options block ({len(SHIPPED)} keys, floor 60)",
+      len(SHIPPED) >= 60)
+check(f"...and a real schema block ({len(SCHEMA)} keys, floor 60)",
+      len(SCHEMA) >= 60)
+check("every shipped default names a schema key, so a mirror pinned to one is "
+      "pinned to the other",
+      sorted(set(SHIPPED) - SCHEMA) == [])
+
+
+# --------------------------------------------------------------------------- #
+print()
+print("2. the mirrors, DISCOVERED from the tree rather than listed here")
+# ⚠️ WHY DISCOVERED. Six suites already pin their own mirror to config.yaml -
+# test_maintenance, test_price, test_watchdog (twice), test_peakguard,
+# test_delivery. Every one of those pins names its table by hand, so a SEVENTH
+# table added tomorrow is pinned by nobody and nothing says so. This walks
+# `app/` for module-level assignments whose name ends in OPTION_DEFAULTS and
+# grades whatever it finds, so the seventh is covered on the commit that
+# introduces it. That is the property the six hand-written pins cannot have.
+MODULES = sorted(p.name for p in APP.glob("*.py"))
+TREES = {name: ast.parse(io.open(APP / name, encoding="utf-8").read(), name)
+         for name in MODULES}
+
+MIRROR = {}          # schema key -> (module, table name, value)
+TABLES = []
+for name in MODULES:
+    for node in TREES[name].body:
+        if not (isinstance(node, ast.Assign) and len(node.targets) == 1
+                and isinstance(node.targets[0], ast.Name)
+                and node.targets[0].id.endswith("OPTION_DEFAULTS")):
+            continue
+        table = node.targets[0].id
+        TABLES.append((name, table))
+        for k, v in zip(node.value.keys, node.value.values):
+            # A literal key with a literal value is the only shape a mirror
+            # has; anything else is not a table this file can grade, and
+            # saying so out loud beats skipping it silently.
+            MIRROR[k.value] = (name, table, ast.literal_eval(v))
+
+print(f"    {len(TABLES)} table(s): " + ", ".join(f"{m}:{t}" for m, t in TABLES))
+print(f"    {len(MIRROR)} mirrored key(s) of {len(SCHEMA)} in `schema:`")
+check(f"the walk found the mirrors that exist ({len(TABLES)} tables, floor 6)",
+      len(TABLES) >= 6)
+_unschemad = sorted(k for k in MIRROR if k not in SCHEMA)
+check(f"every mirrored key is a real schema key {_unschemad or ''}",
+      _unschemad == [])
+_unshipped = sorted(k for k in MIRROR if k not in SHIPPED)
+check(f"...and is shipped in the options block {_unshipped or ''}",
+      _unshipped == [])
+_drift = sorted(k for k, (m, t, v) in MIRROR.items()
+                if k in SHIPPED and v != SHIPPED[k])
+check(f"...and reproduces config.yaml exactly, so ANY mirror - including one no "
+      f"suite names by hand - goes red on drift {_drift or ''}",
+      _drift == [])
+
+
+# --------------------------------------------------------------------------- #
+print()
+print("3. the AST scan: no module carries its own literal fallback for a "
+      "MIRRORED schema default")
+
+
+def literal_fallbacks(tree, name):
+    """Every `f(..., "<schema key>", <constant>, ...)` in one module.
+
+    Yields (module, lineno, callee, key, value). See the header for the two
+    exclusions and the argument for each.
+    """
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Call):
+            continue
+        fn = node.func
+        callee = fn.attr if isinstance(fn, ast.Attribute) else getattr(fn, "id", "")
+        if callee == "getattr":
+            continue
+        for i, arg in enumerate(node.args[:-1]):
+            if not (isinstance(arg, ast.Constant) and isinstance(arg.value, str)
+                    and arg.value in SCHEMA):
+                continue
+            default = node.args[i + 1]
+            if not isinstance(default, ast.Constant):
+                continue
+            if default.value is None or default.value == "":
+                continue
+            yield (name, node.lineno, callee, arg.value, default.value)
+
+
+ALL_SITES = [s for name in MODULES for s in literal_fallbacks(TREES[name], name)]
+MIRRORED_SITES = [s for s in ALL_SITES if s[3] in MIRROR]
+UNMIRRORED_SITES = [s for s in ALL_SITES if s[3] not in MIRROR]
+
+print(f"    scanned {len(MODULES)} modules, {len(ALL_SITES)} literal fallback "
+      f"site(s): {len(MIRRORED_SITES)} over mirrored keys, "
+      f"{len(UNMIRRORED_SITES)} over unmirrored ones")
+# ⚠️ THE SCANNER'S OWN FLOOR, and it is the line that stops this whole file
+# rotting into clause one. A scanner that silently stopped matching - a renamed
+# helper, an ast API change, a `SCHEMA` that came back empty - would report ZERO
+# sites and every check below would pass. So the total is graded against a floor
+# the tree already exceeds by a wide margin, and part 5 additionally feeds the
+# scanner a source it MUST flag.
+check(f"the scanner is not blind: it found {len(ALL_SITES)} sites (floor 20)",
+      len(ALL_SITES) >= 20)
+for site in sorted(MIRRORED_SITES):
+    print(f"      {site[0]}:{site[1]}  {site[2]}({site[3]!r}, {site[4]!r})")
+check(f"no literal fallback survives for any of the {len(MIRROR)} keys a mirror "
+      f"already answers for - those sites must index the mirror",
+      MIRRORED_SITES == [])
+
+
+# --------------------------------------------------------------------------- #
+print()
+print("4. the CFG-03 ledger: the unmirrored keys that still carry a literal, "
+      "frozen at exactly today's set")
+# ⚠️ WHY A FROZEN SET AND NOT A COUNT. A count goes green again the moment
+# somebody deletes one site and adds another, which is the same-size drift this
+# repository keeps finding. The pair (module, key) is stable under every edit
+# that is not the one being guarded against: reformatting, moving the call
+# within the file, renaming the local `opts`. It changes only when a module
+# starts or stops carrying a literal for a key - which is exactly the event.
+#
+# LINE NUMBERS ARE DELIBERATELY NOT IN THE KEY. A ledger that goes red when an
+# unrelated edit shifts a line is a ledger people learn to re-bless without
+# reading, and a ledger re-blessed without reading grades nothing.
+LEDGER = {
+    ("inverter.py", "max_w"),
+    ("main.py", "auto_start"),
+    ("main.py", "cycle_floor_s"),
+    ("main.py", "deadband_w"),
+    ("main.py", "estop_fitted"),
+    ("main.py", "gain"),
+    ("main.py", "heartbeat_s"),
+    ("main.py", "log_level"),
+    ("main.py", "meter_freeze_s"),
+    ("main.py", "meter_max_age_s"),
+    ("main.py", "meter_phases"),
+    ("main.py", "peak_cap_w"),
+    ("main.py", "saturation_cycles"),
+    ("main.py", "saturation_w"),
+    ("main.py", "sign_disagree_s"),
+    ("main.py", "sign_disagree_w"),
+    ("main.py", "slew_w"),
+    ("main.py", "stale_input_s"),
+    ("main.py", "step_w"),
+    ("main.py", "target_grid_w"),
+    ("p1.py", "meter_poll_s"),
+    ("p1.py", "meter_source"),
+    ("validate.py", "max_w"),
+    ("validate.py", "meter_max_age_s"),
+    ("validate.py", "meter_phases"),
+    ("validate.py", "sign_disagree_w"),
+    ("validate.py", "slew_w"),
+}
+_found = {(s[0], s[3]) for s in UNMIRRORED_SITES}
+_new = sorted(_found - LEDGER)
+_gone = sorted(LEDGER - _found)
+print(f"    {len(_found)} (module, key) pair(s) carry a literal for a key no "
+      f"mirror answers for")
+check(f"no NEW module/key pair has started carrying a literal fallback {_new or ''}",
+      _new == [])
+check(f"...and none has been paid off without the ledger being shortened - "
+      f"shrink LEDGER when it has {_gone or ''}",
+      _gone == [])
+
+# --- and every one of them is graded AGAINST config.yaml, which is the half
+#     that catches CFG-04's actual failure: two copies of one number, one moves.
+DIVERGENT = {
+    # main.py:290 `target_grid_w=float(opts.get("target_grid_w", 0))`.
+    # config.yaml ships -10 W. MEASURED here for the first time, 2026-09-07,
+    # and it is a real divergence, not a sentinel: a dict handed to
+    # `Controller` WITHOUT the key holds net exchange at 0 W instead of the
+    # shipped -10 W bias into import. On a Supervisor install the key is always
+    # present - the deep-merge of `options:` guarantees it - so this cannot
+    # reach an operator; the paths that CAN hand over a partial dict are
+    # `sim/inject.py` and this repository's own fixtures. It is recorded rather
+    # than corrected because changing it is a behaviour change to the control
+    # loop's target, which belongs to a card that grades the loop.
+    ("main.py", "target_grid_w"): (0, -10),
+}
+_diverge = {}
+for mod, line, callee, key, value in UNMIRRORED_SITES:
+    want = SHIPPED.get(key)
+    same = (isinstance(value, bool) == isinstance(want, bool)) and value == want
+    if not same:
+        _diverge[(mod, key)] = (value, want)
+for pair, (got, want) in sorted(_diverge.items()):
+    print(f"      {pair[0]}  {pair[1]}: code {got!r} vs config.yaml {want!r}")
+check(f"every remaining literal still AGREES with config.yaml, except the "
+      f"{len(DIVERGENT)} divergence(s) recorded above it",
+      _diverge == DIVERGENT)
+
+
+# --------------------------------------------------------------------------- #
+print()
+print("5. the red direction, RUN on every invocation")
+# ⚠️ A CHECK THAT HAS NEVER BEEN SHOWN FAILING IS NOT A CHECK. Parts 3 and 4
+# both pass on a clean tree, which is exactly what UX-01 AC 4's clause one did
+# while grading nothing. So the scanner is fed a source that is a deliberate
+# violation of each rule and is REQUIRED to flag it - in-process, against the
+# same `literal_fallbacks()` the parts above call, so a scanner that stops
+# working stops working here too.
+_PLANTED = '''
+def build(opts):
+    a = opts.get("maintenance_charge_w", 2000)     # mirrored key -> part 3
+    b = _number(opts, "cycle_floor_s", 15)         # unmirrored, main.py's key
+    c = opts.get("static_profile", "")             # the "" sentinel: NOT a hit
+    d = getattr(unit, "max_w", 0.0)                # attribute guard: NOT a hit
+    e = opts.get("peak_soc_floor", 20)             # mirrored key -> part 3
+    return a, b, c, d, e
+'''
+_hits = list(literal_fallbacks(ast.parse(_PLANTED, "<planted>"), "<planted>"))
+_keys = sorted(h[3] for h in _hits)
+for h in _hits:
+    print(f"      caught  {h[2]}({h[3]!r}, {h[4]!r})  at line {h[1]}")
+check("a planted literal for a MIRRORED key is caught - this is part 3's red "
+      "direction, exercised",
+      "maintenance_charge_w" in _keys and "peak_soc_floor" in _keys)
+check("a planted literal for an unmirrored key is caught too - part 4's",
+      "cycle_floor_s" in _keys)
+check('...and the two documented exclusions still hold: the "" sentinel and '
+      "getattr are not reported",
+      "static_profile" not in _keys and _keys.count("max_w") == 0)
+check(f"exactly the three intended sites, no more {_keys}",
+      _keys == ["cycle_floor_s", "maintenance_charge_w", "peak_soc_floor"])
+
+# ...and the mirror comparison's red direction, likewise executed rather than
+# claimed: part 2 goes red when a table and config.yaml disagree by one.
+_drifted = dict(MIRROR)
+_pick = sorted(MIRROR)[0]
+_mod, _tab, _val = _drifted[_pick]
+_drifted[_pick] = (_mod, _tab, "###drifted###")
+check(f"part 2's comparison rejects a mirror that has drifted (simulated on "
+      f"{_pick!r})",
+      any(v != SHIPPED.get(k) for k, (_, _, v) in _drifted.items()))
+
+
+print()
+if FAILED:
+    print(f"{FAILED} FAILED")
+    sys.exit(1)
+print("all checks passed")
@@ -33,8 +33,9 @@ from datetime import datetime, timedelta, timezone

 from app.control import Tuning
 from app.delivery import (
-    DELIVERY_OPTION_DEFAULTS, REASON_CODE, DeliveryConfig, DeliveryWatchdog,
-    UnitDelivery, account, delinquent, short_by,
+    DELIVERY_OPTION_DEFAULTS, DELIVERY_SHARE_FRACTION, REASON_CODE,
+    DeliveryConfig, DeliveryWatchdog, UnitDelivery, account, delinquent,
+    short_by,
 )
 from app.inverter import LINK_STALE_S
 from app.main import Controller
@@ -151,6 +152,77 @@ check("a unit given no share is not asked and cannot be named, and it is not "
      "a peer either - with only one unit asked there is nothing to compare",
      delinquent([U("a", 0, 0), U("b", 2000, 0)], 500) == {})

+print()
+print("   SAFETY-43: THE RELATIVE ARM - short OF ITS OWN ASK, not just short")
+
+# ⚠️ `share_frac=None` IS SAFETY-42's RULE EXACTLY, and every check in this part
+# above this line is run under BOTH rules below to show the card took nothing
+# away. What it ADDS is one thing: a unit can now be named for delivering almost
+# none of a SMALL share. OPT-09's weights make small shares ordinary - a
+# low-headroom unit is weighted down to 125 W of a 1630 W site target - so
+# before this card a wedged unit could be 100 % delinquent and 125 W short, and
+# 125 W is under a 500 W threshold. test_packshare.py part 8b measures that end
+# to end on the real loop; what is graded here is the predicate.
+_ABS = None                                    # SAFETY-42's rule
+
+_tiny = [U("a", 1505, 1505), U("b", 125, 0)]
+check("A UNIT DELIVERING 0 OF 125 W IS NAMED, AND UNDER THE OLD RULE IT WAS "
+      "NOT. Same three units, same `min_w`, one input changed: the absolute arm "
+      "alone answers {} because 125 W < 500 W, and the shipped rule names it "
+      "because 100 % of its ask is not arriving",
+      delinquent(_tiny, 500, _ABS) == {}
+      and list(delinquent(_tiny, 500)) == ["b"])
+# ⚠️ `.get`, NOT `["b"]`, AND THAT IS NOT STYLE. A bare subscript here raises
+# KeyError on any tree where the relative arm is broken or absent, which ABORTS
+# THE FILE at part 1 and leaves parts 2-8 UNGRADED. The direction is safe - an
+# abort is a red, and MUTATION.md's defect can only fabricate a green - but the
+# recorded coverage of a mutation run then reads larger than it was, which is
+# the shape M11's entry in test_packshare.py part 11 already discloses. Found
+# in review of this card, on M12 and M13.
+_tiny_said = delinquent(_tiny, 500).get("b", "(not named - see the check above)")
+print(f"       {_tiny_said}")
+check("...and the sentence carries the PROPORTION as well as the watts, because "
+      "'125 W is not arriving' reads like a rounding error and '125 W (100 % of "
+      "it)' is the fault",
+      "125 W of its share (100 % of it) is not arriving" in _tiny_said)
+
+check("THE ACCOUNT GATE IS UNTOUCHED BY IT - the same 0 of 125 W with the "
+      "shortfall EXPLAINED names nobody. A relative criterion that could talk "
+      "its way past `account()` would report every pack sitting at its own "
+      "cut-out on a small share, which is the crying-wolf failure this module "
+      "exists to avoid",
+      delinquent([U("a", 1505, 1505), U("b", 125, 0, "its pack is empty")],
+                 500) == {})
+check("...and so is the PROPER-SUBSET rule: both units delivering none of a "
+      "small share is a SITE condition and is still nobody's fault to name",
+      delinquent([U("a", 125, 0), U("b", 125, 0)], 500) == {})
+check("...and the relative bound is STRICT, like the absolute one. A gap of "
+      "EXACTLY the fraction of the share is not over it; one watt more is",
+      delinquent([U("a", 2000, 2000),
+                  U("b", 1000, 1000 * (1.0 - DELIVERY_SHARE_FRACTION))],
+                 5000) == {}
+      and list(delinquent([U("a", 2000, 2000),
+                           U("b", 1000,
+                             1000 * (1.0 - DELIVERY_SHARE_FRACTION) - 1)],
+                          5000)) == ["b"])
+
+_both_rules = [(delinquent(u, w, _ABS), delinquent(u, w)) for u, w in (
+    ([U("a", 2000, 2000), U("b", 2000, 0)], 500),
+    ([U("a", 2000, 2000), U("b", 2000, 0, "its pack is empty")], 500),
+    ([U("a", 2000, 0), U("b", 2000, 0)], 500),
+    ([U("a", 2000, 2000), U("b", 2000, 0), U("c", 2000, 0)], 500),
+    ([U("a", 2000, 0, "at its floor"), U("b", 2000, 0)], 500),
+    ([U("a", 2000, 2000), U("b", 2000, 1500)], 500),
+    ([U("a", 0, 0), U("b", 2000, 0)], 500),
+)]
+check(f"⚠️ AND IT TOOK NOTHING AWAY: all {len(_both_rules)} fixtures above this "
+      f"block answer IDENTICALLY under SAFETY-42's rule and the shipped one. "
+      f"The relative arm is an `or`, so it can only ever ADD a naming - and on "
+      f"every case this part already graded it adds none. That is the "
+      f"difference between adding a second way to be named and lowering "
+      f"`min_w`, which would have moved all seven",
+      all(a == b for a, b in _both_rules))
+
 print()
 print("   AC 5: AT ONE INVERTER IT IS INERT, AND BY ALGEBRA RATHER THAN BY A")
 print("   ROSTER GUARD - `len(asked) < 2` is unreachable-by-construction there")
@@ -323,9 +395,14 @@ class _DeliveryHass:
    same in both.
    """

