goodwe-addon/goodwe_controller/CHANGELOG.md

# Changelog

## 0.7.0

⚠️ **SCHEMA CHANGE, AND IT IS WHY THIS IS A MINOR BUMP RATHER THAN A PATCH.**
[[SAFETY-32]] adds `bms_soc_floor: int(0,50)` as a **required** per-entry key on
the `inverters` roster - no trailing `?`. A config that declares a roster and
omits it on any entry is REFUSED at startup. An install running `inverters: []`,
which is every single-inverter install including the reference one, is
unaffected: the flat `bms_soc_floor` still describes the one pack.

CFG-04 recorded the rule this follows - the Supervisor keys the built image by
`version:`, so a schema change without a bump is a schema nobody reads.

Contents, in merge order since 0.6.0 (`2d025ee`):

| | |
|---|---|
| `3934ad7` | MAINT-05 - one recalibration per pack, migration graded against the roster |
| `67e649a` | SPIKE-04a - inverter B and pack B recorded, from datasheet and serial |
| `3993f9f` | SAFETY-32 - per-pack `bms_soc_floor`, rule 8 per unit, package-wide roster pin |
| `f72c48d` | UX-01 - add-on configuration in English and Dutch |
| `86eaef2` | TEL-07 - a second MQTT device for inverter B; A's identity byte-identical |
| `4482cd8` | SAFETY-41 - the clamp alarm stops narrating normal anti-windup |

⚠️ **0.6.0 WAS NEVER SECTIONED.** Its content sits below this heading rather than
under one of its own, because the bump shipped without cutting the section. That
is a defect in this file, not in the release: the deployed 0.6.0 is `2d025ee` and
the table above is what is new since it. Splitting it retroactively would mean
guessing a boundary; saying so is more useful than a guess.

## Unreleased

⚠️ **THE OUTPUT FREEZE NOW HAS AN EXIT, AND ONLY WHERE IT CANNOT END ON ITS OWN
([[SAFETY-13]]).** While the loop is frozen the command's magnitude may fall but
never rise; at `prev_w == 0` there is no magnitude left to fall, so a charge
cannot start. That refusal is deliberate - the inverter reporting power we did
not command means another writer is on the actuator, the §1.5 non-goal - and it
is kept. What it did not have was an END: nothing cleared it but the other
writer leaving.

The exit is now explicit, and *which* freezes get one is the decision rather
than *how long*:

- A divergence a **limit** could have produced - the inverter doing less than
  the command, in the command's own direction - **clears itself**, because the
  command is pinned at something the inverter can reach. It gets no exit, and
  that is release/1.0's behaviour kept unchanged.
- A divergence a limit could **not** have produced - the inverter overshooting
  the command, or producing the opposite sign - cannot clear itself, because
  clearing it needs the inverter to track *our* command. That one releases once
  the divergence has outlasted `PLANT_SETTLE_S`, i.e. `SETTLE_CYCLES` = 9 cycles
  at the measured tick. No new constant: it is the reciprocal of SAFETY-39's
  `SLEW_RAIL_FRAC`.

⚠️ **RELEASING THE OUTPUT FREEZE DOES NOT UNFREEZE THE LOOP.** The integrator
freeze is untouched, so the command can only walk, one slew step per cycle,
toward an accumulator that stopped accumulating when the freeze engaged.
Measured at the live install's shape (`max_w 5000`, `slew_w 400`,
`target_grid_w -10`) with another writer holding the pack at +700 W and 2 kW of
export standing: the command holds at 0 W, starts on the ninth saturated cycle
and settles at **-1190 W** against a frozen integrator of -1194 W, not at the
5000 W rail. On `release/1.0` @ `c8dcb38` the same run held **0.0 W on every
cycle**, for as long as the other writer stayed.

⚠️ **THIS IS NOT A REGRESSION FIX AND IT ESCALATES AT TWO INVERTERS.** The
unbounded block is `release/1.0` behaviour, confirmed cycle for cycle against
SAFETY-04's branch. At a roster of two, `actual_w` is the SITE SUM, so the
comparison is a site command against two units' power and one unit still moving
after the other has stopped can put the site into the block with no external
actor at all.

Graded against `release/1.0` over the 3024-case sweep in `test_control.py`:
**2592 cases bit-identical, 432 releases, and 27 of those change the command** -
every one of them a charge from `prev_w == 0` that `release/1.0` refused to
start. `saturation_cycles` at 9 or 10 removes the block entirely; DOCS.md says
so.

⚠️ **A PER-INVERTER POWER READING THAT HAS STOPPED FOLLOWING ITS OWN COMMAND IS
NOW `None`, NOT A HELD NUMBER ([[SAFETY-37]]).** `Inverter.batt` (register
`0x0518`) goes through a freshness gate: unchanged for >= 35 s *while* the
standing command is over 250 W *and* has travelled over 250 W since the reading
last moved. A stale reading is a **missing input**, which §11.1 rule 1 already
answers with 0 W and keeps commanding.

**This is a behaviour change at the shipped roster of one, and it is deliberate.**
Before it, a frozen power register read as a healthy measurement for ever - which
is [[HW-01]], where a value held for 6 h 25 m while the pack drained to its floor
and every detector read healthy. A healthy install is byte-identical: sixty reads
of a dithering register under a reversing command all reach the consumers
unchanged, asserted rather than assumed. A false verdict fails toward 0 W, behind
the existing `stale_input_s` hold, and clears itself the moment the register
reports a different value.

At two or more inverters it also removes two false alarms that had no gate at
all: [[SAFETY-30]]'s per-unit sign check no longer trips on a frozen reading
(measured: it trips without the gate and NEVER with it - the tick index is
deliberately not quoted, because it is counted on the wall clock inside
`cycle()` and is therefore load-dependent; the checkable statement is the
presence or absence of the trip, which part 3 grades), and
[[SAFETY-31]]'s circulating-power sensor reports `None` instead of a loop that is
not happening.

**No new configuration key, and no schema change.** The bound is derived from
`circulation_hold_s` and `unit_sign_disagree_*`, which the checks read out of
`config.yaml` at both ends - so moving either default takes the gate red rather
than leaving it silently too slow. `DECISIONS/SAFETY-37-freshness-policy.md`
carries the derivation, what [[TEL-06]] and [[TEL-08]] still owe, and three
named survivors - including the largest: **this detects a value that STOPPED, not
a value that is WRONG.**

⚠️ **`slew_w` IS A RATIO TO THE RAIL NOW, AND ITS DEFAULT CHANGED FROM 1000 TO 0
([[SAFETY-39]]).** `0` does not mean "off" and does not mean "unset" - it means
**follow `max_w`**, exactly as `integrator_max_w: 0` already does, and resolves
to `max_w / 9`: 222 W at the shipped 2000 W cap, 556 W at the live install's
5000, 1067 W at a ~9600 W two-inverter site. The 9 is
`PLANT_SETTLE_S / LOOP_TICK_S` - an 18 s measured step response against a 2 s
meter tick - so the loop takes the plant's own settling time to cross its whole
output range at any rail.

**An install that has never set `slew_w` will therefore command smaller steps
than it did on 0.7.0, and will converge more slowly after a large disturbance.**
That is the point rather than a side effect: `slew_w: 1000` was derived against
a **1500 W** clamp in 2026-08 and the live install has run it against 5000 since,
which crosses the whole range in 10 s against an 18 s plant. Measured on the
owner's system 2026-09-05: a 3.4 kW export drove the setpoint 380 -> 5000 W in
five ticks, the battery to -5154 W past a 5000 VA rating, and the grid from
3.4 kW of **export** to 2.8 kW of **import**.