-    def __init__(self, load_w=_LOAD_W, b_soc=50.0):
+    def __init__(self, load_w=_LOAD_W, b_soc=50.0, b_frac=0.0):
        self.load_w = float(load_w)
        self.b_soc = float(b_soc)
+        # SAFETY-43. What fraction of its command B actually delivers. 0.0 is a
+        # WEDGED unit and is every arm in this file but part 4b(a) - see there
+        # for why that one arm needs a unit that is short WITHOUT being 100 %
+        # short.
+        self.b_frac = float(b_frac)
        self.writes = []
        self.calls = []
        self.ctrl = None
@@ -343,8 +420,11 @@ class _DeliveryHass:
        # A obeys. B produces NOTHING, whatever it is told - and whether that is
        # a wedged power stage or an empty pack is not a question its power
        # register can answer.
+        b = self.ctrl.inverters[1]
+        b_cmd = float((b.last_write if b.last_write is not None else b.target)
+                      or 0.0)
        return [float((a.last_write if a.last_write is not None else a.target)
-                      or 0.0), 0.0]
+                      or 0.0), self.b_frac * b_cmd]

    async def number(self, entity, invert=False):
        m = self.measured()
@@ -630,15 +710,27 @@ print("       schema. Same load, same wedged unit, two configured thresholds.")
 # load is what puts B's shortfall BETWEEN the two: above the shipped threshold,
 # below the raised one. The gap itself is printed by the run that produces it
 # and is never written down here; what is asserted is the ORDERING.
+#
+# ⚠️ AND B IS DERATED HERE RATHER THAN WEDGED, WHICH SAFETY-43 MADE NECESSARY.
+# `delinquent()` now names a unit whose gap exceeds `min_w` OR a fraction of its
+# own share, so a unit delivering NONE of its ask is named at every `min_w`
+# there is - which is the whole point of that card, and which would make the
+# raised arm below fire for a reason that has nothing to do with the watts. A
+# unit delivering HALF of what it is asked for is short in absolute watts and
+# comfortably inside the fraction, so it is the fixture that still isolates
+# `saturation_w`. `_SAT_B_FRAC` is asserted below to be inside the relative
+# arm's silence, so a future change to the fraction turns this red rather than
+# quietly turning the pin vacuous.
 _SAT_RAISED = 1800
 _SAT_LOAD_W = 1600.0
+_SAT_B_FRAC = 0.5
 _SAT_SHIPPED = int(read_option("saturation_w"))
 _sat_lo, _sat_hi = (int(x) for x in re.match(
    r"int\((\d+),(\d+)\)", read_option("saturation_w", "schema")).groups())


-async def _at_saturation_w(sat):
-    hass = _DeliveryHass(load_w=_SAT_LOAD_W)
+async def _at_saturation_w(sat, b_frac=_SAT_B_FRAC):
+    hass = _DeliveryHass(load_w=_SAT_LOAD_W, b_frac=b_frac)
    c = _ctrl(hass, saturation_w=sat)
    return c, await _drive(c, watch=_sample)

@@ -674,6 +766,17 @@ check("the raised arm REACHED the comparison: B was asked for a discharge "
      and not _b_up.sign_wd.tripped
      and account(live=True, parked=False, soc=_b_up.soc,
                  bms_floor=_b_up.bms_soc_floor) is None)
+check("...and SAFETY-43's RELATIVE arm is silent on both arms of this fixture, "
+      "which is what leaves the watts as the only moving part. B delivers half "
+      "of what it is asked for, so its gap is half its share and that is "
+      "inside `DELIVERY_SHARE_FRACTION`. ⚠️ A FIXTURE WHERE THE RELATIVE ARM "
+      "FIRED WOULD NAME B AT EVERY `saturation_w` AND THIS PIN WOULD BE "
+      "VACUOUS, so it is asserted rather than assumed",
+      _gap_up / float(_b_up.standing_w()) <= DELIVERY_SHARE_FRACTION
+      and _gap_ship / float(_b_ship.standing_w()) <= DELIVERY_SHARE_FRACTION
+      and delinquent([U("a", 1000, 1000),
+                      U("b", float(_b_up.standing_w()), _b_up.batt)],
+                     _SAT_RAISED) == {})
 check("...and its shortfall lands STRICTLY BETWEEN the two configured "
      "thresholds, which is the one place the two answers can differ. Part 4's "
      "fixtures are all above both, which is exactly why they cannot see this",
@@ -41,7 +41,9 @@ from app.arbiter import (
    ACTUATOR_SITE, CLAIM_SCOPE, P_LOOP, P_SAFETY, SCOPE_UNIT, Claim, resolve,
 )
 from app.control import Tuning, compute
-from app.delivery import UnitDelivery, account, delinquent
+from app.delivery import (
+    DELIVERY_SHARE_FRACTION, UnitDelivery, account, delinquent,
+)
 from app.peakguard import discharge_permitted

 fails = []
@@ -605,67 +607,87 @@ check("...and SAFETY-42 is where the case this card CANNOT reach is handled. A "
      "site ASKING a pack whose own SoC says it cannot deliver; SAFETY-42 "
      "alarms when a pack that said it could does not. ⚠️ THE WEIGHT, HOWEVER, "
      "DOES MOVE - this check used to say it did not - and part 8b grades what "
-      "that costs the alarm",
+      "that cost the alarm, and what SAFETY-43 did about it",
      not packshare.envelope(PE(Bee, 55.0, B_KWH)).binding)


 # ==============================================================================
 print()
-print("8b. ⚠️ AND WHAT THE WEIGHTS DO TO SAFETY-42's DETECTION WINDOW, WHICH IS "
-      "A REGRESSION AND IS PINNED HERE AS ONE")
+print("8b. ⚠️ WHAT THE WEIGHTS DID TO SAFETY-42's DETECTION WINDOW, AND WHAT "
+      "SAFETY-43 DID ABOUT IT - GRADED UNDER BOTH RULES, ON ONE FIXTURE")

-# ⚠️ READ THIS BEFORE READING THE CHECKS: THEY ASSERT THE UNFLATTERING ANSWER ON
-# PURPOSE. An earlier draft of §5 of the decision said a wedged unit's weight
-# does not move, so OPT-09 "cannot see it at all". The first clause is about the
-# ENVELOPE and is true; the second is FALSE. A wedged unit's weight moves
-# whenever its energy headroom differs from its sibling's - it is a pack fact,
-# not a delivery fact - and `delivery.delinquent()` names a unit only while its
-# `short_by()` gap exceeds `min_w` (`Tuning.saturation_w`, 500 W) IN ABSOLUTE
-# WATTS. Shrinking a unit's share shrinks its gap. Shrink it below 500 W and the
-# alarm stops firing on a unit that is still delivering nothing.
+# ⚠️ THIS PART USED TO PIN THE UNFLATTERING ANSWER ON PURPOSE, AND SAFETY-43 IS
+# THE CARD THAT INVERTED IT. What it said - and it was true when it was written
+# - was: a wedged unit's weight MOVES whenever its energy headroom differs from
+# its sibling's (it is a pack fact, not a delivery fact), `delivery.delinquent()`
+# named a unit only while its `short_by()` gap exceeded `min_w`
+# (`Tuning.saturation_w`, 500 W) IN ABSOLUTE WATTS, and shrinking a unit's share
+# shrinks its gap. Shrink it below 500 W and the alarm stopped firing on a unit
+# that was still delivering nothing. The old text ended "if it ever goes red
+# because the alarm fired, read the card that made it fire before 'fixing' this
+# line". SAFETY-43 is that card, this is that red, and the checks below are the
+# same three measurements with their answers the other way round.
 #
-# So the seam is not "OPT-09 cannot see it". It is: OPT-09 stops the ask AND
-# narrows SAFETY-42's detection window on the same unit. At equal weights the
-# window is `site target < 2 * min_w` = 1000 W; at OPT-09's weights it is
-# `site target < min_w * (w_A + w_B) / w_B`, which on this fixture is 6500 W.
+# ⚠️ AND THE EVIDENCE IS THE DIFF BETWEEN TWO RULES ON ONE RUN, NOT A NUMBER
+# THAT CHANGED IN A COMMIT. `delinquent()` takes `share_frac`, and
+# `share_frac=None` is SAFETY-42's rule EXACTLY - the absolute arm alone. So
+# `wedged_loop` grades every cycle twice, once under each rule, and the red and
+# the green sit side by side in one output:
 #
-# ⚠️ BOTH BOUNDS ARE ARITHMETIC ABOUT `_fill()`'s PROPORTIONAL REGIME AND NEITHER IS A REACHABLE STEADY STATE. Swept on this fixture, the healthy pack's own 5000 W rating binds FIRST: suppressed 0/200 up to a 3000 W load, 4/200 at 4000 W, 41/200 at 4900 W (site target 5320 W, the largest reached with the wedged unit still under 500 W), then 195/200 at 5000 W once `_fill()` pins A at its cap and hands B 3000 W. So the alarm returns BELOW the 6500 W formula bound, not above it. The equal-weights 1000 W bound is unreachable for the same reason - the equal arm names the wedged unit 199/200 at every load swept, 1000 W included.
+#     load 1500 W, OPT-09 weights, absolute arm only   named on   0/200 cycles
+#     load 1500 W, OPT-09 weights, the shipped rule    named on 200/200 cycles
 #
-# ⚠️ THIS IS A DETECTION REGRESSION AT N=2 AND IT IS NOT FIXED HERE. Making
-# `delinquent()` relative to the ask rather than absolute is SAFETY-42's code and
-# a design change; it is filed as card SAFETY-43. What this part does is GRADE THE
-# CURRENT BEHAVIOUR so that the day it changes, a named check moves.
+# A future reader who wants to know what this card did runs this file.
+#
+# ⚠️ WHAT SAFETY-43 DID NOT DO IS LOWER `min_w`, and the check on the window
+# arithmetic below is what holds that open: the ABSOLUTE arm's suppression
+# window is still 500 W wide and still widens to 6500 W under these weights.
+# Nothing that was quiet became noisy; a SECOND way to be named was added, and
+# it is a fraction of the unit's OWN ask, so it fires on a unit delivering none
+# of 125 W exactly as it fires on one delivering none of 1500 W.
 #
 # ⚠️ AND THE FIXTURE IS A SAFETY-42 TRUE POSITIVE BY THAT CARD'S OWN DEFINITION,
 # checked below before anything is concluded from it: B is WEDGED - it accepts
 # its share and produces nothing - at a HEALTHY 15 % against its own 10 % floor,
 # so `envelope()` does not close and `account()` explains nothing. This is
 # exactly the case §5 says the envelope cannot reach, which is why it is the
-# case where the weights' side effect matters.
+# case where the weights' side effect mattered.
 #
 # ⚠️ TWO COPIES OF A. Inverter B does not exist; the plant below is modelled in
 # this file the way part 8's is, and B's ratings are datasheet.

 WEDGED_A_SOC, WEDGED_B_SOC = 90.0, 15.0   # both healthy, both above their floor
 MIN_W = Tuning().saturation_w             # what `delinquent()` is handed
+ABSOLUTE_ONLY = None                      # `share_frac=None` - SAFETY-42's rule


-def wedged_loop(load_w, weighted, cycles=200):
+def wedged_loop(load_w, weighted, cycles=200, share_frac=DELIVERY_SHARE_FRACTION,
+                b_delivers=0.0, b_soc=WEDGED_B_SOC, no_account=False):
    """The closed loop with unit B WEDGED at a healthy SoC. Real everything.

    Real `compute()`, real `plan()`, real `packshare.weights()`, real
    `delivery.account()` and real `delivery.delinquent()`. The only model is the
-    plant: A obeys within one cycle inside its rating, B accepts whatever it is
-    told and delivers 0 W - which is the wedged box, and is what makes this a
+    plant: A obeys within one cycle inside its rating, B delivers `b_delivers`
+    of whatever it is told - 0.0 is the wedged box, and is what makes this a
    SAFETY-42 case rather than an OPT-09 one.

    `weighted=False` is equal weights - what shipped. Envelopes are left OPEN in
    both arms deliberately: neither pack's SoC closes one here, so filing them
    would change nothing and would blur which half of the card is being graded.

+    `share_frac` is SAFETY-43's relative arm. `ABSOLUTE_ONLY` grades the same
+    cycles under SAFETY-42's rule, which is how the red and the green below come
+    out of ONE run over ONE tree instead of out of two commits.
+
+    `no_account=True` DELETES the account gate - every unit is handed
+    `accounted=None` - and it exists to answer "would this arm still be silent
+    if `account()` did nothing at all?". A control that answers the same way
+    with the gate removed is not grading the gate, and the cut-out arm below is
+    exactly that case. See it there; it is the point of that block.
+
    Returns `(cycles on which delinquent() named somebody, last cycle)`.
    """
-    packs = [PE(A, WEDGED_A_SOC, A_KWH), PE(Bee, WEDGED_B_SOC, B_KWH)]
+    packs = [PE(A, WEDGED_A_SOC, A_KWH), PE(Bee, b_soc, B_KWH)]
    tuning = Tuning(gain=1.0, slew_w=1000.0, max_w=A_W + B_W)
    prev, i_w, sat = 0.0, None, 0
    power = [0.0, 0.0]
@@ -678,13 +700,15 @@ def wedged_loop(load_w, weighted, cycles=200):
        r = plan(d.target_w, {A: _B(-A_W, A_W), Bee: _B(-B_W, B_W)}, weights=w)
        assert r.ok, r.refused
        # THE MODELLED PLANT: A within one cycle inside its rating; B WEDGED.
-        power = [max(-A_W, min(A_W, float(r.shares[A]))), 0.0]
+        power = [max(-A_W, min(A_W, float(r.shares[A]))),
+                 b_delivers * max(-B_W, min(B_W, float(r.shares[Bee])))]
        grid = load_w - sum(power)
        who = delinquent(
            [UnitDelivery(p.name, r.shares[p.name], m,
+                          None if no_account else
                          account(live=True, parked=False, soc=p.soc,
                                  bms_floor=p.bms_floor))
-             for p, m in zip(packs, power)], MIN_W)
+             for p, m in zip(packs, power)], MIN_W, share_frac)
        named += bool(who)
        last = (d.target_w, dict(r.shares), grid, who)
    return named, last
@@ -698,75 +722,271 @@ check("PRECONDITION: the wedged unit is a SAFETY-42 TRUE POSITIVE and NOT an "
      not packshare.envelope(_WEDGED_PACK).binding
      and account(live=True, parked=False, soc=WEDGED_B_SOC,
                  bms_floor=FLOOR) is None)
+check("PRECONDITION: the shipped fraction is a fraction, and it is STRICTLY "
+      "below 1.0 - at 1.0 the relative arm could only fire on a unit IMPORTING "
+      "while asked to discharge, which is SAFETY-30's fault and not this one, "
+      "and this card's own case (a unit delivering NONE of its share, gap over "
+      "share exactly 1.0) would be missed",
+      0.0 < DELIVERY_SHARE_FRACTION < 1.0)