An explicit `slew_w` is still accepted and is never silently reduced. What is
new is that a value **faster than the rail allows** is now **REFUSED AT STARTUP**
(`validate.py` rule 10), graded against the ROSTER SUM - so `slew_w: 1000` with
`max_w: 5000` will not start. Set it to `0`, or to any value at or below
`max_w / 9`. Slower is always accepted, including the interim 400 the reference
install has run since 2026-09-05.

⚠️ **This does NOT fix the over-absorption, and the record says so at length.**
Re-deriving the slew removes the >= 900 W setpoint step; the loop still absorbs
~1 kW more than the site needs on a large export step, because that is a `gain`
term and `gain: 0.6` was measured and deliberately left alone. See
`DECISIONS/SAFETY-39-slew-against-the-rail.md` §6, and do not read this entry as
more than it says.


**`test_peakguard.py` answered two of its own checks with the wall clock
(CI-06).** The file sits on the required `checks / unit (pull_request)` context
and had a measured ~20 % flake rate. It cost a wrong merge diagnosis (parents
green, merge red — read as a semantic merge conflict, it was the flake), a false
red on two reviewers' CONTROL runs, and a false "mutation caught" that reached a
merged card in four places. **A flaky check does not waste a re-run; it
manufactures evidence for whichever hypothesis the reader already holds.**

Both causes are the control law's own gain `900 / remaining` applied to a
wall-clock gap the file never intended to have. Neither fix widens anything.

* **The cross-time projection.** The "one cycle of the intent form is exactly
  the delivered form plus the command standing on the wire" check decided the
  intent form from a **second** `PeakTracker.project(..., datetime.now())` taken
  after the real cycle had returned, and compared it against a floor decided
  from the **first** projection — three lines under a comment saying both run
  "against the same frozen projection". `excess = 900 × (projected − ceiling) /
  remaining` moves at `excess / remaining` W/s, so at a 1500 W floor the 1.5 W
  tolerance bought exactly **`remaining / 1000` seconds**: 333 ms with 880 s
  left in the quarter, **6 ms with 10 s left**. The gap between the two
  statements measured **12.7–39.5 ms** over 26 instrumented runs. The check was
  decided by *where in the quarter it ran*. It now keeps the `Projection`
  **object** the real cycle graded and decides both forms from it — the clock is
  out of the comparison and the tolerance is **exact**, not widened. Measured on
  a real cycle: re-deciding the delivered form from that object reproduces the
  controller's own floor to **0.000000000000 W**.
* **The fixture's own room.** `wired_over_limit()` laid telegrams with
  `range(0, span + 1, 5)`, leaving **0–5 s of the open block uncovered by
  construction**, and `excess` multiplies an uncovered second back up by
  `900 / remaining`. Measured on the shipped file: at **12 s of remaining a
  fixture that asked for a 1500 W floor delivered 613 W (−59 %)**, and at 2 s
  the guard **did not engage at all** — which fails "the real cycle files the
  peak claim", "…and the number written to the inverter is the guard's floor"
  and "the STATUS PAGE row for a floor-limited guard" **together**. The
  four-attempt retry could not help: four attempts take ~40 ms and all land in
  the same dying quarter. Fixed with one telegram stamped at `now` (what a real
  meter does, and what the docstring already claimed) plus `FIXTURE_ROOM_S`,
  **derived** from the 1 % accuracy the helper promises rather than picked:
  `net_w × lag / (remaining × want) ≤ 0.01 − lag/900` → **23.6 s → 24 s**, at
  the measured 0.1 s worst-case lag.

**Nothing was skipped, disabled or marked xfail** — the file still carries all
of SAFETY-07 and SAFETY-25's forced-discharge SoC floor, and it gains a check:
the delivered form re-decided from the cycle's own projection must reproduce the
real controller's number to the bit, which pins the test's `decide()` kwargs as
main.py's actual call rather than an approximation of it. **One tolerance got
stricter** (1.5 W → exact), because removing the clock made exactness possible.

**`site_standing` — the one quantity BOTH site guards are asked about — is now
pinned by a fixture whose per-unit shares cannot answer for it (SAFETY-38).**
`app/main.py`'s `site_standing = float(sum(standing.values()))` is the first
argument of [[SAFETY-30]]'s **latching, site-stopping** sign detector *and* the
`commanded_w` handed to [[SAFETY-18]]'s `looks_frozen()`. It was pinned by
**nothing**, and three mutations of it disarmed both guards with the whole suite
green. **This card changes coverage only: not one byte of `app/` moves — the
shipped `sum` is correct and must not be "fixed".**

* **⭐ THE FIXTURE THE SUITE COULD NOT BUILD: per-unit shares EACH BELOW the
  site gate whose SUM clears it.** Two simulated units rated 220 W and 280 W,
  both obeying exactly, inverted gain, house importing 2000 W — the site
  commands **−500 W against a +2500 W meter** and no single share clears the
  300 W gate. That is the case the site detector exists for, and it is the only
  arrangement that can separate `sum` from the mean and from the largest share.
  `test_watchdog.py` part 9a.
* **WHY THE EXISTING FIXTURE COULD NOT.** `_sign_site`, the suite's only N≥2
  site fixture, runs shares of about −2500 W **each** — every one individually
  over the gate — so sum, mean and largest-share all agree on it. *Its values
  happen to be equal, and a check comparing two things that happen to be equal
  cannot see the difference it names.* `MUTATION.md` habit 4, ninth instance,
  inside the site half of the very card that names it.
* **FOUR CANDIDATE REDUCTIONS, FOUR DISTINCT ANSWERS, ONE OVER THE GATE** —
  `sum` −500, mean −250, `max(key=abs)` −280, `[0]` −220 — and that is asserted
  as a precondition, not described. The asymmetric ratings are load-bearing:
  equal shares collapse three of the four onto one number, and the **smaller**
  unit is first so `[0]` stays distinct from `max(key=abs)`. The **gate is the
  shipped 300 W**; only the hold is shortened, because the hold is a wall clock.
* **⭐ THE SECOND CONSUMER IS COVERED, NOT SCOPED OUT.** The same value is
  `looks_frozen()`'s `commanded_w` at `main.py:1926`. At the same roster the
  site rail is 0.95 × 500 = **475 W**, so `sum` = 500 W rails it while 250, 280
  and 220 do not. Shipped: the peak guard goes blind on SAFETY-18's conjunction
  and files `floor_w = 0.0`. Under both surviving mutations it is **not blind at
  all and files no floor**, so the battery may keep importing through a quarter
  whose billed average is unobservable. `test_watchdog.py` part 9b. The open
  quarter is deliberately backfilled with the stuck reading so the mutant cannot
  reach the same action by the *coverage* route and mask the difference.
* **AT N=1 NOTHING CHANGES, AND IT IS ASSERTED.** On a real one-inverter
  controller `standing` has one entry, so `sum`, the mean, `max(key=abs)` and
  `[0]` are **the same number** — all three mutations are genuinely *equivalent*
  on a shipped install and no N=1 fixture could ever separate them. Printed and
  checked, beside the site detector still being absent there. Part 9c.
* **⚠️ A CORRECTION, AND IT IS THE MOST USEFUL LINE IN THIS CARD.** An earlier
  revision of this entry said one of the three — `list(standing.values())[0]` —
  was **already closed** by [[SAFETY-32]]'s package-wide structural pin, and
  told future authors not to re-cover it. **That was false.** It had been
  measured **once**, that run hit `test_peakguard.py`'s [[CI-06]] flake (a
  measured **~20%** false-red rate), and the red was recorded as a mutation
  **kill**. Re-measured on `release/1.0` @ `86eaef2` carrying only that edit:
  the **full 21-file suite is green** and `test_peakguard.py` passes **five for
  five** in isolation — **the mutation survives.** The reason SAFETY-32 cannot
  see it is not the shape: part 12j taints names *bound* to roster values and
  lists *accumulated* in a loop, and `standing` is a dict filled by **subscript
  assignment**, so it never enters the taint set and `sum(standing.values())` is
  not a roster site at all. 12j's own enumeration on the mutated tree lists
  `main.py:1553`, `1555`, `1690` — **not `1599`**. **All three mutations needed
  this card; all three are now covered by it.** *A single red from
  `test_peakguard.py` proves nothing — re-run it five times in isolation and
  name the failing check.*
* **⚠️ AND A ZERO-BYTE SURVIVOR ONE LINE BELOW THE ONE THIS CARD PINNED.**
  Pinning the *value* `site_standing` said nothing about **where it goes**.
  Swapping the detector's first two arguments —
  `evaluate(site_standing, self.grid, …)` → `evaluate(self.grid, site_standing, …)`
  — is **zero bytes** and left the tree fully green, because the check asserted
  only that `"-500 W"` appeared **somewhere** in the banner. It still does:
  **in the meter slot.** The swapped banner reads *"commanded +2500 W against a
  meter reading −500 W"* — a site charging at −500 W into a +2500 W import, told
  to the operator **exactly backwards**, on the one line somebody woken by a
  latching site-stopping alarm reads first. The check now asserts the **ordered
  pair** rather than a substring, and is red on it. **No `app/` change:** the
  hole was in the test.
* **⭐ A SURVIVOR THIS CARD WENT LOOKING FOR AND CLOSED.**
  `site_standing = float(self.target)` — the arbiter's site *target* instead of
  the sum of what actually **stands on the wire** — is **−11 bytes** and
  survived the full suite, because in every fixture that existed the two are the
  same number. They are not when a device's reported range is narrower than its
  `max_w`: the `device` claim is filed **per actuator**, so it bounds each
  unit's share and never the site target. A site asking for −500 W while −250 W
  stands would then take a **latching site-wide false trip** under its own
  300 W gate. Closed by part 9a's device-clamped counterfactual.
* **⚠️ AND THE SAME ZERO-BYTE DEFECT ONE CALL SITE OVER — THE ONE A SHIPPED
  INSTALL CAN ACTUALLY SEE.** The ordered pair went onto `main.py:1620`, the
  **site** detector. That detector is armed only at N≥2, and part 9c asserts it
  is **absent at a roster of one** — so on every install this add-on has, the
  **per-unit** banner is the only sign-disagreement sentence that can ever be
  printed, and its call at **`main.py:1560`** was still unpinned. The identical
  swap there is **0 bytes** and left the tree **21/21 green**, and nothing
  behavioural could ever have caught it: `disagrees()` is **symmetric** in those
  two arguments (`commanded_w * against_w < 0`, `abs()` gate on both), so the
  mutant takes **identical actions on every input**. What moves is the sentence
  — measured, **9 emitted banner lines invert in 3 distinct sentences**,
  including the one persisted to `sign_reason` and read back after a restart:
  *"commanded +1000 W against its own measured power −1000 W"* becomes
  *"commanded −1000 W against its own measured power +1000 W"*. Closed by the
  ordered pair on **both** per-unit banners — the N≥2 `its own measured power`
  one and the **N=1** `a meter reading` one — with both operands read off the
  **observed cycle** rather than off the arguments the detector was handed, plus
  a precondition on each that the two operands really are different and opposite
  in sign. **Red on four named checks, green in the other twenty files. No
  `app/` change.**
* **⭐ AND THE FAMILY IS NOW ENUMERATED BY MACHINE, because it had been
  enumerated by hand twice and was short both times.**
  `tools/symmetric_message_scan.py` finds every call whose argument order is
  load-bearing for a **message** rather than for a decision: it renames each
  parameter pair, erases every formatted string, canonicalises the commutative
  operators and iterates to a **fixpoint** (the symmetry is usually one call
  deep — `evaluate()` is symmetric only *through* `disagrees()`), then
  classifies each sentence by where it goes. A second census covers sentences
  the **caller** builds. It found **one open survivor this card does not close
  and now discloses** — `rail_unreachable()`'s `dev_min`/`dev_max` at
  `main.py:3208`, **0 bytes, 21/21 green**, a status-page row — and **twelve**
  unpinned caller-built sentences including SAFETY-31's `circ_reason`, which is
  the same shape and is also persisted. The list, the pinned status of every
  member and the instrument's own blind spots are in
  `DECISIONS/SAFETY-38-site-standing-pin.md` §7.
* **⚠️ AND A LEDGER SENTENCE THIS CARD ITSELF GOT WRONG, CORRECTED WITH THE
  MEASUREMENT.** An earlier revision said *"before part 9 existed, **three** of
  these left the whole suite green — M1, M2 and M3 — and the remaining seven
  were caught by SAFETY-30's own N≥2 checks"*. **Eight** did. Re-run one at a
  time on clean `release/1.0` @ `bbc903b`, fresh copy each, bytecode purged,
  byte delta and `main.py` fingerprint printed, **full 21-file suite** every
  time: M1, M2, M3, M4, M7, M8, M9 and M10 are all **21/21 PASS**; only M5 and
  M6 go red. It is the same species of error this round exists to fix — **a
  survivor recorded as a kill** — and the file contradicted itself three times
  while saying it. **And the two that ARE caught make the card's case:** M5
  (`sum(abs …)`) and M6 (`-sum`) are the only two that reverse the **sign**.
  Every other entry changes only the **magnitude**, and the old `_sign_site`
  fixture's ~−2500 W shares put nine different magnitudes on the **same side of
  the same 320 W gate**. The pre-existing coverage could see a sign and nothing
  else.
* **13 mutations — 11 of that line and its call site, plus 2 zero-byte edits to
  the arbiter claim wiring — each on a fresh full-repo copy mounted `:ro`,
  bytecode purged by `tools/mutation_run.sh`, the edit re-read from disk and the
  byte delta printed. All 13 red.** Every one was **re-run independently** in
  this round rather than carried forward from the previous round's report. The
  claim-wiring pair (`sign_site` and `circulation` moved from site scope to unit
  scope) closes two gaps a reviewer disclosed — though **`circulation`'s kill
  belongs entirely to pre-existing checks and is not claimed by this card.**
  ⚠️ One **false red** from `test_peakguard.py` on M10 did not reproduce in four
  subsequent runs; it is CI-06's known flakiness, no claim here rests on it, and
  it is recorded rather than quietly dropped — unlike the one that *was*
  believed, two bullets up. Ledger, and an **enumerated list of what was not
  mutated**, in `DECISIONS/SAFETY-38-site-standing-pin.md` §4a, §4b and §6a.
* **⭐ ROUND 4: THE SCANNER COVERED ONE SYNTACTIC SINK AND CALLED IT THE
  FAMILY, AND TWO LIVE MEMBERS WERE SITTING IN THE HOLE.** Every one of census
  B's 16 members was a sentence that is **appended, assigned, or logged**. Not
  one was **RETURNED** — and census A's return bucket only ever held a callee
  formatting its own *parameters*, so a method formatting its own **attributes**
  and handing the string back was invisible to both. That is the whole
  `describe()` / `explain()` family, which is where this codebase keeps its
  operator sentences. **Worse than the gap: it was in none of the five blind
  spots the tool printed about itself.** The tool now scans the return sink
  (census C), resolving it through the returning method's **call sites** so that
  a returned *key* — `store.unit_key()`, `mqtt.unit_state_key()` — is not
  mistaken for a returned *sentence*; it also scores a sentence handed to a
  constructor as `reason=` or placed under a `"text"` dict key, which added
  **six more caller-built members** it had been stepping past. Re-run, it
  **rediscovered both of the members this round was sent to fix without being
  told where to look**.
* **⭐ `ClampResult.describe()` — the clamp alarm the owner has been complaining
  about all day — had a SECOND independent way to lie, and no `test_*.py`
  mentioned `describe()` at all.** Transposing `requested_w` and `applied_w` in
  `clamp.py:253` is a **zero-byte** edit that leaves the guard reporting a
  setpoint clamped **UPWARD**, in the measured case to +9999 W by a bound whose
  rating the same sentence gives as 2000 W. It is the §15 action-log row
  (`main.py:1366`) **and** the ClampWatchdog notification reason
  (`main.py:2452`). Measured **21/21 green** before; now **red on 4 checks** in
  `test_clamp.py` part 6d, pinned on the **ordered pair** on the contiguous
  literal at **both** consumers, with operands read off `ctl.target` and
  `hass.writes[-1]` — what the loop asked and what the device was sent — rather
  than off the `ClampResult` the swap lives in.
* **⭐ `PeakDecision.explain()` — the status-page banner — the same shape.**
  `projected_w` ↔ `ceiling_w` at `peakguard.py:493`, **zero bytes**, and the
  row then reads *"quarter projects 5013 W under a 2300 W ceiling"*: a quarter
  **2713 W over** its ceiling, on a row whose own words say *idle* and whose own
  flag says **ok**. `"peak guard idle"` appeared in no `test_*.py`. Measured
  **21/21 green** before; now **red on 3 checks** in `test_peakguard.py` part 13,
  pinned on the decision and on the shipped status row at `main.py:3313`.
  ⚠️ **A correction the ticket is owed:** the idle sentence does not reach
  `main.py:3299/3302` — those are the *floor-limited* branch, where `explain()`
  returns `self.reason` before the idle literal is built. It reaches
  **`main.py:3313`** and, on the release edge, `main.py:1951`.
* **`test_arbiter.py`'s scope-violation check made a claim WIDER than its
  assertion** — it said `scope_violations` *"names the source and the actuator
  it does not belong on"* and asserted `"peak" in scope_errors[0]`, which is
  still true after the two swap roles. Measured this round: **0 bytes, 21/21
  green**; now red on 2 checks, pinned on the ordered pair.
* **14 more zero-byte mutations measured, and 10 SURVIVORS ARE DISCLOSED RATHER
  THAN CLOSED**, each with its file fingerprint and its 21/21 verdict, in
  `DECISIONS/SAFETY-38-site-standing-pin.md` §7e–§7g: **two** in
  `allocator.py`'s `UnitPlan.explain()`, three across `arbiter.py`'s two
  `explain()`s, two in `tariff.py`, two in `arbitrage.py`'s trade reasons and
  one in `main.py`'s SoC-floor notification. The other four of the fourteen are
  kills: `allocator.py:162` was already caught by `test_alloc.py`, and
  `clamp.py:253`, `peakguard.py:493` and `arbiter.py:264` are closed here. **An enumerated list with honest
  gaps beats a fifth round finding a fifth instance.**
  ⚠️ **And the probe that grades those mutations was itself the defect, once.**
  Its first cut numbered the operands **by order of appearance**, so after a
  transposition the other leaf took the same number and all thirteen renders came
  back **byte-identical** — which reads exactly like *the edit is inert*. Keying
  the value to the **name** fixed it. A measurement comparing the swap against
  itself, in the instrument this time.
* **The tool now states what it does NOT scan**, largest hole first:
  `%s`-style lazy logging (`_LOG.error("… %s … %s", a, b)` puts the operands in
  *arguments*, so no census here looks at them and the transposition is the same
  zero-byte edit — **31 such calls across 8 modules, counted and listed by the
  tool rather than asserted**), `.format()`/`%`/`+`/`.join()` sentences, a sentence that
  leaves through a dict or dataclass and is rendered elsewhere, `yield`ed
  sentences, message attributes outside the `reason`-shaped set, and everything
  outside `app/`. Its five over/under-report gaps stay, and every census row now
  prints whether the two interpolations **share a format spec** — the difference
  between a transposition that is silent and one that merely looks broken.