 _wedge = {}
 for _load in (1500.0, 6000.0):
    for _weighted in (False, True):
-        _n, _t = wedged_loop(_load, _weighted)
-        _wedge[(_load, _weighted)] = (_n, _t)
-        print(f"    load {_load:.0f} W, B wedged at 15 %, "
-              f"{'OPT-09 weights' if _weighted else 'equal weights ':<14} "
-              f"named on {_n:3d}/200 cycles, site target {_t[0]:+.0f} W, shares "
-              + ", ".join(f"{u}{v:+d}" for u, v in sorted(_t[1].items()))
-              + f", grid {_t[2]:+.0f} W, named="
-              + (str(sorted(_t[3])) if _t[3] else "(nobody)"))
+        for _frac in (ABSOLUTE_ONLY, DELIVERY_SHARE_FRACTION):
+            _n, _t = wedged_loop(_load, _weighted, share_frac=_frac)
+            _wedge[(_load, _weighted, _frac)] = (_n, _t)
+            print(f"    load {_load:.0f} W, B wedged at 15 %, "
+                  f"{'OPT-09 weights' if _weighted else 'equal weights ':<14} "
+                  + ("absolute arm only" if _frac is ABSOLUTE_ONLY
+                     else f"frac {_frac:.2f}         ")
+                  + f", named on {_n:3d}/200 cycles, site target {_t[0]:+.0f} W,"
+                  + " shares "
+                  + ", ".join(f"{u}{v:+d}" for u, v in sorted(_t[1].items()))
+                  + f", grid {_t[2]:+.0f} W, named="
+                  + (str(sorted(_t[3])) if _t[3] else "(nobody)"))

-check("AT EQUAL WEIGHTS - WHAT SHIPPED - the wedged unit is named on every "
-      "cycle but the first, at a load its sibling can carry on its own. The "
-      "loop winds the site target to twice the load, the 50/50 split stands "
-      "+1500 W on a box producing nothing, and 1500 W > 500 W is what "
-      "`delinquent()` reads",
-      _wedge[(1500.0, False)][0] == 199
-      and sorted(_wedge[(1500.0, False)][1][3]) == [Bee]
-      and _wedge[(1500.0, False)][1][1] == {A: 1500, Bee: 1500})
+check("⚠️ THE CARD, RED: UNDER SAFETY-42's RULE AND OPT-09's WEIGHTS THE WEDGED "
+      "UNIT IS NAMED ON NONE OF THE 200. Same wedged box, same healthy SoC, "
+      "same load, same detector: the split moves to 1505/125, the gap on the "
+      "wedged unit falls to 125 W, and 125 W < 500 W so the absolute arm "
+      "returns {}. This is the defect SAFETY-43 was filed for, and it is "
+      "measured here rather than remembered",
+      _wedge[(1500.0, True, ABSOLUTE_ONLY)][0] == 0
+      and _wedge[(1500.0, True, ABSOLUTE_ONLY)][1][3] == {}
+      and _wedge[(1500.0, True, ABSOLUTE_ONLY)][1][1] == {A: 1505, Bee: 125})

-check("⚠️ AND WITH OPT-09's WEIGHTS IT IS NAMED ON NONE OF THE 200. Same wedged "
-      "box, same healthy SoC, same load, same detector: the split moves to "
-      "1505/125, the gap on the wedged unit falls to 125 W, and 125 W < 500 W "
-      "so `delinquent()` returns {}. THIS CHECK PINS A REGRESSION - it asserts "
-      "what the code DOES, not what anyone wants, and if it ever goes red "
-      "because the alarm fired, read the card that made it fire before "
-      "'fixing' this line",
-      _wedge[(1500.0, True)][0] == 0
-      and _wedge[(1500.0, True)][1][3] == {}
-      and _wedge[(1500.0, True)][1][1] == {A: 1505, Bee: 125})
+check("⚠️ THE CARD, GREEN: THE SHIPPED RULE NAMES IT ON ALL 200, ON THE SAME "
+      "RUN. The unit is standing at +125 W and delivering 0 W - 100 % of its "
+      "ask is not arriving - and `gap > frac * share` does not care that 125 W "
+      "is a small number. The ONLY input that moved between this check and the "
+      "one above is `share_frac`",
+      _wedge[(1500.0, True, DELIVERY_SHARE_FRACTION)][0] == 200
+      and sorted(_wedge[(1500.0, True, DELIVERY_SHARE_FRACTION)][1][3]) == [Bee]
+      and _wedge[(1500.0, True, DELIVERY_SHARE_FRACTION)][1][1] \
+          == {A: 1505, Bee: 125})
+
+check("AT EQUAL WEIGHTS - WHAT SHIPPED BEFORE OPT-09 - the wedged unit was "
+      "already named under the absolute rule on every cycle but the first, at "
+      "a load its sibling can carry on its own. The loop winds the site target "
+      "to twice the load, the 50/50 split stands +1500 W on a box producing "
+      "nothing, and 1500 W > 500 W is what the absolute arm reads",
+      _wedge[(1500.0, False, ABSOLUTE_ONLY)][0] == 199
+      and sorted(_wedge[(1500.0, False, ABSOLUTE_ONLY)][1][3]) == [Bee]
+      and _wedge[(1500.0, False, ABSOLUTE_ONLY)][1][1] == {A: 1500, Bee: 1500})
+
+check("...and the shipped rule adds exactly the FIRST cycle to that arm and "
+      "nothing else - 199 becomes 200. On cycle one the slew limit holds the "
+      "split at 500/500, a 500 W gap is NOT over a 500 W threshold (the "
+      "absolute bound is strict), and the relative arm names a box delivering "
+      "none of its 500 W. That one cycle is the whole of what this card costs "
+      "an arm that was already firing",
+      _wedge[(1500.0, False, DELIVERY_SHARE_FRACTION)][0] == 200
+      and _wedge[(1500.0, False, DELIVERY_SHARE_FRACTION)][0]
+          - _wedge[(1500.0, False, ABSOLUTE_ONLY)][0] == 1)

 check("...and this is the regime where NOTHING ELSE TELLS THE OPERATOR "
      "ANYTHING: the meter is satisfied (grid within 5 W of zero), `unmet_w` is "
      "0 because both units accepted their shares, no envelope closed and no "
-      "action-log line fires. The site is quietly running one inverter and the "
-      "one instrument built to say so has gone quiet with it",
-      abs(_wedge[(1500.0, True)][1][2]) <= 5.0
-      and abs(_wedge[(1500.0, False)][1][2]) <= 5.0)
+      "action-log line fires. The site is quietly running one inverter, and "
+      "before this card the one instrument built to say so had gone quiet with "
+      "it",
+      abs(_wedge[(1500.0, True, DELIVERY_SHARE_FRACTION)][1][2]) <= 5.0
+      and abs(_wedge[(1500.0, False, DELIVERY_SHARE_FRACTION)][1][2]) <= 5.0)

 _w = packshare.weights([PE(A, WEDGED_A_SOC, A_KWH), _WEDGED_PACK], 1000.0)
 _window_equal = MIN_W * 2.0
 _window_opt09 = MIN_W * sum(_w.values()) / _w[Bee]
-print(f"    suppression window on this fixture: equal weights, site target "
-      f"< {_window_equal:.0f} W; OPT-09 weights, site target "
-      f"< {_window_opt09:.0f} W of a {A_W + B_W:.0f} W site")
-check("...and the arithmetic behind it, stated as a window rather than as one "
-      "fixture: `delinquent()` reads an ABSOLUTE gap, so a unit weighted to a "
-      "fraction f of the site target is invisible while `f * target < min_w`. "
-      "Equal weights suppress below 1000 W; these weights suppress below "
-      "6500 W of a 9600 W site. ⚠️ BOTH ARE PROPORTIONAL-REGIME BOUNDS AND "
-      "NEITHER IS REACHABLE: the healthy pack's own rating binds first, so the "
-      "alarm returns BELOW 6500 W (swept: 0/200 to a 3000 W load, 41/200 at "
-      "4900 W, 195/200 at 5000 W) and the equal arm never suppresses at all. "
-      "That is the regression, and it is a property "
-      "of the threshold being absolute rather than of any number chosen here",
-      abs(_window_equal - 1000.0) < 1.0 and abs(_window_opt09 - 6500.0) < 1.0)
+print(f"    the ABSOLUTE arm's suppression window is UNCHANGED by SAFETY-43: "
+      f"equal weights, site target < {_window_equal:.0f} W; OPT-09 weights, "
+      f"site target < {_window_opt09:.0f} W of a {A_W + B_W:.0f} W site")
+check("⚠️ AND `min_w` WAS NOT LOWERED, WHICH IS THE HALF OF THIS CARD THAT IS "
+      "EASIEST TO GET WRONG. The absolute arm still suppresses below 1000 W at "
+      "equal weights and below 6500 W at these ones - the same arithmetic this "
+      "part has always printed - because SAFETY-43 ADDED a second way to be "
+      "named rather than loosening the first. Lowering `saturation_w` would "
+      "have reported normal saturation on every unit at once, which is the "
+      "false-trip regime SAFETY-41 spent four attempts in",
+      abs(_window_equal - 1000.0) < 1.0 and abs(_window_opt09 - 6500.0) < 1.0
+      and MIN_W == Tuning().saturation_w == 500.0)
+
+check("⚠️ AND THE ABSOLUTE ARM WAS NEVER SILENT EVERYWHERE - which is what a "
+      "check that only ever ran the quiet load would be pinning. Push the load "
+      "past the healthy pack's OWN RATING and `_fill()`'s last-resort pass "
+      "hands the wedged unit 3000 W anyway, the gap goes back over 500 W, and "
+      "both arms named it under SAFETY-42's rule too. The window was real and "
+      "it had an upper edge; the hole was the quiet regime below that edge, "
+      "and the shipped rule now names the unit on every cycle of all four runs",
+      _wedge[(6000.0, True, ABSOLUTE_ONLY)][0] == 195
+      and _wedge[(6000.0, False, ABSOLUTE_ONLY)][0] == 199
+      and _wedge[(6000.0, True, DELIVERY_SHARE_FRACTION)][0] == 200
+      and _wedge[(6000.0, False, DELIVERY_SHARE_FRACTION)][0] == 200
+      and sorted(_wedge[(6000.0, True, DELIVERY_SHARE_FRACTION)][1][3]) == [Bee])
+
+print()
+print("   THE POSITIVE CONTROL: what the relative arm must STILL NOT say")
+
+# ⚠️ A CARD THAT ADDS A WAY TO RAISE AN ALARM HAS TO SHOW WHAT IT DID NOT ADD,
+# or it has traded a missed alarm for a false one - which in this system is not
+# a rhetorical risk. SAFETY-41 took four attempts on notification cadence and
+# the owner's words were "i'm getting really annoyed by the notifications in my
+# HA". So: the same fixture, the same weights, the same tiny 125 W share, with
+# the ONE fact that makes the shortfall legitimate changed.
+
+# ⚠️ AND THE FIRST DRAFT OF THIS BLOCK WAS VACUOUS, WHICH IS WHY IT NOW PRINTS
+# ITS SHARES. It ran the cut-out arm, saw 0/200, and concluded "`account()`
+# explains it" and "identical trace to the wedged arm". BOTH WERE FALSE. At its
+# own cut-out the pack's usable energy ABOVE that cut-out is zero, so
+# `weights()` gives it 0.0, `_fill()`'s preferred pass gives it nothing, the
+# healthy sibling covers the load alone so the LAST-RESORT pass never runs, and
+# its share is 0 W on ALL 200 CYCLES. A unit with a 0 W share is not in `asked`,
+# so `len(asked) < 2` returns {} BEFORE EITHER ARM IS EVALUATED and `account()`
+# is never consulted - measured below by deleting the gate and getting the same
+# answer. The wedged arm's share is the 125 W printed on its own line above;
+# this one's is 0, printed and asserted below. Nothing was identical and nothing
+# reached the comparison.
+#
+# ⚠️ THE FACT THAT REPLACES IT IS BETTER, NOT WEAKER: at its own cut-out OPT-09
+# STOPS ASKING THE PACK AT ALL, so SAFETY-43's relative arm is never reached.
+# That is a SECOND, INDEPENDENT protection sitting UPSTREAM of `account()` - the
+# allocator declines to ask before the detector has to decline to alarm. The
+# account gate itself is graded where it is REAL, on a positive 125 W share, in
+# `test_delivery.py` part 1 ("THE ACCOUNT GATE IS UNTOUCHED BY IT").
+#
+# ⚠️ AND IT WAS INVISIBLE FOR ONE REASON: this was the only line in part 8b that
+# did not print its shares. `inverter-b+0` would have been on screen from the
+# first run. Every line in this part prints them now.
+_sat_named, _sat_last = wedged_loop(1500.0, True, b_soc=FLOOR)
+_sat_nogate, _sat_nogate_last = wedged_loop(1500.0, True, b_soc=FLOOR,
+                                            no_account=True)
+print(f"    load 1500 W, B AT ITS OWN {FLOOR:.0f} % CUT-OUT, OPT-09 weights, "
+      f"frac {DELIVERY_SHARE_FRACTION:.2f}: named on {_sat_named:3d}/200 "
+      f"cycles, shares "
+      + ", ".join(f"{u}{v:+d}" for u, v in sorted(_sat_last[1].items()))
+      + ", named="
+      + (str(sorted(_sat_last[3])) if _sat_last[3] else "(nobody)")
+      + f"; with account() DELETED: {_sat_nogate:3d}/200")
+# ⚠️ AND THE `no_account=True` SWITCH ITSELF WAS GRADED BY NOTHING. It is used
+# in exactly one place - the fixture above - where it PROVABLY cannot change the
+# answer, because a cut-out pack is never asked and the gate is never reached.
+# Review neutered the switch so it removed nothing: the file stayed GREEN and
+# printed "with account() DELETED: 0/200" while `account()` had not been
+# deleted. A line asserting a mechanism that did not happen - B1's shape, one
+# level down, in the fix for B1.
+#
+# The switch DOES bite, on a fixture already reachable from this helper: at the
+# cut-out with EQUAL weights, `plan()` hands B a 50/50 share, so the gate is the
+# only thing keeping it silent. Measured: gate on 0/200, gate deleted 200/200.
+# That proves the switch removes the gate, AND gives this part the end-to-end
+# account-gate control on a POSITIVE share that it otherwise has to outsource to
+# test_delivery.py part 1.
+_eq_gate, _eq_last = wedged_loop(1500.0, False, b_soc=FLOOR)
+_eq_nogate, _ = wedged_loop(1500.0, False, b_soc=FLOOR, no_account=True)
+print(f"    ...and the SWITCH is real: same cut-out, EQUAL weights so B IS "
+      f"asked ({_eq_last[1][Bee]:+d} W last cycle) - gate on {_eq_gate:3d}/200, "
+      f"gate DELETED {_eq_nogate:3d}/200")
+check("⭐ THE `no_account` SWITCH ACTUALLY REMOVES THE GATE, on a fixture where "
+      "the gate is the only thing keeping the pack silent. Without this the "
+      "line above prints 'with account() DELETED' about a run in which nothing "
+      "was deleted - which is exactly the defect this block was rewritten for",
+      _eq_gate == 0 and _eq_nogate == 200 and _eq_last[1][Bee] > 0)
+
+check("A PACK AT ITS OWN CUT-OUT IS NEVER ASKED, SO THE RELATIVE ARM IS NEVER "
+      "REACHED. Its usable energy above its own cut-out is zero, `weights()` "
+      "answers 0.0 for it, the healthy sibling covers the load so `_fill()`'s "
+      "last-resort pass never runs, and its share is 0 W on every one of the "
+      "200 cycles - so it is not in `asked` and `len(asked) < 2` returns {} "
+      "before either arm is compared. ⚠️ THE SHARE IS ASSERTED, not inferred "
+      "from the silence: a control that cannot say WHY it is quiet is a control "
+      "that cannot fail",
+      _sat_named == 0 and _sat_last[3] == {} and _sat_last[1][Bee] == 0)
+check("⚠️ AND THAT IS A SECOND, INDEPENDENT PROTECTION - NOT THE ACCOUNT GATE, "
+      "WHICH THIS ARM DOES NOT TOUCH. Delete `account()` entirely, hand every "
+      "unit `accounted=None`, and this arm is STILL silent on all 200 cycles, "
+      "because the protection is UPSTREAM of it: the allocator declines to ask "
+      "before the detector has to decline to alarm. ⚠️ SO DO NOT READ THIS "
+      "BLOCK AS EVIDENCE ABOUT `account()` - it is evidence about the ask. The "
+      "account gate on a POSITIVE share is graded in test_delivery.py part 1",
+      _sat_nogate == 0 and _sat_nogate_last[3] == {})
+
+_derate_named, _derate_last = wedged_loop(1500.0, True, b_delivers=0.5)
+print(f"    load 1500 W, B delivering HALF its share, OPT-09 weights, frac "
+      f"{DELIVERY_SHARE_FRACTION:.2f}: named on {_derate_named:3d}/200 cycles, "
+      f"shares "
+      + ", ".join(f"{u}{v:+d}" for u, v in sorted(_derate_last[1].items()))
+      + ", named="
+      + (str(sorted(_derate_last[3])) if _derate_last[3] else "(nobody)"))
+check("...and ORDINARY TRACKING ON A SMALL SHARE IS NOT DELINQUENCY EITHER. "
+      "The same unit at the same weights delivering HALF of what it is asked "
+      "for is short by 50 % - well inside the shipped fraction - and its gap "
+      "in watts is far under `min_w`, so NEITHER arm names it. A relative "
+      "criterion that fired here would be the saturation-shaped alarm this "
+      "whole module is written against",
+      _derate_named == 0 and _derate_last[3] == {})
+
+check("...and the DERATED control is not vacuous, which is asserted on the ONE "
+      "thing that could make it so: unlike the cut-out arm above, its unit IS "
+      "asked - a positive share on every cycle - so it really does reach both "
+      "arms and is turned away by the size of its shortfall. ⚠️ THIS GUARD USED "
+      "TO RE-ASSERT THE WEDGED ARM's 200/200 INSTEAD, WHICH IS WHY IT DID NOT "
+      "CATCH THE CUT-OUT ARM BEING VACUOUS: a check on a DIFFERENT run cannot "
+      "say anything about whether THIS one reached the comparison",
+      _derate_last[1][Bee] > 0
+      and _wedge[(1500.0, True, DELIVERY_SHARE_FRACTION)][0] == 200)
+
+print()
+print("   N=1 IS BYTE-IDENTICAL, AND THAT IS ASSERTED RATHER THAN ARGUED")
+
+# ⚠️ THE LIVE INSTALL IS N=1 AND THIS CARD MUST NOT REACH IT. Two independent
+# statements, and each is checked against the thing that would break it:
+#   1. `plan()` RETURNS BEFORE `weights` IS READ AT ALL, so no weighting - and
+#      no change to how one is derived - can move a one-inverter share. Graded
+#      with a `weights` object that RAISES if anything reads it.
+#   2. `delinquent()` returns {} at a roster of one for every input, including
+#      the input the RELATIVE arm was built for. The proper-subset algebra runs
+#      before either arm, so the fraction is never even multiplied.
+
+
+class _Poison(dict):
+    """A `weights` that cannot be read without saying so. Not a mock: a tripwire."""
+
+    def get(self, *a, **k):
+        raise AssertionError("plan() read weights at a roster of one")
+
+    def __getitem__(self, k):
+        raise AssertionError("plan() read weights at a roster of one")
+
+
+_n1_plain = plan(1234.5, {A: _B(-A_W, A_W)})
+_n1_poison = plan(1234.5, {A: _B(-A_W, A_W)}, weights=_Poison())
+check("`plan()` at a roster of ONE returns the site target unread, and it "
+      "returns it WITHOUT TOUCHING `weights` - handed a mapping that raises on "
+      "any read, it answers identically to the same call with no weights at "
+      "all. The N=1 pass-through is above the weighting, so OPT-09 could not "
+      "move it and SAFETY-43 does not either",
+      _n1_poison.shares == _n1_plain.shares == {A: 1234.5}
+      and _n1_poison.unmet_w == _n1_plain.unmet_w)
+try:
+    plan(1234.5, {A: _B(-A_W, A_W), Bee: _B(-B_W, B_W)}, weights=_Poison())
+except AssertionError:
+    _poison_bites = True
+else:
+    _poison_bites = False
+check("...and the tripwire is a real one: the SAME poisoned mapping at a roster "
+      "of TWO is read and raises. A check that could not fail would be "
+      "measuring nothing",
+      _poison_bites)
+
+_n1_rel = [delinquent([UnitDelivery(A, s, m)], MIN_W, f)
+           for s in (0, 125, 500, 2000, 5000, -2000, None)
+           for m in (0, 60, 125, 2000, -2000, None)
+           for f in (ABSOLUTE_ONLY, DELIVERY_SHARE_FRACTION)]
+check(f"...and `delinquent()` answers {{}} at a roster of one on all "
+      f"{len(_n1_rel)} of (share, measured, rule) - INCLUDING 0 of 125 W, the "
+      f"exact input this card added an arm for. `len(asked) < 2` returns "
+      f"before either threshold is compared, so the one-inverter install this "
+      f"add-on is running today cannot tell that SAFETY-43 happened",
+      all(r == {} for r in _n1_rel))