**Inverter B gets its own MQTT device, and inverter A keeps the one it has
(TEL-07).** MQTT discovery announced **one** device, and every entity hung off
one `unique_id` prefix, so at two inverters B's per-unit telemetry had nowhere
to go. It has one now — **without touching a single entity that already
exists.**

* **⭐ A's discovery payloads are byte-identical to `release/1.0`.** Same topics,
  same `unique_id`s, same payload bytes, same retain flags, in the same order —
  verified by **diffing the announced payloads**, not by reading the code:
  `9 topics, 0 missing, 0 changed, 0 added` at a roster of one, and
  `0 missing, 0 changed, 8 added` at a roster of two. That is what protects the
  live install's recorded history and the three watchdog automations that name
  its entities.
* **⚠️ THE ASYMMETRY IS DELIBERATE — DO NOT "CLEAN IT UP".** The first roster
  entry's per-unit rows go on **the device it already has**; every later entry
  gets its **own** device (`via_device` the controller). "One device per
  inverter" is tidier and is the change that renames every entity on the live
  system — and because HA keeps the id it composed at first discovery, a rename
  produces a **new entity beside a dead one**: the history stays behind and
  every dashboard, template and automation goes on pointing at a sensor that
  quietly stopped updating. The composed ids come out symmetric anyway
  (`sensor.goodwe_inverter_a_battery_power` /
  `sensor.goodwe_inverter_b_battery_power`).
* **Four per-unit rows** — setpoint, battery power, SoC and the SAFETY-02 sign
  latch — announced **only from a roster of two**. At N=1 nothing is announced
  and nothing is published: no second device, no renamed entity, no added topic,
  asserted rather than assumed.
* **Site-scoped sensors stay site-scoped.** `grid` and `p1_sample_age_s` are
  announced **once**, on A's device, at every roster length 1–6. There is one P1
  and one grid connection, and `p1_sample_age_s` is a **watchdog input** — a
  second copy is a second answer to "is the meter fresh?", one of which is
  always the one nobody is looking at. Circulating power, controller status,
  maintenance phase and the capacity-tariff peak are site for the same reason.
* **`inverter.roster_names()`** is now the one answer to "what are the units
  called", read by `build_roster()` **and** by the publisher — which has to know
  the roster **before** a `Controller` exists, because announcing happens on the
  broker's CONNACK.
* **`test_mqtt.py`**, a new suite: the recorded `release/1.0` baseline
  (generated by running that commit's module, not transcribed), the payload
  diffs, the composed entity ids modelled **independently** of `app/mqtt.py` and
  anchored on the one id verified against the live system
  (`sensor.goodwe_p1_sample_age`), the announcement-vs-publication agreement on
  a real `Controller` with **different** SoC, battery sign and latch state on
  the two units, and **twelve mutations of the shipped module, all red, each on
  a named property**.
* **MEASURED, NOT ASSERTED.** Three of those properties exist because the
  mutation table found their absence: dropping the unit from the *state topic*,
  moving `grid` into the per-unit set, and `units=()` in main()'s publisher call
  were all **green** on the first run. So was a roster fallback that named two
  nameless entries `inverter-a`. Sixteen mutations were run through
  `tools/mutation_run.sh`; the survivors are enumerated **with their size** in
  `test_mqtt.py` part 9.
* ⚠️ **Inverter B does not physically exist** (~2026-09-12). Every fixture here
  is **simulated**, from the datasheet and serial in `NOTES.md` §1.1b. Nothing
  in this card was measured on a second inverter.

**The add-on configuration screen now speaks English and Dutch instead of raw
option keys (UX-01).** There was no `translations/` directory at all, so Home
Assistant fell back to rendering the option KEYS — an operator opening
*Configuration* read `unit_sign_disagree_w`, `saturation_cycles`,
`integrator_max_w`, seventy of them, in file order, with no description and no
way to tell a constant derived from 1890 closed-loop test runs from their own
inverter's rating. **This card changes presentation only: not one default, not
one range and not one line of behaviour moved.**

* **`translations/en.yaml` and `translations/nl.yaml`**, with a `name` and a
  `description` for **all 70 options** and for **all 7 per-entry keys** of the
  `inverters:` repeatable section. The Dutch is written for a Flemish
  homeowner rather than transliterated: *afname* and *injectie*, *kwartierpiek*
  and *capaciteitstarief*, *laadniveau*, *uitschakelgrens* — the words that are
  on the bill and on the meter.
* **⭐ EVERY DERIVED CONSTANT SAYS WHERE IT CAME FROM AND WHETHER TO TOUCH IT.**
  `unit_sign_disagree_s` now reads *"derived from 1890 closed-loop test runs
  (SAFETY-30): 85 s is 8.5 times the worst legitimate opposition those runs
  produce… do not change it without re-deriving"*; `circulation_w` cites its 924
  runs and the measured 400 W knee; `meter_freeze_s` cites the 2399 s capture;
  `bms_soc_floor` says it is register `0x0560` read off the pack. The two
  numbers that are genuinely the operator's — `max_w` and `peak_cap_w` — say
  **"THIS IS YOUR NUMBER"** instead, and `cycle_cost_threshold_eur`, which is
  neither measured nor the operator's, admits it is an open product question
  (OPT-08) rather than borrowing the others' authority.
* **`config.yaml` reordered into task order** — hardware/entities → meter →
  ratings and limits → tariff → maintenance → safety → advanced — in **both**
  halves, with every comment carried with its key. ⚠️ The rendered order is the
  **`schema:`** half's (`App.schema_ui` is `UiOptions(...)(data["schema"])`);
  the `options:` half is reordered to match so the file reads the way the tab
  renders. ⚠️ **`test_inverter.py` part 12's first-match trap is preserved**:
  the nested per-entry `max_w:` still appears BEFORE the flat one under
  `schema:`, and `inverters: []` still ships in the `options:` half. Moving the
  nested block below the flat one is a **1-byte** edit and it was run: red in
  `test_inverter.py` and in `test_translations.py`.
* **MEASURED, NOT ASSERTED.** `tools/supervisor_shape_probe.py` gains a section
  F that loads both files with Supervisor 2026.08.0's *own*
  `_read_app_translations()` out of the release image (a file that fails
  `SCHEMA_APP_TRANSLATIONS` is **dropped with a log warning and nothing else**),
  runs its own `UiOptions`, replays the frontend's label resolution transcribed
  from the shipped bundle, and prints every rendered row in both languages.
* **⚠️ ONE MEASURED NEGATIVE RESULT, RECORDED RATHER THAN GLOSSED.** The
  per-entry keys inside the `inverters:` repeatable section still render RAW.
  `_convertSchemaElement` sends `type: "schema"` **with** `multiple` to
  `ha-selector-object`, which labels its fields `fields[name].label || name`,
  and the converter sets neither `label` nor `translation_key`. The `fields:`
  translations are shipped and validated because that is the shape Supervisor's
  own schema defines for exactly this, and a frontend that wires `label` through
  would consume them — but on 20250925.1 they do not reach the screen, and this
  card does not claim they do. The `inverters` label and description themselves
  do.
* **A new check, `test_translations.py`, so this cannot rot.** A translation key
  naming an option that does not exist, or an option with no translation, fails
  CI — in both directions, for both languages, top-level and per-entry. It is
  **mutation-tested both ways**: eleven mutations applied in-process, each
  required to NAME the key it broke rather than merely to go red, plus a
  negative control that the unmutated input is clean. Two of the eleven are the
  ones that cost the WHOLE FILE rather than one label — an unrecognised
  top-level block, and a `network:` value that is not a string. Supervisor
  validates each translation file through `SCHEMA_APP_TRANSLATIONS` and, on
  refusal, **logs a warning and drops all of it**, so those silently return the
  entire tab to raw keys; neither is reachable through the `configuration:`
  coverage, which is why they were hunted for and closed. Six more were run **on disk**, through the
  purge-and-re-read protocol: deleting a translation (−65 B), renaming an option
  in `config.yaml` (+4 B), adding an orphan key (+47 B), emptying a Dutch
  description (−711 B), breaking one label's group prefix (−4 B) and deleting a
  provenance citation (−19 B) each go red.
* **The parser has a licence.** There is no PyYAML in the add-on's CI image, so
  `tools/translations_lint.py` reads these files by hand — and a hand parser
  that quietly disagrees with the real loader is the defect this repository
  keeps finding. The probe imports **that same module** and asserts, key by key
  and string by string, that it returns what `_read_app_translations()` returns.
  The parser is deliberately strict and **raises** on anything outside the
  subset; nine malformed shapes are fed to it to prove it does.
* **⚠️ WHAT THIS CHECK CANNOT SEE, stated with its size.** It grades coverage
  and provenance *citations*, not truth. Falsifying a figure inside a
  description that is not one of the cited tokens survives (run: *"8.5 times"* →
  *"3 times"*, −2 B, green), and swapping two options' descriptions survives
  (0 B, green). Separately and out of this card's scope: changing a shipped
  default that no suite pins survives — `step_w: 10` → `20` (0 B) leaves
  `test_control.py`, `test_clamp.py`, `test_validate.py` and
  `test_peakguard.py` green. That is the same family as CFG-04's DEFECT 2 and
  the UX-01 rescope moved it to its own card on purpose.