-check("⚠️ AND IT IS NOT THAT THE WEIGHTED ARM NEVER NAMES ANYBODY - which is "
-      "what a check that only ever ran the quiet load would be pinning. Push "
-      "the load past the healthy pack's OWN RATING and `_fill()`'s last-resort "
-      "pass hands the wedged unit 3000 W anyway, the gap goes back over 500 W, "
-      "and BOTH arms name it. The window is real and it has an upper edge; the "
-      "hole is the quiet regime below that edge",
-      _wedge[(6000.0, True)][0] == 195
-      and sorted(_wedge[(6000.0, True)][1][3]) == [Bee]
-      and _wedge[(6000.0, False)][0] == 199
-      and sorted(_wedge[(6000.0, False)][1][3]) == [Bee])

 # ==============================================================================
 print()
@@ -978,9 +1198,12 @@ print("        a preference and are honest about being one.")
 print("      * A WEDGED UNIT IS NOT REACHABLE BY THE ENVELOPE. Its SoC reads")
 print("        healthy, so nothing closes, and SAFETY-42 owns the detection.")
 print("        ⚠️ BUT IT IS REACHABLE BY THE WEIGHTS, WHICH MOVE ON ENERGY")
-print("        HEADROOM AND NOT ON DELIVERY - part 8b grades what that does")
-print("        to SAFETY-42's window, and the answer is a REGRESSION that is")
-print("        pinned rather than fixed.")
+print("        HEADROOM AND NOT ON DELIVERY. That narrowed SAFETY-42's")
+print("        detection window - a regression this file PINNED rather than")
+print("        fixed - and SAFETY-43 closed it by making delinquency relative")
+print("        to the ask as well as absolute. Part 8b now grades both rules")
+print("        on one fixture: the absolute arm still says nothing, and the")
+print("        shipped one names the unit on all 200 cycles.")

 # ==============================================================================
 print()
@@ -1023,13 +1246,41 @@ MUT = [
     "mutation took it down with a TypeError three lines into part 1"),
    ("M11 weights(): keep the shape, force every weight to 1.0  <- THE PROOF "
     "THAT PART 8b CAN GO RED",
-     "test_packshare 8b, three named checks: the wedged unit goes back to being "
-     "named on 199/200 at 1500 W, the split reverts to 1500/1500 and the "
-     "window collapses from 6500 W to 1000 W. ⚠️ RUN, NOT REASONED ABOUT - a "
-     "check that pins an UNFLATTERING behaviour has to be shown capable of "
-     "failing, or it is indistinguishable from a check that pins nothing. "
+     "test_packshare 8b, FIVE named checks - re-run for SAFETY-43 rather than "
+     "carried across, because the part it grades was rewritten: the split "
+     "reverts to 1500/1500 so the absolute arm names the unit on 199/200 "
+     "(taking the card's RED check down), the shipped rule's 200/200 comes "
+     "from the wrong shares (taking the GREEN one down), the OPT-09 window "
+     "collapses from 6500 W to 1000 W, the 6000 W arm's 195 becomes 199, and "
+     "the derated positive control moves from 0/200 to 3/200. ⚠️ RUN, NOT "
+     "REASONED ABOUT - a check that pins a behaviour has to be shown capable "
+     "of failing, or it is indistinguishable from a check that pins nothing. "
     "(It also takes parts 1, 3, 7 and 8 down, which is M3's territory; 8b is "
     "the part that names the SAFETY-42 interaction)"),
+    # --- SAFETY-43's three, and they pin the new arm from BOTH sides --------
+    ("M12 delinquent(): drop the relative arm, keep `gap > min_w`  <- THE "
+     "CARD ITSELF REVERTED",
+     "SEVEN named checks: test_delivery 1 (THREE - the 0-of-125 W naming, the "
+     "proportion in the sentence, and the strict relative bound) and "
+     "test_packshare 8b (FOUR - the GREEN check goes back to 0/200 under "
+     "OPT-09's weights, the equal arm's 200 back to 199, the 6000 W arm back "
+     "to 195/199, and the derated control's vacuity guard falls with them). "
+     "test_alloc and test_control stay green, which is the point - nothing "
+     "outside this check moves"),
+    ("M13 DELIVERY_SHARE_FRACTION 0.85 -> 1.0  <- THE TOO-LOOSE DIRECTION",
+     "SEVEN named checks: test_delivery 1 (TWO) and test_packshare 8b (FIVE), "
+     "including the PRECONDITION that the fraction is strictly below 1.0. At "
+     "1.0 the arm can only fire on a unit IMPORTING while asked to discharge, "
+     "so the card's own case - 0 of 125 W, gap over share exactly 1.0 - is "
+     "missed and the whole card is a no-op"),
+    ("M14 DELIVERY_SHARE_FRACTION 0.85 -> 0.05  <- THE TOO-SENSITIVE ONE",
+     "SIX named checks: test_delivery 1 (TWO - the strict bound, and 'IT TOOK "
+     "NOTHING AWAY' stops being true) and 4b (THREE - the `saturation_w` pin "
+     "goes VACUOUS, because the relative arm names B at every threshold), plus "
+     "test_packshare 8b's DERATED positive control, where a unit delivering "
+     "half its share starts being named. ⚠️ THIS IS THE ROW THAT MATTERS - the "
+     "fraction is pinned from BOTH directions, so it is a derived number with "
+     "checks around it rather than a literal nobody can move"),
 ]
 for _n, _w in MUT:
    print(f"    caught  {_n}")
@@ -1039,8 +1290,17 @@ check("every mutation in the table above was RUN and every one of them is "
      "M6: the whole defect this card fixes - a parameter nothing passes - is "
      "invisible to every behavioural fixture in this repository, because they "
      "all build their own claim list and call `resolve()` themselves. That is "
-      "why the call site is pinned statically as well",
-      len(MUT) == 11)
+      "why the call site is pinned statically as well. ⚠️ M12-M14 ARE "
+      "SAFETY-43's and were run the same way, on trees copied fresh and "
+      "mounted :ro so no `.pyc` could go stale - MUTATION.md's defect. All "
+      "three are REDS, which is the direction that defect cannot fabricate. "
+      "⚠️ AND THEIR COUNTS WERE RE-RUN IN REVIEW AND TWO OF THEM MOVED: "
+      "test_delivery.py part 1 used to subscript `delinquent(...)['b']` for a "
+      "print, which raised KeyError under M12 and M13 and ABORTED THE FILE at "
+      "part 1, leaving parts 2-8 ungraded and the recorded coverage smaller "
+      "than the truth. It is a `.get` now. M14's count was simply wrong: five "
+      "named checks in test_delivery, not four",
+      len(MUT) == 14)