**Inverter B's rating is 4600 W, FROM ITS DATASHEET, and every documented copy
of it now says so (SPIKE-04a).** Three places stated B's rating — `config.yaml`'s
commented roster example, `DOCS.md`'s, and `tools/supervisor_shape_probe.py`'s
"a blob an operator would POST" — and all three said **3000 W**, a number that
was never measured, never read off a datasheet, and simply wrong. The owner
supplied B's model and serial on 2026-09-04: a **GoodWe GW5048D-ES**, serial
`55048ESU22AW0992`, max charging and discharging power **4600 W**. No control
behaviour changes here; this is documentation and its pin.

* **Serial characters 6-8 read `ESU`** → platform 105, AA55, the supported set.
  Not `ESN`/`ESC` (ES G2, platform 745). SPIKE-01 does not gate this site.
  ⚠️ A passed platform gate is **not** a proven control surface: `specs.md`
  already calls other platform-105 serials *"same protocol family; untested
  silicon"*, and this repo's `0x056E`/`0x056F` map was reverse-engineered from a
  passive tap on **one BPS unit**. That check is the bench half of SPIKE-04.
* **⚠️ 4600 W IS FROM A DATASHEET AND IS NOT A MEASUREMENT, and one ambiguity is
  recorded rather than guessed.** The datasheet prints it under *Battery Input
  Data*, so it is probably **DC**, while `max_w` is graded against `0x0518` —
  the inverter's own **AC** power. If it is DC the AC ceiling is nearer
  4600 × 0.94 ≈ **4324 W**, and the charge direction moves the other way
  (≈ 4894 W drawn AC-side). **Neither reading is written in as fact.**
  SAFETY-20's `min(max_w, dev_max)` already bounds it correctly under either,
  once `dev_max` is read off the unit — so `max_w: 4600` is what to configure
  today. `goodwe/NOTES.md` §1.1b carries the full record, with the date and the
  method on every figure.
* **A NEW CHECK, BECAUSE A DOCUMENTATION FIX WITHOUT ONE DRIFTS AGAIN.**
  `test_validate.py` now pins all three documented rosters: they must be
  findable, they must agree, and they must say **4600**. It also grades the
  **provenance beside the number** — the datasheet it came from, the `0x0518`
  it is graded against, and the 4324 W the DC reading would make it — **within
  2500 characters of the roster, not file-wide.** The first version searched
  whole files and was already vacuous on two of them, because `0x0518` appeared
  in `config.yaml` and `DOCS.md` for unrelated reasons before this card: a
  check whose data could not reach the failure it named, in the file that
  exists to complain about those. Four mutations were run against the fixed
  version — the probe back to 3000 (0-byte delta), the `DOCS.md` caveat
  deleted, `config.yaml` alone "corrected" to 4324 (0-byte delta), and the
  probe's entity renamed so the regex finds nothing — and all four are red.
* **THE TEST FIXTURES KEEP 3000 W ON PURPOSE, AND NOW SAY WHY AT THE LINE.**
  `test_inverter.py`, `test_maintenance.py`, `test_peakguard.py` and
  `test_watchdog.py` build a second inverter at 3000 W against A's 5000 W. The
  only property those checks consume is that B's number is **unmistakably** not
  A's, and the real pair is a **400 W** gap — narrower than `slew_w`, inside the
  range where an off-by-one-unit bug and ordinary control arithmetic produce
  indistinguishable numbers. Collapsing them onto the real values would buy
  realism those checks do not use and cost them their separating power, which
  is the defect `test_peakguard.py`'s own §10g precondition exists to prevent.
  They are **illustrative, not B's rating, and no longer a placeholder for it**.
* **Two consequences with numbers, recorded in `goodwe/NOTES.md` §1.1b for the
  cards that own them.** At 94 % battery-to-AC a round trip is ≈ 88 %, so A
  charging 3 kW while B discharges 3 kW burns ≈ **360 W continuously** with the
  meter at zero — SAFETY-31 argues its threshold from *"a few watts"* and has no
  figure. And **< 13 W night self-consumption per unit** (≈ 26 W for two,
  ≈ 0.6 kWh/day), which nothing in this add-on accounts for anywhere.
* **The site is asymmetric in opposite directions:** A = 5000 W / 15.36 kWh
  nominal, B = 4600 W / 20.48 kWh — ~33 % more energy, ~8 % less power. No
  stability implication (ARCH-01's ḡ = Σαᵢgᵢ is a convex combination), but a
  full recalibration of B moves 33 % more energy and **MAINT-05's 12 h `drain`
  phase timeout was derived on A**; it should be re-checked against B rather
  than assumed to carry.

**Every pack is recalibrated on its own schedule, and only one moves power at a
time (MAINT-05).** §12's monthly cycle drains a pack to empty and refills it to
true its coulomb counter. There was exactly one `Maintenance` state machine and
one un-namespaced history, so at two packs **only the first was ever
recalibrated** and the second's counter drifted forever — and nothing prevented
the cycle draining pack A while the loop or arbitrage charged pack B, which is
two power-moving phases on one grid connection pulling against each other.

* **ONE `Maintenance` PER ROSTER ENTRY**, each with its own schedule, phase and
  history under `unit/<name>/...`. `store.migrate_maint_keys()` moves an
  existing install's flat history into the first pack's namespace under
  `migrate_unit_keys()`'s three rules — shared in code, not restated — with the
  same read-back postcondition, and the flat keys are kept **live** as the
  rollback view so a downgrade finds a current history rather than one dated
  migration day.
* **ONE MAINTENANCE SLOT PER SITE.** At most one pack may be in a power-moving
  phase (`drain`, `charge` **and** `hold` — `charge` is in the set because a
  pack pulled full from the grid beside a pack exporting to empty is 4 kW
  circulating with the meter at zero). The holder is **derived** from the packs'
  own stored phases, not persisted, so a restart mid-drain re-derives it. When
  several packs are due in the same hour the slot goes to the one that has gone
  **longest without a completed cycle**; roster order breaks a numeric tie and
  nothing else. Nothing preempts a cycle in flight.
* **THE DEFERRED PACK IS NOT STARVED AND IS NOT A FAULT.** A deferral does not
  stamp `last_start_attempt` (that is the 90 min post-abort cooldown, inside a
  60 min start hour), and does not move the phase, so neither the 12 h `drain`
  nor the 36 h `charge` timeout — *"the battery is not following commands"* —
  can be reached by waiting. It says WAITING, names the pack it waits for, and
  recalibrates at the next start hour; a test drives the winner's whole cycle
  and observes it start.
* **THE SCHEDULE IS NOT STAGGERED BY CONSTRUCTION.** The interlock defers and
  the stagger is what comes out of it: one contested month separates two packs
  by one cycle length and they stay separated. `DECISIONS/MAINT-05-per-pack-maintenance.md`
  §3 carries the argument.
* **`validate.py` rule 6 now grades `maintenance_charge_w` against the WEAKEST
  cap on the roster**, like rule 5. Its own comment said it would have to move
  the day `maint_unit` stopped being `inverters[0]`; that day is this card. It
  refuses configurations that used to start — a 3000 W charge beside a 2000 W
  pack that is now recalibrated too — which is the `lo > hi` contradiction rule
  6 exists to refuse, caught when the config is written. Identical at N=1.
* **AT A ROSTER OF ONE, NOTHING CHANGES**, and it is asserted rather than
  assumed: the state file keeps the flat keys and gains none, the migration does
  not run, the four cycle events are the shipped strings character for
  character, and the status banner is the bare phase with no pack name.
* **THE MIGRATION IS GRADED AGAINST THE ROSTER, NOT AGAINST THE NAME IT WAS
  HANDED** (fix round). `migrate_maint_keys()` takes the roster rather than a
  pack count and `_verify_moves()` checks the namespace it wrote against the
  names the machines actually read. A postcondition that reads back the
  namespace it was given cannot detect being given the wrong one: handed
  `inverter.UNIT_A` on a roster of `["garage", "shed"]`, the first version of
  this card wrote all four keys under `unit/inverter-a/...`, read all four back
  equal, reported **verified**, and stamped `maint_schema_version` — so both
  packs came up believing they had never been recalibrated and the next start
  skipped the migration too. That is now `misdirected`: not `ok`, the marker is
  **not** written, and the next start re-runs and recovers (rule 1 makes the
  re-run a no-op). `build_roster()` numbers a nameless roster entry
  `inverter-1`, so `inverter-a` is not even the fallback for a declared roster.
  Every two-pack fixture that could see this now names its first pack
  `garage`; with them all called `inverter-a`, `self.inverters[0].name` and
  `UNIT_A` were the same string and the swap left all 19 suites green.
* **AND THE REFUSAL NOW SURVIVES THE NEXT START** (second fix round). The fix
  above graded only the keys it had just *written*, so it refused exactly once:
  measured at Controller level over one persisted store, roster
  `["garage", "shed"]`, with the call site still wrong on both starts —
  construction 1 `ok=False misdirected=4 moved=4 kept=0 marker=<absent>`,
  construction 2 `ok=True misdirected=0 moved=0 kept=4` and the marker
  **stamped**. On the second start rule 1 finds the wrong keys already there and
  routes all four to `kept`, so `moved` is empty and the roster loop grades
  nothing — and start 3 skips the migration forever, *including after the call
  site is fixed*. Loud-and-recoverable became silent-and-permanent after one
  restart. `_verify_moves()` now grades `set(moved) | set(kept)`; both lists are
  built from the same `unit` argument, so grading `kept` grades the same
  namespace decision. `test_maintenance.py` part **9c** drives two starts with
  the same wrong unit over one store — the shape nothing in the suite had, since
  9b measured recovery along the path where the operator has *already* fixed the
  call site.
* **AND THE MARKER GATE IS PINNED ON BOTH GUARDS.** `if rep.ok:` is
  `not lost and not misdirected`, and the mutation `if rep.ok or rep.kept:`
  stamped the marker on a migration that **lost a month of history** with all 19
  suites green. Part **9d** is the fixture that separates the two guards: two
  keys already namespaced (`kept` non-empty) and two whose write the store
  drops (`lost` non-empty), `misdirected` empty throughout.
* ⚠️ **AND THE SAME MUTATION ONE FUNCTION UP WAS A SURVIVOR, FOUND HERE AND
  CLOSED.** `migrate_unit_keys()` — SAFETY-02's **latch** migration — gates its
  own marker on `rep.ok` too, and every check that reached that gate ran on a
  store where nothing was namespaced yet, so `rep.kept` was empty in all of
  them. `if rep.ok or rep.kept:` there stamps `schema_version` on a migration
  that **dropped the latch**: the guard is off, the file still says
  `sign_tripped: true`, and the next start skips the migration that would have
  re-armed it. Measured a survivor of all 19 files; now red on
  `test_inverter.py` part 10, which drops the `sign_tripped` write beside an
  already-namespaced `sign_reason`. No source change — the shipped gate was
  right; nothing could see it.
* ⚠️ **FOUND, RECORDED, NOT FIXED (SECOND ONE).** `migrate_unit_keys()` —
  SAFETY-02's latch migration — has the identical hole and it costs a **guard**
  rather than a history. It runs before `build_roster()` (each `Inverter`
  restores its latch in its constructor), so it is handed the constant `UNIT_A`
  and there is no roster to grade against. On a declared roster a pre-DEV-02
  `sign_tripped: true` lands under `unit/inverter-a/sign_tripped`, which no unit
  reads — and `_grade_store_migration()`, whose job is to re-arm exactly this,
  is **defeated by the key the defect wrote**: its "is anything namespaced
  latched" test is satisfied by the misdirected value itself. Fixing it means
  naming the migration's target from the config before any `Inverter` exists,
  which is `inverter.py` and SAFETY-02's semantics, with its own battery.
  Measured and pinned as a running check (`test_maintenance.py` part 13), whose
  assertions state the WRONG behaviour on purpose.
* ⚠️ **FOUND, RECORDED, NOT FIXED.** A maintenance drain **pins** its pack while
  the loop's target is a **site** claim, so when the site target is below the
  pin the only split asks the sibling to charge against the drain and
  `allocator.plan()` refuses it — the site parks at 0 W and the pack's 12 h
  timeout would eventually report *"drain stuck"* about a healthy battery.
  Pre-existing on `release/1.0` at `2d025ee` with the unit names swapped; fixing
  it means changing what the loop's target means while a unit is pinned, which
  is loop and allocator territory. Pinned as a running check
  (`test_maintenance.py` part 12) so it cannot be forgotten.

**`bms_soc_floor` describes a BATTERY, so at two packs it stops being one number
(SAFETY-32).** The peak guard's energy backstop rests on two figures: the EMS
floor `peak_soc_floor` (the depth a money-saving discharge may never reach) and
`bms_soc_floor`, the percent **that pack's own BMS** cuts out at — register
`0x0560`, read as 10 on the reference GW5000S-BP. Rule 8 refuses an EMS floor at
or below the cut-out, because such a floor can never bind: the BMS wins and the
status page reports a backstop that does not exist. With ONE cut-out for two
packs of different models that rule graded the EMS floor against whichever
battery happened to be written into the flat key, and a site with cut-outs of
10 % and 30 % under `peak_soc_floor: 20` **started** — the guard would then
command a forced discharge that pack B's BMS, not the operator's floor, is the
thing that stops. Rule 8's own failure mode, reached through rule 8.