 print()
 if fails:
@@ -0,0 +1,743 @@
+"""Runnable check for the roster rehearsal.  `python3 test_rehearse.py`
+
+ARCH-02. `tools/rehearse_roster.py` is the instrument that decides whether the
+flat -> roster switch on 2026-09-12 loses a safety latch, a month of maintenance
+history, or a Home Assistant entity that an automation is watching. This file
+grades the instrument.
+
+⚠️ AN INSTRUMENT THAT HAS NEVER BEEN SEEN TO GO RED IS NOT EVIDENCE. This
+repository's most-repeated defect, `MUTATION.md` habit 4 and now FIFTEEN
+recorded instances, is "a check whose DATA cannot reach the failure it names" -
+and a rehearsal that reports "migration OK" because it compared a structure to
+itself is the purest example of it there is. So this file does two things in
+order, and the first one is the one that makes the second mean anything:
+
+  Part 2  TWO CLEAN CONTROLS. Two (state file, roster) pairs on which the
+          shipped code is CORRECT and the harness must report NOTHING. A
+          harness that is red on every input has measured nothing, and every
+          red below it would be free.
+
+  Part 3  ⭐ SIX DELIBERATE MUTATIONS OF THE SHIPPED MIGRATION AND OF THE
+          ENTITY-ID COMPOSITION, each applied to a private copy of the tree,
+          each required to flip its control from green to red, AND each
+          required to produce the SPECIFIC finding it should - not merely a
+          non-zero exit. A mutation that goes red for the wrong reason is a
+          harness that would go red on the real defect for the wrong reason
+          too.
+
+  Part 4  TWO MUTATIONS THAT DELIBERATELY SURVIVE, with the reason written
+          down. A battery in which everything dies is a battery whose author
+          stopped as soon as the number looked good - and the `object_id` one
+          is what BOUNDS the four `app/mqtt.py` mutations above it: without a
+          change to that file that must NOT be caught, "the reference notices
+          any edit to mqtt.py" and "the reference notices what HA keys on" are
+          the same green.
+
+  Part 5b THE PER-LEG CHECK'S OWN POSITIVE CONTROL: a build in which exactly
+          ONE of the two legs is clean. ⚠️ A per-thing check that has never
+          seen the things DISAGREE is a total with extra steps, and this file
+          shipped that defect twice - see `legs()`.
+
+  Part 5  SAFETY-44's exact path, pinned as a finding, by name, in both
+          directions.
+
+⚠️ THE MUTATIONS ARE SOURCE EDITS IN A COPIED TREE, NOT MONKEYPATCHES. A
+monkeypatch grades the name a module was asked for; an edit grades the code that
+runs. Each copy is made with `mkdtemp()`, so two runs of this file cannot
+overwrite each other's work - habit 18 - and `__pycache__` is never copied, so
+no mutant is ever graded through a stale `.pyc` - habit 1 / TEST-15.
+"""
+
+import json
+import os
+import re
+import shutil
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+
+HERE = Path(__file__).resolve().parent
+FIX = HERE / "tools" / "rehearsal"
+
+fails = []
+
+
+def check(name, cond):
+    print(f"  {'ok  ' if cond else 'FAIL'} {name}")
+    if not cond:
+        fails.append(name)
+
+
+# --------------------------------------------------------------------------- #
+print("1. the harness's own preconditions")
+# --------------------------------------------------------------------------- #
+sys.path.insert(0, str(HERE / "tools"))
+import rehearse_roster as RR  # noqa: E402
+
+check("the tool refuses rather than passing when it cannot run: its three exit "
+      "codes are distinct and 2 is not 0",
+      (RR.EXIT_OK, RR.EXIT_FINDING, RR.EXIT_COULD_NOT_RUN) == (0, 1, 2))
+
+# ⚠️ THE ANCHOR. `app/mqtt.py` records `sensor.goodwe_p1_sample_age` as VERIFIED
+# against the live system on 2026-08-26, and the ESP32's ingest guard subscribes
+# to it. The tool's entity-id model is written independently of `mqtt._slug` on
+# purpose (habit 4), so this is the one external fact that says the independent
+# model is right. If it fails, every orphan claim the tool makes is worthless
+# and the tool exits 2 rather than reporting.
+try:
+    RR._check_anchor()
+    anchored = True
+except RR.Refused:
+    anchored = False
+check("the entity-id model reproduces the ONE id verified against the live "
+      f"system ({RR.ANCHOR_ID})", anchored)
+
+check("...and it is NOT built out of `mqtt._slug`, so a mutation in the "
+      "shipped slugifier moves one side of the comparison and not both",
+      "_slug" not in RR.compose.__code__.co_names)
+
+_opts = RR.load_options(str(FIX / "roster.yaml"))
+check("the built-in YAML reader parses the roster fixture the check image "
+      "cannot parse with PyYAML (there is no `yaml` module in it)",
+      [e["name"] for e in _opts["inverters"]] == ["garage", "shed"]
+      and _opts["max_w"] == 5000 and _opts["batt_invert"] is False
+      and _opts["meter_source"] == "off")
+try:
+    RR._mini_yaml("inverters: [a, b]\n")
+    _refused = False
+except RR.Refused:
+    _refused = True
+check("...and it REFUSES a construct it does not understand rather than "
+      "guessing, because a config silently mis-parsed here rehearses a "
+      "configuration nobody is going to deploy", _refused)
+
+_flat = RR.load_options(str(FIX / "flat.yaml"))
+check("⚠️ the two configs differ in the ONE way the transition is about: the "
+      "BEFORE roster is empty and the AFTER roster is declared - if they did "
+      "not, every diff below would be a structure compared to itself",
+      _flat.get("inverters") == [] and len(_opts["inverters"]) == 2)
+check("...and the AFTER roster's first entry is NOT `inverter-a`, which is "
+      "habit 4's seventh instance in one line: with it, `UNIT_A` and "
+      "`inverters[0].name` are the same string and a misdirected namespace "
+      "cannot be seen",
+      _opts["inverters"][0]["name"] != RR.UNIT_A)
+
+_legacy = json.loads((FIX / "state-legacy.json").read_text())
+check("⚠️ the synthetic state carries MAINTENANCE HISTORY - state built by the "
+      "happy path cannot detect a migration that drops it",
+      _legacy.get("last_completed") and _legacy.get("phase"))
+check("⚠️ ...and a LEGACY-SHAPED KEY WHOSE LOSS WOULD BE SILENT: a LATCHED "
+      "`sign_tripped` with no `schema_version` beside it, i.e. written by a "
+      "controller that had never heard of unit namespaces",
+      _legacy.get("sign_tripped") is True
+      and "schema_version" not in _legacy
+      and "maint_schema_version" not in _legacy)
+
+
+# --------------------------------------------------------------------------- #
+print()
+print("2. THE CLEAN CONTROLS: three transitions on which the shipped code is "
+      "right and the harness must say nothing")
+# --------------------------------------------------------------------------- #
+# ⚠️ THIS PART IS WHAT MAKES PART 3 EVIDENCE. Six reds prove nothing about an
+# instrument that is red on its own fixtures. Each control below is a genuine
+# flat -> roster switch that the shipped migration handles correctly, and the
+# required answer is EXIT 0. CONTROL-C earned its place by failing: it found the
+# harness reporting a defect on every fresh install.
+#
+#   CONTROL-A  a PRE-DEV-02 flat store, latch tripped, onto a roster whose
+#              first entry IS `inverter-a`. This is the path SAFETY-44's
+#              documented reasoning describes, and on it that reasoning holds.
+#   CONTROL-B  a store that has already run flat under 0.7.0 (schema stamped,
+#              no latch), onto a roster named `garage`/`shed`. This is the
+#              MAINT-05 half, and the names are deliberately not `inverter-a`.
+#   CONTROL-C  a FRESH INSTALL - an empty state file - onto `garage`/`shed`.
+#              ⚠️ THIS ONE FOUND A DEFECT IN THE HARNESS. A fresh store's
+#              migration writes `sign_tripped = False` and `sign_reason = ""`
+#              under `inverter-a` before any roster exists, and the first draft
+#              reported both as STRANDED - red on every fresh install. Losing a
+#              DEFAULT loses nothing; that is `MigrationReport.lost`'s own
+#              distinction, and an alarm that fires on the harmless case is one
+#              an operator learns to dismiss before the real one arrives.
+CONTROLS = {
+    "CONTROL-A": ("state-legacy.json", "roster-inverter-a.yaml"),
+    "CONTROL-B": ("state-flat-070.json", "roster.yaml"),
+    "CONTROL-C": ("state-fresh.json", "roster.yaml"),
+}
+
+
+def run(tree: Path, state: str, after: str, *extra) -> "tuple[int, str]":
+    """The harness, as the owner runs it, over `tree`.
+
+    ⚠️ ALWAYS WITH `--live-entities`. That file is the only part of the
+    entity-id comparison that does not move with the code, so a run without it
+    is a run that cannot see a renamed device.
+    """
+    fixtures = tree / "tools" / "rehearsal"
+    proc = subprocess.run(
+        [sys.executable, str(tree / "tools" / "rehearse_roster.py"),
+         "--state", str(fixtures / state),
+         "--before", str(fixtures / "flat.yaml"),
+         "--after", str(fixtures / after),
+         "--live-entities", str(fixtures / "live-entities-recorded.txt"),
+         "--reference", str(fixtures / "announced-reference.txt"),
+         *extra],
+        capture_output=True, text=True,
+        env={**os.environ, "PYTHONDONTWRITEBYTECODE": "1"})
+    return proc.returncode, proc.stdout + proc.stderr
+
+
+def mutant_tree(edits) -> str:
+    """A private copy of the add-on with `edits` applied.
+
+    ⚠️ EVERY EDIT IS ASSERTED TO HAVE MATCHED. Four defects in this session came
+    from `str.replace` calls whose pattern did not match: the call silently
+    returns the original string, the mutation never happens, and the harness
+    reports a SURVIVOR on code it never changed. That is a fabricated result.
+    """
+    tmp = tempfile.mkdtemp(prefix="rehearse-mut-")
+    dest = Path(tmp) / "tree"
+    ignore = shutil.ignore_patterns("__pycache__", "*.pyc")
+    (dest / "tools").mkdir(parents=True)
+    shutil.copytree(HERE / "app", dest / "app", ignore=ignore)
+    shutil.copytree(FIX, dest / "tools" / "rehearsal", ignore=ignore)
+    shutil.copyfile(HERE / "tools" / "rehearse_roster.py",
+                    dest / "tools" / "rehearse_roster.py")
+    for rel, old, new in edits:
+        path = dest / rel
+        text = path.read_text(encoding="utf-8")
+        hits = text.count(old)
+        if hits != 1:
+            raise AssertionError(
+                f"mutation target appears {hits} times in {rel}, expected 1: "
+                f"{old!r} - the mutation did NOT happen and any result from it "
+                f"would be fabricated")
+        path.write_text(text.replace(old, new), encoding="utf-8")
+    return str(dest)
+
+
+def legs(out: str) -> dict:
+    """The report, split into its legs on the tool's `LEG <id>` markers.
+
+    ⚠️ THIS FUNCTION IS THE FIX FOR A DEFECT IN THIS FILE, AND THE DEFECT IS
+    HABIT 4 ONE LEVEL UP. Part 5's check that the disarmed latch is "NOT THE
+    SAME-STEP PATH ALONE" was written as
+    `out.count("LATCH CAME UP DISARMED") >= 2`. The tool prints every finding
+    TWICE per red leg - once in that leg's own FINDINGS block and once in the
+    closing summary - so ONE red leg yields exactly 2 and the check passes.
+    Measured on three builds:
+
+        shipped code, both legs red                            count 4
+        fix at main.py:250 only        -> STAGED alone red     count 2  ok
+        partial guard fix on rep.ran   -> SAME-STEP alone red  count 2  ok
+
+    The check named "NOT THE SAME-STEP PATH ALONE" reported `ok` on the build
+    whose remaining failure IS the same-step path alone - and, worse, applying
+    the fix DOCS.md names as the cause left all five part-5 checks green. A
+    count over a whole report cannot distinguish one leg from two; only reading
+    the legs separately can.
+
+    ⚠️ AND IT CLOSES ON `END LEG`, WHICH IS THE SAME DEFECT ONE LEVEL DOWN.
+    The first version of this splitter opened on `LEG <id>` and never closed,
+    so every line after the LAST marker went to that leg - including the
+    closing summary, which repeats every finding from every leg. The last leg
+    then read red whenever ANY leg was red. Measured on a build whose SAME-STEP
+    leg is genuinely clean: it still reported SAME-STEP disarmed. Part 5b is
+    the running control for exactly that.
+    """
+    out_legs, current, closed = {}, None, set()
+    for line in out.splitlines():
+        if line.startswith("  LEG "):
+            # ⚠️ AND THIS ASSERTION IS THE ONE THAT KILLS THE CLASS. Review
+            # attacked part 5b directly and got through in three edits: delete
+            # `END LEG` (5b catches it), ALSO reorder the legs (5b goes blind -
+            # it only ever saw them disagree with the clean leg LAST), and ALSO
+            # reword the closing summary (the backstop is a transcription of the
+            # tool's own sentence, so rewording blinds that too). Result: "all
+            # checks passed" with a splitter that closed ZERO legs.
+            #
+            # Two of those three are ordinary refactors - reordering report legs
+            # and rewording a summary line - neither of which reads as touching
+            # a safety check. A fixture cannot cover every arrangement, so the
+            # STRUCTURE is asserted instead: a leg may not open while one is
+            # open, and every leg that opens must be explicitly closed. Under
+            # all three attacks nothing is ever closed, so this fires on the
+            # FIRST edit regardless of order or wording.
+            assert current is None, (
+                f"leg {line.strip()[4:].strip()!r} opened while {current!r} was "
+                "still open - the tool stopped closing its legs, and a splitter "
+                "that never closes swallows the closing summary into the last "
+                "leg. That is the defect this function exists to fix.")
+            current = line.strip()[4:].strip()
+            out_legs[current] = []
+        elif line.startswith("  END LEG "):
+            assert current is not None, "END LEG without an open leg"
+            closed.add(current)
+            current = None
+        elif current is not None:
+            out_legs[current].append(line)
+    assert current is None, f"leg {current!r} was never closed"
+    assert closed == set(out_legs), (
+        f"legs opened {sorted(out_legs)} but closed {sorted(closed)} - a leg "
+        "that is never closed absorbs everything after it")
+    return {k: "\n".join(v) for k, v in out_legs.items()}
+
+
+for label, (state, after) in CONTROLS.items():
+    code, out = run(HERE, state, after)
+    check(f"{label}: the shipped code, {state} -> {after}, exits 0 with no "
+          f"findings", code == RR.EXIT_OK and "NO FINDINGS" in out)
+
+# ⚠️ AND THE SOFTENING IS NOT A HOLE. The SAME key stranded with a NON-default
+# value is still a finding - that is the whole of part 5 - so the two cases are
+# separated by the VALUE and not by the key name.
+_fresh_code, _fresh_out = run(HERE, "state-fresh.json", "roster.yaml")
+# ⚠️ THE NAMESPACE IS READ OUT OF THE TOOL'S OWN REPORT, NOT SPELLED HERE. A
+# hardcoded `unit/inverter-a/...` makes this check fail for the RIGHT reason on
+# the shipped code and for the WRONG reason on a build that fixes the migration
+# to use the roster's own name - it would look like the default-softening broke
+# when what changed is the namespace. What this check is about is the VALUE.
+_stranded_default = set(re.findall(
+    r"^\s+stranded\s+(unit/\S+/sign_tripped) = False$", _fresh_out, re.M))
+check("a fresh install DOES strand a `.../sign_tripped` key - it is written "
+      "and nothing on the roster reads it - and the report SAYS SO under its "
+      f"own heading (found: {sorted(_stranded_default)})",
+      _fresh_code == RR.EXIT_OK and len(_stranded_default) == 1)
+check("...and does NOT call it a finding, because the value is the DEFAULT; "
+      "the same key at True is the finding part 5 turns on, so the two cases "
+      "are separated by the VALUE and not by the key name",
+      not any(f"STRANDED STORE KEY: {k}" in _fresh_out
+              for k in _stranded_default)
+      and "STRANDED STORE KEY: unit/inverter-a/sign_tripped = True"
+      in run(HERE, "state-legacy.json", "roster.yaml")[1])
+
+# --------------------------------------------------------------------------- #
+print()
+print("2b. ⚠️ OMITTING THE INSTALL'S ENTITY LIST IS A REFUSAL, NOT A QUIETER "
+      "RUN")
+# --------------------------------------------------------------------------- #
+# ⚠️ THIS PART EXISTS BECAUSE THE ASYMMETRY WAS BACKWARDS AND IT WAS MEASURED.
+# A `--live-entities` file that could not be READ refused with exit 2; the flag
+# simply not being PASSED produced "NO FINDINGS" and exit 0 - and
+# `grep -ci "live-entities|recorded list|composition"` over that report returned
+# ZERO, so nothing in the artifact an owner reads said half the check had not
+# run. Measured on the two entity mutations at the time:
+#
+#     M5 device rename   with the flag: exit 1     without: exit 0
+#     M6 sensor rename   with the flag: exit 1     without: exit 0
+#
+# The owner runs this ONCE, at commissioning, under time pressure. A forgotten
+# flag must not hand back a clean bill of health for the half of the transition
+# that orphans entities.
+def raw(*args):
+    """`(exit code, THE REPORT, the log)` - and the split matters.
+
+    ⚠️ THE REPORT IS STDOUT AND THE CONTROLLER'S LOGGING IS STDERR. Grading
+    "is the degraded banner near the end of the report" over the two
+    CONCATENATED reads the last lines of the LOG - `SAFETY-02 was TRIPPED...`,
+    `unparseable timestamp...` - and reports a missing banner that is in fact
+    printed. That is not a hypothetical; it is how this check first failed.
+    """
+    fx = HERE / "tools" / "rehearsal"
+    proc = subprocess.run(
+        [sys.executable, str(HERE / "tools" / "rehearse_roster.py"),
+         "--state", str(fx / "state-legacy.json"),
+         "--before", str(fx / "flat.yaml"),
+         "--after", str(fx / "roster-inverter-a.yaml"), *args],
+        capture_output=True, text=True,
+        env={**os.environ, "PYTHONDONTWRITEBYTECODE": "1"})
+    return proc.returncode, proc.stdout, proc.stderr
+
+
+_no_code, _no_out, _no_err = raw()
+check("omitting --live-entities REFUSES with exit 2 rather than reporting a "
+      "clean transition it did not check",
+      _no_code == RR.EXIT_COULD_NOT_RUN and "REFUSED" in _no_out + _no_err)
+
+_deg_code, _deg_out, _ = raw("--no-live-entities")
+check("...and the explicit opt-out RUNS, so the escape hatch exists",
+      _deg_code in (RR.EXIT_OK, RR.EXIT_FINDING))
+check("⚠️ ...carrying a standing banner in the report HEADER and again in "
+      "the closing block, because a report is skimmed from the bottom",
+      _deg_out.count("DEGRADED RUN") >= 2
+      and "DEGRADED RUN" in "\n".join(
+          _deg_out.strip().splitlines()[-4:]))
+check("...and the degraded report is not silent about what it did not do - "
+      "the words are in it, which is the grep that returned 0 before",
+      sum(w in _deg_out.lower() for w in
+          ("live-entities", "orphan check", "not graded")) == 3)
+check("contradictory flags are refused rather than one silently winning",
+      raw("--no-live-entities", "--live-entities",
+          str(FIX / "live-entities-recorded.txt"))[0]
+      == RR.EXIT_COULD_NOT_RUN)
+check("...and the banner really is at the END of the report, which is the "
+      "half of that claim a concatenated stdout+stderr could not have graded",
+      "DEGRADED RUN" in "\n".join(_deg_out.strip().splitlines()[-4:]))
+
+# ⚠️ AND THE COMPOSITION HALF IS NOT BEHIND THE FLAG AT ALL ANY MORE. The
+# reference is loaded on EVERY run, so a degraded run still catches a renamed
+# device - which is what makes `--no-live-entities` a genuinely bounded
+# concession rather than "half the tool off".
+_ref_missing = raw("--no-live-entities", "--reference", "/nonexistent.txt")
+check("a missing or malformed reference is exit 2 on every run, because it is "
+      "the only reference in the tool that does not move with the code",
+      _ref_missing[0] == RR.EXIT_COULD_NOT_RUN)
+
+# ⚠️ AND THE VANISHED CHECK CARRIES THE SAME DEFAULT/NON-DEFAULT RULE, BOUNDED
+# THE SAME WAY. `M9` below drops a NON-default key and must still be a finding;
+# without it, "vanished at the default is not a finding" could widen into
+# "vanished is not a finding" and nobody would see it.
+_m9 = mutant_tree([("app/store.py",
+                    "        value = data.get(legacy, DEFAULTS.get(legacy))\n"
+                    "        store.set(target, value)",
+                    "        value = data.get(legacy, DEFAULTS.get(legacy))\n"
+                    "        store.data.pop(legacy, None)\n"
+                    "        store.set(target, value)")])
+_m9_code, _m9_out = run(Path(_m9), "state-legacy.json", "roster-inverter-a.yaml")
+_m9_vanished = set(re.findall(r"VANISHED STORE KEY: (\S+) =", _m9_out))
+print(f"      deleting the flat key as it migrates -> exit {_m9_code}, "
+      f"vanished {sorted(_m9_vanished)}")
+check("⭐ a NON-default key that vanishes is STILL a finding - a migration "
+      "made to DELETE the flat rollback view store.py rule 2 requires it to "
+      "leave in place is caught, so the default-softening above is bounded by "
+      "its own positive",
+      _m9_code == RR.EXIT_FINDING
+      and {"last_completed", "phase_started", "last_start_attempt"}
+      <= _m9_vanished)
+check("...and `phase` is NOT among them, because its value 'idle' IS the "
+      "default - the same mutation, the same file, and the rule splitting the "
+      "two on the VALUE rather than on the key",
+      "phase" not in _m9_vanished)
+shutil.rmtree(Path(_m9).parent, ignore_errors=True)
+
+# ⚠️ AND THE INPUT FILE IS NEVER WRITTEN. The owner runs this against a copy of
+# a live state file; a tool that edited it would be a tool that had to be
+# trusted rather than one that cannot do harm.
+_before_bytes = (FIX / "state-legacy.json").read_bytes()
+run(HERE, "state-legacy.json", "roster.yaml", "--reverse", "--round-trip")
+check("the state file the harness was pointed at is BYTE-IDENTICAL afterwards, "
+      "including after the reverse and round-trip legs, which write",
+      (FIX / "state-legacy.json").read_bytes() == _before_bytes)
+
+
+# --------------------------------------------------------------------------- #
+print()
+print("3. ⭐ THE MUTATIONS: each one breaks the shipped migration or the "
+      "entity-id composition, and each must turn its control RED")
+# --------------------------------------------------------------------------- #
+MUTATIONS = [
+    # (id, control, what it breaks, edits, the finding it must produce)
+    ("M1", "CONTROL-B",
+     "the maintenance migration is handed the hardcoded `UNIT_A` instead of "
+     "the roster's own first name - MAINT-05's exact shipped regression, which "
+     "left all 19 suites green",
+     [("app/main.py",
+       "store, self.inverters[0].name, [inv.name for inv in self.inverters])",
+       "store, UNIT_A, [inv.name for inv in self.inverters])")],
+     "MAINTENANCE HISTORY LOST"),
+
+    ("M2", "CONTROL-B",
+     "the maintenance namespace is DROPPED: the migration moves no keys at all, "
+     "so a month of history stays under a flat spelling the new packs do not "
+     "read",
+     [("app/store.py",
+       "_move_flat_keys(store, unit, MAINT_UNIT_KEYS, rep)",
+       "_move_flat_keys(store, unit, (), rep)")],
+     "MAINTENANCE HISTORY LOST"),
+
+    ("M3", "CONTROL-A",
+     "the SAFETY-02 pair loses HALF of itself: `sign_reason` drops out of "
+     "`LEGACY_UNIT_KEYS`, so the latch migrates without the sentence that says "
+     "why it tripped",
+     [("app/store.py",
+       'LEGACY_UNIT_KEYS = ("sign_tripped", "sign_reason")',
+       'LEGACY_UNIT_KEYS = ("sign_tripped",)')],
+     "LATCH KEY NOT UNDER A ROSTER NAME"),
+
+    ("M4", "CONTROL-A",
+     "the READER's key composition changes while the migration's does not - "
+     "`Inverter.key()` upper-cases the unit - so the latch is written "
+     "perfectly, read back perfectly by the migration, and looked for under a "
+     "name that is not there",
+     [("app/inverter.py",
+       "return unit_key(self.name, name)",
+       "return unit_key(self.name.upper(), name)")],
+     "LATCH CAME UP DISARMED"),
+
+    ("M5", "CONTROL-A",
+     "the entity-id composition changes on the DEVICE half: the MQTT device is "
+     "renamed, which is the 'tidy it up' change `app/mqtt.py` spends thirty "
+     "lines in capitals refusing. ⚠️ THE BEFORE/AFTER DIFF IS BLIND TO THIS - "
+     "both sides are composed by the mutant, so every id moves equally and "
+     "nothing looks orphaned. Only the recorded list catches it",
+     [("app/mqtt.py", '"name": "GoodWe",', '"name": "GoodWe Controller",')],
+     "COMPOSITION DRIFT"),
+
+    ("M6", "CONTROL-A",
+     "the entity-id composition changes on the ENTITY half: one site sensor is "
+     "renamed, which orphans exactly one id and leaves the other seven alone",
+     [("app/mqtt.py",
+       '("grid", "goodwe_grid_power", "Grid power", "W", "power", "measurement", None),',
+       '("grid", "goodwe_grid_power", "Grid power W", "W", "power", "measurement", None),')],
+     "COMPOSITION DRIFT: this build announces sensor.goodwe_grid_power_w"),
+
+    # ⚠️ M7 AND M8 ARE THE TWO CHANGES THAT ORPHAN AN INSTALL WITHOUT MOVING A
+    # SINGLE COMPOSED entity_id, so an entity_id-only reference cannot see them.
+    # Both came back exit 0 before the reference carried three columns.
+    ("M7", "CONTROL-A",
+     "the `unique_id` composition changes and NOT the entity_id: `mqtt.BASE` "
+     "is renamed, so every announced unique_id moves while every composed "
+     "entity_id stays byte-identical. HA's MQTT entity registry keys on "
+     "unique_id, so this is a NEW entity beside a dead one - and discovery is "
+     "retained with no removal payload, so the dead one stays",
+     [("app/mqtt.py", 'BASE = "goodwe_ctl"', 'BASE = "goodwe_controller"')],
+     "UNIQUE_ID DRIFT"),
+
+    ("M8", "CONTROL-A",
+     "the DEVICE IDENTIFIER changes: HA's device registry keys on it, so the "
+     "whole install moves onto a new device and the old one keeps its dead "
+     "entities. Again no composed entity_id moves",
+     [("app/mqtt.py", '"identifiers": ["goodwe_rs485_controller"],',
+       '"identifiers": ["goodwe_rs485_ctl_v2"],')],
+     "DEVICE IDENTIFIER DRIFT"),
+]
+
+for mid, control, what, edits, must_say in MUTATIONS:
+    state, after = CONTROLS[control]
+    tree = mutant_tree(edits)
+    code, out = run(Path(tree), state, after)
+    red = code == RR.EXIT_FINDING
+    named = must_say in out
+    print()
+    print(f"  --- {mid} ({control}) {what}")
+    print(f"      exit {code} (0=clean 1=finding 2=could-not-run)")
+    # ⚠️ THE FINDINGS, DE-DUPLICATED AND IN ORDER. The tool prints each finding
+    # once per transition leg and once more in its summary, and its own header
+    # is full of ⚠️ paragraphs - printing every line that starts with one would
+    # bury the evidence this part exists to show under the tool's docstring.
+    seen = []
+    for line in out.splitlines():
+        text = line.strip()
+        # ⚠️ `_` IS IN THE CLASS. Without it `UNIQUE_ID DRIFT:` does not match
+        # its own finding kind, and M7 - the mutation that renames `mqtt.BASE`
+        # - printed its verdict with NO evidence under it. A filter that
+        # silently drops the very finding the mutation exists to produce is a
+        # small version of the same defect this file is about.
+        if text.startswith("⚠️ ") and re.match(r"⚠️ [A-Z][A-Z _-]+[A-Z()]:",
+                                                 text):
+            if text not in seen:
+                seen.append(text)
+    for text in seen:
+        print(f"      {text[:160]}")
+    check(f"{mid} turns {control} RED", red)
+    check(f"{mid} says WHY, and it is the right why: {must_say!r}", named)
+    shutil.rmtree(Path(tree).parent, ignore_errors=True)
+
+
+# --------------------------------------------------------------------------- #
+print()
+print("4. THE MUTATION THAT SURVIVES, ON PURPOSE, WITH THE REASON WRITTEN DOWN")
+# --------------------------------------------------------------------------- #
+# ⚠️ A BATTERY IN WHICH EVERYTHING DIES IS A BATTERY SOMEBODY STOPPED WRITING
+# AS SOON AS THE NUMBER LOOKED GOOD. `store.unit_key()` is the ONE spelling of
+# the namespaced key, read by `Inverter.key()`, by `UnitHistory.key()` and by
+# both migrations - that is store.py's stated design, "one spelling, in one
+# place". Changing it therefore moves the writer AND every reader together, and
+# the store comes out internally consistent under a different name.
+#
+# The harness is BLIND to it and cannot be otherwise: nothing inside one version
+# of this add-on can see that the key shape stopped matching what an OLDER
+# version reads. That is a ROLLBACK contract, and grading it needs two versions
+# of the code, which is a different instrument (`tools/predeploy_check.py`'s
+# territory). Stated here rather than left as an unexplained green.
+_tree = mutant_tree([("app/store.py",
+                      'return f"unit/{unit}/{key}"',
+                      'return f"ns.{unit}.{key}"')])
+_code, _out = run(Path(_tree), *CONTROLS["CONTROL-A"])
+print(f"      renaming the ONE key spelling -> exit {_code}")
+check("renaming `unit_key()` survives, because writer and readers move "
+      "together - a fact about the design, recorded, not a gap being hidden",
+      _code == RR.EXIT_OK)
+check("...and the store really did change shape underneath it, so this is a "
+      "survivor of a mutation that HAPPENED rather than of one that did not",
+      "ns.inverter-a.sign_tripped" in _out and "unit/inverter-a/" not in _out)
+shutil.rmtree(Path(_tree).parent, ignore_errors=True)
+
+# ⚠️ THE SECOND DECLARED SURVIVOR, AND IT IS THE ONE THAT BOUNDS M5-M8. Four
+# mutations of `app/mqtt.py` are now required to go red; without a mutation of
+# the same file that must NOT, "the reference notices any edit to mqtt.py"
+# would be indistinguishable from "the reference notices what HA keys on".
+# `object_id` is the case: `app/mqtt.py` lines 36-42 and 61-76 record that HA
+# 2026.8 does NOT honour it in a discovery payload - the entity_id is composed
+# from device name plus entity name - so renaming one changes nothing an
+# install can see, and a tool that refused it would refuse a free change.
+_tree2 = mutant_tree([("app/mqtt.py",
+                       '("grid", "goodwe_grid_power", "Grid power",',
+                       '("grid", "goodwe_grid_pwr", "Grid power",')])
+_code2, _out2 = run(Path(_tree2), *CONTROLS["CONTROL-A"])
+print(f"      renaming an object_id (HA does not honour it) -> exit {_code2}")
+check("renaming an `object_id` survives, because HA composes the entity_id "
+      "from the device and entity NAMES and ignores it - so the four mqtt.py "
+      "mutations above are red for what HA keys on, not for being edits to "
+      "mqtt.py",
+      _code2 == RR.EXIT_OK)
+shutil.rmtree(Path(_tree2).parent, ignore_errors=True)
+
+
+# --------------------------------------------------------------------------- #
+print()
+print("5. SAFETY-44: a pre-DEV-02 flat store meeting a DECLARED ROSTER, which "
+      "is the path nobody had established either way")
+# --------------------------------------------------------------------------- #
+# The question on the card: `migrate_unit_keys(store, UNIT_A)` takes a hardcoded
+# constant and has no `misdirected` guard, while `migrate_maint_keys(store,
+# unit, roster)` takes the roster and does. Is that asymmetry safe?
+#
+# ⚠️ MEASURED, NOT ARGUED. The two checks below are the answer, and they are
+# written as REQUIREMENTS so that a fix to `migrate_unit_keys` turns them red
+# and this file has to be updated deliberately rather than drifting.
+code44, out44 = run(HERE, "state-legacy.json", "roster.yaml")
+print(f"      flat store + roster ['garage','shed'] -> exit {code44}")
+check("⚠️ SAFETY-44 IS A DEFECT AGAIN: a store that ran pre-DEV-02 flat and "
+      "declares a multi-inverter roster comes up with its SAFETY-02 latch "
+      "DISARMED - the file says `sign_tripped: true` and not one inverter is "
+      "latched",
+      code44 == RR.EXIT_FINDING and "LATCH CAME UP DISARMED" in out44)
+check("...and the harness names the two latch keys SAFETY-44 asks about, by "
+      "the namespace they were stranded in",
+      "unit/inverter-a/sign_tripped" in out44
+      and "unit/inverter-a/sign_reason" in out44)
+check("...and the mechanism is visible in the report: the keys are stranded "
+      "under `inverter-a`, which is not on the roster, so no machine reads them",
+      "STRANDED STORE KEY" in out44 and "['garage', 'shed']" in out44)
+# ⚠️ PER LEG, NOT A COUNT OVER THE WHOLE REPORT. See `legs()` for the three
+# builds this distinction was measured on and for what the count missed.
+_L = legs(out44)
+print(f"      legs in the report: {sorted(_L)}")
+for _leg in ("STAGED", "SAME-STEP"):
+    print(f"      LEG {_leg}: "
+          f"{'DISARMED' if 'LATCH CAME UP DISARMED' in _L.get(_leg, '') else 'clean'}")
+check("the report is split into legs this check can grade SEPARATELY - the "
+      "tool prints a `LEG <id>` marker, because counting a finding's text over "
+      "the whole report cannot tell one red leg from two",
+      set(_L) == {"STAGED", "SAME-STEP"})
+check("⚠️ SAME-STEP is disarmed: the roster is declared in the SAME step as "
+      "the upgrade, the migration runs and misdirects the latch to "
+      "`inverter-a`",
+      "LATCH CAME UP DISARMED" in _L.get("SAME-STEP", ""))
+check("⚠️ AND IT IS NOT THE SAME-STEP PATH ALONE - the STAGED leg reaches the "
+      "same disarmed latch by a DIFFERENT route: the flat start stamps "
+      "`schema_version`, so the roster start's migration takes its early "
+      "return and the latch never moves at all",
+      "LATCH CAME UP DISARMED" in _L.get("STAGED", ""))
+
+# ⚠️ AND THE CONTRAST, WHICH IS WHAT STOPS THE CHECK ABOVE BEING VACUOUS. The
+# SAME state file onto a roster whose first entry IS `inverter-a` comes up
+# LATCHED. The two runs differ in one thing - the first roster name - so the
+# finding is attributable to it and to nothing else.
+code_ok, out_ok = run(HERE, "state-legacy.json", "roster-inverter-a.yaml")
+check("...and no leg's text has swallowed the CLOSING SUMMARY, which repeats "
+      "every finding from every leg - the failure that made the last leg read "
+      "red whenever any leg was",
+      not any("FINDING(S) - this transition is not clean" in v
+              for v in _L.values()))
+
+# --------------------------------------------------------------------------- #
+print()
+print("5b. THE SPLITTER'S POSITIVE CONTROL: a build in which exactly ONE leg "
+      "is clean")
+# --------------------------------------------------------------------------- #
+# ⚠️ A PER-LEG CHECK THAT HAS NEVER SEEN THE LEGS DISAGREE IS A COUNT WITH
+# EXTRA STEPS. Every fixture above has both legs red, so nothing so far
+# separates "reads each leg" from "reads the whole report twice". This part
+# builds the one case that does: `migrate_unit_keys()` handed the roster's own
+# first name instead of the hardcoded `UNIT_A`.
+#
+# ⚠️ AND WHAT IT MEASURES IS ALSO THE ANSWER TO "IS THAT THE FIX". It is NOT.
+# It repairs the SAME-STEP path only. The STAGED path is untouched, because
+# there the flat start stamps `schema_version = 1` and the roster start's
+# migration takes its `from_version >= SCHEMA_VERSION` early return before the
+# name it would be handed matters at all. A fix that stops at this line would
+# close the half an operator is least likely to take and leave the half that
+# "upgrade first, declare the roster later" produces.
+_one_leg = mutant_tree([(
+    "app/main.py",
+    "        self.store_migration = migrate_unit_keys(store, UNIT_A)",
+    "        self.store_migration = migrate_unit_keys(store, "
+    "roster_names(opts)[0])")])
+_ol_code, _ol_out = run(Path(_one_leg), "state-legacy.json", "roster.yaml")
+_OL = legs(_ol_out)
+for _leg in ("STAGED", "SAME-STEP"):
+    print(f"      LEG {_leg}: "
+          f"{'DISARMED' if 'LATCH CAME UP DISARMED' in _OL.get(_leg, '') else 'clean'}")
+check("⭐ with `migrate_unit_keys()` handed the roster's own first name, the "
+      "SAME-STEP leg comes up CLEAN - so the splitter really does read the "
+      "legs separately, and a check that only counted would have missed this",
+      "LATCH CAME UP DISARMED" not in _OL["SAME-STEP"])
+check("⚠️ ...while the STAGED leg is STILL DISARMED on the same build: that "
+      "line is not the whole fix, because the staged path never reaches it - "
+      "the first start stamps `schema_version` and the second start's "
+      "migration returns early",
+      "LATCH CAME UP DISARMED" in _OL["STAGED"])
+check("...and the run as a whole is still a finding, so nobody reads this "
+      "partial fix as done",
+      _ol_code == RR.EXIT_FINDING)
+shutil.rmtree(Path(_one_leg).parent, ignore_errors=True)
+
+print()
+print("5c. back to the shipped code")
+check("the SAME state file onto a roster whose first entry IS `inverter-a` "
+      "comes up latched and clean - so the finding above is caused by the "
+      "roster NAME and not by the fixture being tripped",
+      code_ok == RR.EXIT_OK and "LATCH CAME UP DISARMED" not in out_ok)
+
+
+# --------------------------------------------------------------------------- #
+print()
+print("6. THE REVERSE DIRECTION: the retreat, and the return from it")
+# --------------------------------------------------------------------------- #
+code_r, out_r = run(HERE, "state-flat-070.json", "roster.yaml",
+                    "--reverse", "--round-trip")
+_RL = legs(out_r)
+print(f"      legs in the reverse run: {sorted(_RL)}")
+_orphaned = set(re.findall(r"ORPHANED ENTITY: (sensor\.\S+)",
+                           _RL.get("RETREAT", "")))
+print(f"      the RETREAT leg orphans: {sorted(_orphaned)}")
+check("the RETREAT is rehearsed as its own leg, and THAT LEG reports the "
+      "entities it orphans - the eight per-unit rows, which is the cost of "
+      "backing out",
+      {"RETREAT", "ROUND-TRIP"} <= set(_RL) and _orphaned == {
+          f"sensor.goodwe_{u}_{k}"
+          for u in ("garage", "shed")
+          for k in ("setpoint", "battery_power", "battery_soc", "sign_latch")})
+check("...and NOT one site entity: the eight ids the install has today survive "
+      "the retreat, which is what makes the eight above attributable to the "
+      "per-unit rows and not to a blanket rename",
+      not any(o in _orphaned for o in
+              ("sensor.goodwe_grid_power", "sensor.goodwe_setpoint",
+               "sensor.goodwe_battery_power", "sensor.goodwe_p1_sample_age")))
+check("...and the first pack's maintenance history SURVIVES the retreat, "
+      "because `UnitHistory.mirror_flat` keeps the flat keys live - so the "
+      "retreat itself does not lose state. ⚠️ Read from the RETREAT LEG "
+      "alone, so the round trip's own loss cannot be mistaken for it",
+      "MAINTENANCE HISTORY LOST" not in _RL["RETREAT"])
+check("⚠️ ...but the ROUND TRIP does not: a maintenance cycle completed while "
+      "the install is rolled back is DISCARDED when it re-advances, because "
+      "`maint_schema_version` is already stamped and the migration skips",
+      "RETREAT-WINDOW WORK DISCARDED" in _RL["ROUND-TRIP"])
+check("...and that is a leg whose data can reach the failure it names: the "
+      "retreat window actually contains a completion, written through the "
+      "shipped `UnitHistory`, at a timestamp the starting fixture does not have",
+      RR.RETREAT_WINDOW_COMPLETION in out_r
+      and RR.RETREAT_WINDOW_COMPLETION
+      not in (FIX / "state-flat-070.json").read_text())
+
+
+print()
+if fails:
+    print(f"FAILED {len(fails)}:")
+    for f in fails:
+        print(f"  - {f}")
+    sys.exit(1)
+print("all checks passed")
@@ -0,0 +1,48 @@
+# WHAT THIS BUILD IS SUPPOSED TO ANNOUNCE, recorded so that a change to the
+# composition has something to disagree with. THREE COLUMNS:
+#
+#     <entity_id>   <unique_id>   <device identifier>
+#
+# ⚠️ ALL THREE, AND THE SECOND AND THIRD ARE THE HALF THE FIRST DRAFT MISSED.
+# A reference that carries only entity_ids cannot see the two changes that
+# orphan an install without changing a single composed id:
+#
+#   `BASE` renamed (`goodwe_ctl` -> anything)      every unique_id changes
+#   `DEVICE["identifiers"]` changed                HA sees a NEW device
+#
+# Home Assistant's MQTT integration keys its entity registry on `unique_id` and
+# its device registry on the device identifier. Change either and HA creates a
+# NEW entity beside the old one - and because discovery is published with
+# `retain=True` and nothing publishes a removal payload, the OLD config stays on
+# the broker and the dead entity stays in the registry, still named by every
+# automation that referenced it, never updating again. The composed entity_id is
+# identical on both, so an entity_id-only comparison reports a clean transition.
+# That is the same shape as the habit-4 this file's own card already caught
+# once: a reference that does not carry the thing that moved.
+#
+# ⚠️ AND `object_id` IS DELIBERATELY NOT A COLUMN. `app/mqtt.py` lines 36-42 and
+# 61-76 record that HA 2026.8 does NOT honour `object_id` in a discovery
+# payload - the id is composed from device name plus entity name - so renaming
+# one is a true no-op and a column for it would make this file refuse a change
+# that costs nothing. The check is graded on what HA actually keys on.
+#
+# PROVENANCE, stated exactly:
+#   - recorded 2026-09-08 from `release/1.0` at `meter_source: firmware`,
+#     roster of one, by reading `mqtt.discovery_payloads(units=[])`. It is what
+#     the code announces.
+#   - `sensor.goodwe_p1_sample_age` is additionally recorded in `app/mqtt.py` as
+#     VERIFIED AGAINST THE LIVE SYSTEM on 2026-08-26, and the ESP32's ingest
+#     guard subscribes to that exact string.
+#   - ⚠️ IT IS NOT A DUMP OF THE LIVE INSTALL'S ENTITY REGISTRY. That is what
+#     `--live-entities` is for. This file is about THE BUILD.
+#
+# Editing this file is how you say "the announcement changed on purpose". Do not
+# edit it to make a check pass.
+sensor.goodwe_setpoint            goodwe_ctl_setpoint     goodwe_rs485_controller
+sensor.goodwe_grid_power          goodwe_ctl_grid         goodwe_rs485_controller
+sensor.goodwe_battery_power       goodwe_ctl_battery      goodwe_rs485_controller
+sensor.goodwe_battery_soc         goodwe_ctl_soc          goodwe_rs485_controller
+sensor.goodwe_circulating_power   goodwe_ctl_circulation  goodwe_rs485_controller
+sensor.goodwe_maintenance_phase   goodwe_ctl_phase        goodwe_rs485_controller
+sensor.goodwe_controller_status   goodwe_ctl_status       goodwe_rs485_controller
+sensor.goodwe_p1_sample_age       goodwe_ctl_p1_age       goodwe_rs485_controller
@@ -0,0 +1,22 @@
+# ARCH-02 fixture: THE SHAPE THE LIVE BOX IS IN TODAY (a flat, single-inverter
+# config). ⚠️ ILLUSTRATIVE ENTITY NAMES, NOT THE LIVE ONES - the owner runs the
+# rehearsal against a copy of the real options blob, and this file exists so
+# that the harness has a synthetic transition it can be graded on.
+setpoint_entity: number.goodwe_setpoint
+soc_entity: sensor.goodwe_battery_soc_pct
+batt_entity: sensor.goodwe_battery_power_w
+batt_invert: false
+max_w: 5000
+bms_soc_floor: 10
+meter_source: 'off'
+meter_entity: sensor.p1_active_power
+inverters: []
+gain: 0.6
+slew_w: 0
+deadband_w: 15
+target_grid_w: 0
+step_w: 10
+heartbeat_s: 10
+stale_input_s: 15
+maintenance_discharge_w: 2000
+maintenance_charge_w: 2000
@@ -0,0 +1,39 @@
+# THE ENTITY IDS THE INSTALL HAS — a stand-in for the owner's real export.
+#
+# ⚠️ THIS FILE AND `announced-reference.txt` ANSWER DIFFERENT QUESTIONS AND ARE
+# NOT INTERCHANGEABLE:
+#
+#   announced-reference.txt   what the BUILD announces. Loaded on EVERY run.
+#                             Three columns, because HA keys on `unique_id` and
+#                             on the device identifier as well as on the
+#                             composed entity_id.
+#   this file (--live-entities)
+#                             what the INSTALL has. One id per line. Orphans
+#                             are graded against it: an id the install has and
+#                             the AFTER shape does not announce is an orphan
+#                             whatever the build's own history says.
+#
+# ⚠️ OMITTING `--live-entities` IS A REFUSAL (exit 2), NOT A QUIETER RUN. It
+# used to be the other way round - an unreadable file refused, a missing flag
+# passed with exit 0 and said nothing about the half it had not checked - and
+# the owner runs this ONCE, at commissioning, under time pressure.
+# `--no-live-entities` is the deliberate opt-out and it prints a standing
+# banner in the report header and again on the exit line.
+#
+# PROVENANCE: these eight are what `release/1.0` announces at
+# `meter_source: firmware`, roster of one, recorded 2026-09-08. They stand in
+# for the export so the checks run in CI. ⚠️ THE REAL INSTALL CARRIES ROUGHLY 49
+# `goodwe*` ENTITIES (card ARCH-02, read 2026-09-06); the other ~41 belong to
+# the ESPHome firmware, to template sensors and to helpers, and this add-on
+# announces none of them. The tool separates the ids it owns from the ids it
+# does not rather than claiming orphans it cannot cause - which is the half of
+# this file the eight-line stand-in cannot exercise, and the reason DOCS.md's
+# owner step asks for the real export.
+sensor.goodwe_battery_power
+sensor.goodwe_battery_soc
+sensor.goodwe_circulating_power
+sensor.goodwe_controller_status
+sensor.goodwe_grid_power
+sensor.goodwe_maintenance_phase
+sensor.goodwe_p1_sample_age
+sensor.goodwe_setpoint
@@ -0,0 +1,42 @@
+# ARCH-02 fixture: THE SHAPE AFTER INVERTER B IS COMMISSIONED.
+#
+# ⚠️ THE FIRST ROSTER ENTRY IS **NOT** CALLED `inverter-a`, AND THAT IS THE
+# WHOLE FIXTURE. `MUTATION.md` habit 4's seventh instance is MAINT-05 shipping
+# `migrate_maint_keys(store, UNIT_A, ...)` under a comment saying it must not:
+# every two-pack fixture in the suite named its first entry `inverter-a`, so
+# `UNIT_A` and `inverters[0].name` were the same string in every check that
+# existed and the defect was invisible. Names that are not `inverter-a` are what
+# make a misdirected namespace something a rehearsal can SEE.
+#
+# `roster-inverter-a.yaml` beside this file is the same roster with the first
+# entry named `inverter-a`, so the two can be compared - see DOCS.md.
+setpoint_entity: number.goodwe_setpoint
+soc_entity: sensor.goodwe_battery_soc_pct
+batt_entity: sensor.goodwe_battery_power_w
+batt_invert: false
+max_w: 5000
+bms_soc_floor: 10
+meter_source: 'off'
+meter_entity: sensor.p1_active_power
+gain: 0.6
+slew_w: 0
+deadband_w: 15
+target_grid_w: 0
+step_w: 10
+heartbeat_s: 10
+stale_input_s: 15
+maintenance_discharge_w: 2000
+maintenance_charge_w: 2000
+inverters:
+  - name: inverter-a
+    setpoint_entity: number.goodwe_setpoint
+    soc_entity: sensor.goodwe_battery_soc_pct
+    batt_entity: sensor.goodwe_battery_power_w
+    max_w: 5000
+    bms_soc_floor: 10
+  - name: inverter-b
+    setpoint_entity: number.goodwe_b_setpoint
+    soc_entity: sensor.goodwe_b_battery_soc_pct
+    batt_entity: sensor.goodwe_b_battery_power_w
+    max_w: 4600
+    bms_soc_floor: 10
@@ -0,0 +1,42 @@
+# ARCH-02 fixture: THE SHAPE AFTER INVERTER B IS COMMISSIONED.
+#
+# ⚠️ THE FIRST ROSTER ENTRY IS **NOT** CALLED `inverter-a`, AND THAT IS THE
+# WHOLE FIXTURE. `MUTATION.md` habit 4's seventh instance is MAINT-05 shipping
+# `migrate_maint_keys(store, UNIT_A, ...)` under a comment saying it must not:
+# every two-pack fixture in the suite named its first entry `inverter-a`, so
+# `UNIT_A` and `inverters[0].name` were the same string in every check that
+# existed and the defect was invisible. Names that are not `inverter-a` are what
+# make a misdirected namespace something a rehearsal can SEE.
+#
+# `roster-inverter-a.yaml` beside this file is the same roster with the first
+# entry named `inverter-a`, so the two can be compared - see DOCS.md.
+setpoint_entity: number.goodwe_setpoint
+soc_entity: sensor.goodwe_battery_soc_pct
+batt_entity: sensor.goodwe_battery_power_w
+batt_invert: false
+max_w: 5000
+bms_soc_floor: 10
+meter_source: 'off'
+meter_entity: sensor.p1_active_power
+gain: 0.6
+slew_w: 0
+deadband_w: 15
+target_grid_w: 0
+step_w: 10
+heartbeat_s: 10
+stale_input_s: 15
+maintenance_discharge_w: 2000
+maintenance_charge_w: 2000
+inverters:
+  - name: garage
+    setpoint_entity: number.goodwe_setpoint
+    soc_entity: sensor.goodwe_battery_soc_pct
+    batt_entity: sensor.goodwe_battery_power_w
+    max_w: 5000
+    bms_soc_floor: 10
+  - name: shed
+    setpoint_entity: number.goodwe_b_setpoint
+    soc_entity: sensor.goodwe_b_battery_soc_pct
+    batt_entity: sensor.goodwe_b_battery_power_w
+    max_w: 4600
+    bms_soc_floor: 10
@@ -0,0 +1,11 @@
+{
+ "_comment": "ARCH-02 SYNTHETIC FIXTURE - a store that has ALREADY run the flat config under 0.7.0+, so DEV-02's `schema_version` is stamped and the SAFETY-02 latch is not tripped. NOT a copy of the live box's /data/state.json. This is the CLEAN CONTROL for the MAINT-05 half of the rehearsal: the flat->roster transition over it must come out with NO findings, which is what makes each deliberate mutation's RED mean something. A harness that is red on every input has not measured anything.",
+ "schema_version": 1,
+ "phase": "idle",
+ "phase_started": "2026-08-07T10:00:00+00:00",
+ "last_completed": "2026-08-07T18:42:00+00:00",
+ "last_start_attempt": "2026-08-07T10:00:00+00:00",
+ "sign_tripped": false,
+ "sign_reason": "",
+ "auto": true
+}
@@ -0,0 +1,3 @@
+{
+ "_comment": "ARCH-02 SYNTHETIC FIXTURE - a FRESH INSTALL. No maintenance history, no latch, no schema markers: what /data/state.json looks like before the add-on has ever completed a cycle. It is a clean control, and it earned its place: the first draft of the harness reported this transition RED, because the DEV-02 migration writes both SAFETY-02 defaults under `inverter-a` before a roster exists and the tool called them stranded. Losing a default loses nothing."
+}
@@ -0,0 +1,10 @@
+{
+ "_comment": "ARCH-02 SYNTHETIC FIXTURE - a PRE-DEV-02 flat state file. NOT a copy of the live box's /data/state.json; this session cannot read that one and did not try. It is built to carry, deliberately, the two shapes that state built by the happy path does not have: a LATCHED SAFETY-02 failsafe under the flat spelling, and a real maintenance history under the flat spelling. Both are LEGACY-SHAPED: there is no schema_version and no maint_schema_version, so this file was last written by a controller that had never heard of unit namespaces. Losing either one is silent.",
+ "phase": "idle",
+ "phase_started": "2026-08-07T10:00:00+00:00",
+ "last_completed": "2026-08-07T18:42:00+00:00",
+ "last_start_attempt": "2026-08-07T10:00:00+00:00",
+ "sign_tripped": true,
+ "sign_reason": "SIGN_DISAGREEMENT: commanded -1800 W against a meter reading -1450 W for 190 s",
+ "auto": true
+}
@@ -315,6 +315,114 @@ print(f"  the Configuration tab renders {len(ui)} fields, in `schema:` order "
      "(App.schema_ui is UiOptions(...)(data['schema']), so the ORDER IS THE "
      "SCHEMA BLOCK'S and not the options block's)")

+# --------------------------------------------------------------------------- #
+# ⚠️ UX-04: THAT COUNT IS NOW GRADED, AND IT IS WHY UX-01's AC 3 IS WITHDRAWN.
+#
+# UX-01 AC 3 said the visible count would drop to ~15, with the rest behind
+# HA's "show unused optional configuration options" toggle. The line above has
+# printed the refutation since the day it was written - 71, not 15 - and PRINTED
+# IS NOT GRADED: this probe exited 0 through the whole of UX-01's review while
+# displaying the number that contradicted the criterion it was evidence for.
+# That is the defect, not the count.
+#
+# THE FRONTEND RULE, transcribed from the bundle in THIS image at the top of
+# this section and replayed rather than paraphrased:
+#
+#     _filteredSchema = (opts, schema) => schema.filter(e => e.name in opts
+#                                                            || e.required)
+#
+# `opts` is the add-on's OPTIONS AS SUPERVISOR SERVES THEM - the operator's
+# stored blob after `options:` has been deep-merged into it - so a key is
+# hideable only when it is BOTH absent from `options:` AND not required.
+#
+# WHY THE CRITERION IS WITHDRAWN RATHER THAN IMPLEMENTED, in two measurements:
+#
+#   1. IT IS ARITHMETICALLY UNREACHABLE. Supervisor's own UiOptions marks
+#      `required` on every non-`?` key, and the filter above shows a required
+#      key UNCONDITIONALLY. So the floor - what the tab renders with `options:`
+#      EMPTIED ALTOGETHER, which is the most hiding that is even expressible -
+#      is the required count below. It is nowhere near 15. No arrangement of
+#      `options:` can reach the target; the criterion was never satisfiable.
+#
+#   2. THE 12 KEYS THAT *COULD* BE HIDDEN COST THE DEEP-MERGE DEFAULT. A key
+#      absent from `options:` has nothing for the Supervisor to merge into an
+#      existing box's stored blob. That is exactly the `delivery_hold_s`
+#      mechanism the 0.9.0 upgrade rests on: a key that 0.7.0's stored blob
+#      cannot contain resolves to its `options:` default instead of failing
+#      validation, and moving it out of `options:` is precisely how you get the
+#      exit 78 that mechanism exists to prevent. Worse, it would promote the 27
+#      unpinned in-code literal fallbacks that `test_defaults_pin.py` part 4
+#      ledgers into the LOAD-BEARING source of those defaults on a live install
+#      - which is the CFG-03 defect the same card's AC 4 exists to police.
+#
+# So: four fewer clicks, bought with the upgrade path and with the defaults
+# pin. The decision is written up in
+# DECISIONS/UX-04-the-configuration-tab-and-the-defaults-pin.md
+# and the three assertions below are what keeps it a decision rather than an
+# omission - if anyone empties `options:` to shorten the form, this goes red.
+_hideable = sorted(set(schema) - set(defaults))
+_required = [n["name"] for n in ui if n.get("required")]
+
+
+def _filtered(opts):
+    """The frontend's own `_filteredSchema`, replayed on this manifest."""
+    return [n for n in ui if n["name"] in opts or n.get("required")]
+
+
+show(f"every one of the {len(schema)} `schema:` keys is ALSO in `options:`, so "
+     f"schema-only - the only set HA's toggle can hide - is EMPTY. UX-01 AC 3 "
+     f"is not partially met, it is unstarted. Hideable: {_hideable}",
+     not _hideable and len(schema) == len(defaults))
+show(f"...so replaying the frontend's own _filteredSchema on the served options "
+     f"hides NOTHING: {len(_filtered(defaults))} of {len(ui)} fields render",
+     len(_filtered(defaults)) == len(ui))
+# ⚠️ DERIVED, NOT TYPED, AND NOT A FLOOR. The assertion above read
+# `len(_required) > 15` while the prose said 59 - so a manifest change dropping
+# the required count to 20 would have passed while the sentence beside it stayed
+# wrong. That is the same shape as the count this repo already fixed in
+# `test_translations.py`: a number displayed but not graded.
+#
+# So the expected value is computed from config.yaml by Supervisor's OWN rule -
+# a key is optional iff its schema string ends in `?`, plus `inverters`, which
+# `UiOptions` marks optional because it is a list - and the assertion is an
+# EQUALITY against that. Typing 59 here would rot the first time a key is added.
+_OPTIONAL = {k for k, v in schema.items()
+             if isinstance(v, str) and v.rstrip().endswith("?")} | {"inverters"}
+_EXPECT_REQ = len(schema) - len(_OPTIONAL)
+show(f"...and that floor is DERIVED: {len(schema)} schema keys minus "
+     f"{len(_OPTIONAL)} optional ones (the `?` suffix, plus `inverters`) is "
+     f"{_EXPECT_REQ}, and Supervisor's own required flag agrees",
+     _EXPECT_REQ == len(_required))
+
+# ⚠️ AND THE WITHDRAWAL'S OWN PREMISE, GRADED SEPARATELY. The equality above is
+# deliberately NOT enough on its own: both sides read the same `?` convention,
+# so marking a key optional moves BOTH and it stays green. That is correct - it
+# makes the printed number impossible to leave stale - but it grades Supervisor
+# agreeing with us, not the thing AC 3 was withdrawn for.
+#
+# AC 3 was withdrawn because the floor is FAR above the ~15 it asked for. If
+# enough keys ever became optional that the floor approached 15, the withdrawal
+# stops being true and this file should say so rather than keep printing a
+# sentence that has quietly become wrong.
+_AC3_ASKED = 15
+show(f"...and the floor of {_EXPECT_REQ} is still far above the ~{_AC3_ASKED} "
+     f"AC 3 asked for, which is the whole reason it was withdrawn. If this ever "
+     f"goes red the withdrawal needs re-arguing, not re-asserting",
+     _EXPECT_REQ > 2 * _AC3_ASKED)
+
+
+# ⚠️ AN ISOLATED COUNTERFACTUAL, and the number is Supervisor's own `required`
+# flag rather than a count typed here. `options:` is emptied and NOTHING else
+# varies, so the answer cannot be about anything but the merge.
+show(f"...and with `options:` EMPTIED ENTIRELY - the maximum hiding that is "
+     f"expressible - the tab still renders {len(_filtered({}))} required "
+     f"fields. THAT is the floor, and AC 3 asked for ~15. The criterion is "
+     f"withdrawn as unsatisfiable, not deferred.",
+     len(_filtered({})) == len(_required) and len(_required) == _EXPECT_REQ)
+
+
+
+

 def _entry(lang, name, path=()):
    node = TRANSLATIONS.get(lang, {}).get("configuration")
@@ -239,7 +239,7 @@ configuration:
    description: "TWO OR MORE INVERTERS ONLY. DERIVED (SAFETY-31): 75 s is fifteen control cycles - 15 times the worst legitimate episode the rig can produce, 2.5 times the firmware's own 30 s link veto, and twelve times inside the 900 s billing quarter. It is deliberately not a round 60 or 120, which would be a guess wearing a measurement's clothes. Do not change it without re-deriving it."
  delivery_hold_s:
    name: "Safety · Not-delivering report after (s) — derived, do not change"
-    description: "TWO OR MORE INVERTERS ONLY; with one inverter this check cannot fire at all, and that is deliberate rather than a limitation. How long ONE inverter may stand at a discharge share it is not producing - short by more than the saturation threshold, with nothing this add-on can read to explain it - while at least one sibling is NOT in that state, before it is named. REPORTS ONLY (SAFETY-42): nothing is commanded differently, the loop is untouched, and the site goes on running short by that unit's share until you look. DERIVED FROM 2640 CLOSED-LOOP TEST RUNS in which nothing was faulty (SAFETY-42, plus 1890 runs at one inverter as the control, which report nothing at any threshold): the worst legitimate episode those runs produce is 30.0 s, the floor of 30 s is the firmware's own link veto below which a unit is already being dropped from the split, and the ceiling is the 900 s quarter that bills the capacity tariff. 185 s is 6.2 times that, and the margin is sized on a measurement rather than a feeling: the second inverter in those tests is a copy of the first, and a second unit four times slower produces 90.0 s. Do not change it without re-deriving it. A pack at or below its own bms_soc_floor is never reported: an empty pack and a wedged one are identical watt for watt, and that setting is the only thing that tells them apart - so if this fires on a pack you believe is empty, that setting does not describe your battery."
+    description: "TWO OR MORE INVERTERS ONLY; with one inverter this check cannot fire at all, and that is deliberate rather than a limitation. How long ONE inverter may stand at a discharge share it is not producing - short by more than the saturation threshold, with nothing this add-on can read to explain it - while at least one sibling is NOT in that state, before it is named. REPORTS ONLY (SAFETY-42): nothing is commanded differently, the loop is untouched, and the site goes on running short by that unit's share until you look. DERIVED FROM 2640 CLOSED-LOOP TEST RUNS in which nothing was faulty (SAFETY-42, plus a one-inverter arm as the control, which reports nothing at any threshold): the worst legitimate episode those runs produce is 35.0 s, and the ceiling is the 900 s quarter that bills the capacity tariff. 185 s is 5.29 times that. ⚠️ The schema lets you set this as low as 30 s, which is BELOW that measured 35.0 s - at 30 s you should expect reports on legitimate saturation. The second inverter in those tests is a COPY OF THE FIRST, because the second inverter does not exist yet: what a unit four times slower would produce (75.0 s) is a sensitivity study on that idea, not a measurement of any real second inverter. Do not change it without re-deriving it - and it can be re-derived now: admin/goodwe sim/delivery_window.py --campaign --blag (TEST-21). A pack at or below its own bms_soc_floor is never reported: an empty pack and a wedged one are identical watt for watt, and that setting is the only thing that tells them apart - so if this fires on a pack you believe is empty, that setting does not describe your battery."
  auto_start:
    name: "Advanced · Start controlling automatically on boot"
    description: "Off by default. Leave it off until the site has been commissioned: with it off, somebody has to press Start after every restart of the add-on, which is what you want while you are still checking signs and entities."
@@ -219,7 +219,7 @@ configuration:
    description: "ALLEEN BIJ TWEE OF MEER OMVORMERS. AFGELEID (SAFETY-31): 75 s zijn vijftien regelcycli - 15 keer de langste terechte episode die de testopstelling kan produceren, 2,5 keer het eigen veto van 30 s in de firmware, en twaalf keer binnen het factuurkwartier van 900 s. Het is bewust geen ronde 60 of 120, want dat zou een gok zijn in het jasje van een meting. Verander dit niet zonder de afleiding opnieuw te maken."
  delivery_hold_s:
    name: "Veiligheid · Melding niet-leverende omvormer na (s) — afgeleid, niet wijzigen"
-    description: "ALLEEN BIJ TWEE OF MEER OMVORMERS; bij één omvormer kan deze controle helemaal niet aanslaan, en dat is een bewuste keuze en geen tekortkoming. Hoe lang één omvormer mag blijven staan op een ontlaadaandeel dat hij niet levert - meer dan de verzadigingsdrempel te weinig, zonder dat deze add-on er een verklaring voor kan uitlezen - terwijl minstens één andere omvormer dat NIET doet, voordat hij bij naam wordt gemeld. ALLEEN MELDEN (SAFETY-42): er wordt niets anders gestuurd, de regeling blijft ongemoeid, en de installatie blijft draaien met het aandeel van die omvormer te weinig tot u gaat kijken. AFGELEID UIT 2640 GESLOTEN-LUSTESTS waarbij niets defect was (SAFETY-42, plus 1890 tests met een omvormer als controlegroep, die bij geen enkele drempel iets melden): de langste terechte episode in die tests duurt 30.0 s, de ondergrens van 30 s is het eigen veto van de firmware waaronder een omvormer al uit de verdeling wordt gehaald, en de bovengrens is het factuurkwartier van 900 s. 185 s is 6,2 keer zoveel, en die marge is op een meting gebaseerd en niet op een gevoel: de tweede omvormer in die tests is een kopie van de eerste, en een tweede eenheid die vier keer trager is levert 90,0 s op. Verander dit niet zonder de afleiding opnieuw te maken. Een accu op of onder zijn eigen bms_soc_floor wordt nooit gemeld: een lege accu en een vastgelopen omvormer zijn watt voor watt identiek, en die instelling is het enige wat ze uit elkaar houdt - slaat dit toch aan bij een accu die volgens u leeg is, dan beschrijft die instelling uw batterij niet."
+    description: "ALLEEN BIJ TWEE OF MEER OMVORMERS; bij één omvormer kan deze controle helemaal niet aanslaan, en dat is een bewuste keuze en geen tekortkoming. Hoe lang één omvormer mag blijven staan op een ontlaadaandeel dat hij niet levert - meer dan de verzadigingsdrempel te weinig, zonder dat deze add-on er een verklaring voor kan uitlezen - terwijl minstens één andere omvormer dat NIET doet, voordat hij bij naam wordt gemeld. ALLEEN MELDEN (SAFETY-42): er wordt niets anders gestuurd, de regeling blijft ongemoeid, en de installatie blijft draaien met het aandeel van die omvormer te weinig tot u gaat kijken. AFGELEID UIT 2640 GESLOTEN-LUSTESTS waarbij niets defect was (SAFETY-42, plus een arm met één omvormer als controlegroep, die bij geen enkele drempel iets meldt): de langste terechte episode in die tests duurt 35,0 s, en de bovengrens is het factuurkwartier van 900 s. 185 s is 5,29 keer zoveel. ⚠️ Het schema laat u zo laag als 30 s instellen, en dat is ONDER die gemeten 35,0 s - op 30 s moet u meldingen bij terechte verzadiging verwachten. De tweede omvormer in die tests is een KOPIE VAN DE EERSTE, want de tweede omvormer bestaat nog niet: wat een vier keer tragere eenheid zou opleveren (75,0 s) is een gevoeligheidsstudie op dat idee en geen meting aan een echte tweede omvormer. Verander dit niet zonder de afleiding opnieuw te maken - en dat kan nu: admin/goodwe sim/delivery_window.py --campaign --blag (TEST-21). Een accu op of onder zijn eigen bms_soc_floor wordt nooit gemeld: een lege accu en een vastgelopen omvormer zijn watt voor watt identiek, en die instelling is het enige wat ze uit elkaar houdt - slaat dit toch aan bij een accu die volgens u leeg is, dan beschrijft die instelling uw batterij niet."
  auto_start:
    name: "Geavanceerd · Automatisch beginnen te regelen bij opstarten"
    description: "Staat standaard uit. Laat het uit tot de installatie in dienst is genomen: zolang het uit staat, moet iemand na elke herstart van de add-on op Start drukken, en dat is precies wat u wilt terwijl u tekens en entiteiten nog aan het nakijken bent."