* **`bms_soc_floor` becomes a per-entry roster key, and it is REQUIRED there.**
  Nothing is inherited: an entry that omits it has no cut-out, does not borrow
  the flat one, and is refused by name at startup. A per-entry *default* was
  considered and declined — a default is one number for every pack, which is the
  defect this card removes; `DECISIONS/SAFETY-32-per-pack-bms-floor.md` records
  the decision and its migration cost. The flat key is unchanged and is now
  documented as the single-inverter shape's own cut-out.
* **Rule 8 grades EVERY unit and NAMES the one it refuses**, through a new
  `unit_bms_floors_indexed()` that carries the **roster** index the value came
  off. CFG-04 had just fixed a rule 5/6 bug of exactly this shape — a caps-list
  index fed to a function that indexes the roster, so the message named the wrong
  inverter — and this reader skips nothing, so there is no second list for an
  index to desynchronise from. Graded on a roster whose offender is the SECOND
  entry *and* on the same two numbers in the other order.
* **The permission is answered PER PACK.** `Controller.pack_floor_soc` — the
  emptiest measured pack — is untouched and is still what `decide()` grades;
  what changed is the floor. `peakguard.pack_permissions()` answers one
  `(name, reason)` per pack against that pack's own floor, and the floor-limited
  notification now names the pack that stopped the site rather than "the
  battery". At a roster of one the alarm is byte-for-byte what SAFETY-25 shipped.
* **An unstated cut-out withholds rather than defaults.** `None`, not `0.0` —
  `discharge_permitted()` already treats an unknown floor as no permission, so
  the absence costs a shave and never a pack, including on `sim/inject.py`'s path
  which does not run the startup gate.
* ⚠️ **Stated rather than hidden:** under any config rule 8 accepts, the per-pack
  *floor* reduces to `peak_soc_floor` for every pack — rule 8 makes it so. The
  reduction is written for `unit_caps_indexed()`'s reason ("unreachable" is a
  claim about rule ordering, not a property of a function) and is graded on both
  sides: 1775 rule-8-valid `(peak_soc_floor, bms_soc_floor)` pairs answer the
  option, and the pair rule 8 refuses answers the cut-out. `peak_soc_floor`
  itself is still one number for the site, so the "deeper pack needlessly
  withheld" half is filed as an open question, not closed.
* **`tools/supervisor_shape_probe.py` re-run and repaired.** MUTATION.md habit 8:
  a committed transcript is a check and it decays. The required key broke §C and
  left §D printing PASS on a refusal about a *different* key. §D is now an
  isolated counterfactual over an accepted, merged baseline, one key varied per
  case, asserting the key **name** in the refusal — `all cases passed` on
  `amd64-hassio-supervisor:2026.08.0`.
* ⚠️ **SECOND FIX ROUND — the round-2 fix carried the defect it was written to
  close, in its prose.** Two more rewrites survived a green 19-file suite on
  `8501e05`: `binding_soc_floor()`'s `max(floors)` → `max(floors[-2:])` (+5
  bytes) and the call site's `self.packs` → `self.packs[:2]` (+4 bytes). Both are
  checks whose TEXT claimed more than their FIXTURE could reach — 12c2's only
  counterfactual varied the **last** pack, which every window ending at the
  roster's end follows, and all three 12g runs drove a **two**-pack roster, where
  `[:2]` *is* the roster. **The general law, now stated in the file: a fixture of
  N packs cannot separate two reductions that agree on N packs**, and the regress
  does not terminate in a fixture — `[:1]` needed two packs, `[:2]` three, `[:3]`
  four. So the reduction is now graded by a **length sweep** (78 simulated
  rosters, lengths 1–12, deepest cut-out at every position, closing head and tail
  windows of width 1–11 at once) plus a **computed family table** of fourteen
  rewrites that goes red naming any member no roster kills; the call site is
  graded by **three-pack runs** with the deep cut-out at index 0, 1 and 2, plus
  an **exact structural pin** (the roster argument must unparse to `self.packs`)
  for the slice widths no bounded roster can reach. Round 3's own battery found
  `self.packs[:3]` / `[-3:]` surviving all six behavioural runs — the structural
  pin is what closes them — and one row is recorded **NOT CAUGHT**: a reduction
  over a window twelve packs wide, which needs a 13-pack roster to observe.
  No shipped behaviour changed in this round; only what observes it.
* **Twelve source mutations, eleven red, one argued survivor**, through
  `tools/mutation_run.sh` (bytecode purged). The battery also found a real
  weakness of its own: `_flat_bms_floor()` had the default in two places and each
  covered for the other, so deleting either survived — masking, in three lines of
  option reading. One defaulting line now.
* ⚠️ **FIX ROUND — that battery was INCOMPLETE, and the review found what it
  missed.** Neither of the two mutations below appears in the twelve rows above:
  the ledger graded "the reduction dropped" and "the bare option handed over"
  and never asked what happens when the reduction is simply the *wrong* one, or
  when the roster handed to it is *sliced*. No shipped behaviour changed here;
  two *checks* whose data could not reach the failure they named did. `max(floors)` → `min(floors)` in
  `binding_soc_floor()` (zero byte delta) and `self.packs` → `self.packs[:1]` at
  the `decide()` call site (+4 bytes) were each green across all 19 files. Both
  survived for the same reason: **under any rule-8-valid config every pack's
  `pack_floor_pct()` equals `peak_soc_floor`**, so `floors` is a list of
  identical numbers and `max`, `min`, first and last are four spellings of one
  answer — and the one check using a cut-out above the EMS floor ran on a roster
  of ONE, where every reduction agrees. The second also walked through part 12g,
  which asserted `"self.packs" in <the unparsed call site>`: a slice still
  contains that substring, so the pin passed on the mutation that grades the
  whole site against pack A. MUTATION.md habit 4 — a source-text test is a check
  on how code is spelled, not on what it does.
  * **New part 12c2** evaluates `binding_soc_floor()` on a roster whose per-pack
    floors **all differ** — four simulated cut-outs of 30/25/40/35 %, every one
    above `peak_soc_floor`, chosen so that `max` (40), `min` (25), first (30),
    last (35) and the mean (32.5) are five different numbers and the two extremes
    sit at *interior* roster positions. Two packs cannot separate those four;
    four is the smallest roster that can. The precondition is its own check
    (habit 1) and one pack — the last, neither deepest nor shallowest — is then
    raised past the old maximum and the answer follows it (habit 7).
  * **Part 12g is now behaviour, not source text.** Real `Controller.cycle()`s on
    a two-pack roster, both packs at the same SoC, with the cut-outs as the only
    thing varied: move the **non-first** pack's cut-out and the floor reaching
    `decide()` moves 20 % → 35 %, with the site's verdict flipping from a sized
    discharge to FLOOR-LIMITED. `self.packs[:1]` answers 20 % both times. Which
    option key the floor comes from stays pinned on the source where it already
    was, because that claim genuinely is about the code's shape.
  * ⚠️ **Two MORE survivors, found by this round's own battery and recorded
    rather than quietly repaired.** (a) The first cut of part 12g varied only
    pack B's cut-out, so `self.packs[-1:]` — grade the site against the **last**
    pack — agreed with the full roster on both runs and stayed green across all
    19 files (+5 bytes). A two-run fixture pins one slice and not the other.
    Part 12g now runs the **same pair of cut-outs in both orders**, so the
    deeper one is on the second pack in one run and on the first in another and
    the site floor is 35 % in both; no single-position slice passes all three
    runs. (b) `binding_soc_floor()`'s empty-roster branch rewritten to
    `return None` (−43 bytes) was also green — a defensive branch nothing in the
    suite reached. Both directions of it are pinned now, with the note that it
    cannot permit anything either way because `pack_floor_soc` is `None` at an
    empty roster.

* ⚠️ **THIRD FIX ROUND — the same family was alive at the SIBLING ARGUMENT of
  the same call.** Rounds 2 and 3 hardened the FLOOR axis; `decide()` is handed
  a second reduction over the same roster beside it, `soc=self.pack_floor_soc`,
  and `peakguard.py`'s own docstring names the two as a pair. Three rewrites of
  `Controller.pack_floor_soc` were green across all 19 files on `52380c0` and
  again after the rebase onto `3934ad7`: `min(float(v) for v in vals)` →
  `vals[:2]` (+4), `vals[-2:]` (+5) and `vals[:3]` (+4), while `[:1]`, `[-1:]`,
  `[1:]`, `[:-1]`, `max()` and the mean were all red. **The suite's separating
  power on the SoC axis stopped at exactly N=2** — the state the floor axis was
  in before round 3. It changes a real answer: a three-pack roster at
  50 / 60 / 11 % with the emptiest pack THIRD answers 11 % shipped and 50 %
  mutated, so the site is graded for discharge PERMISSION against a pack 39
  points fuller than the emptiest and spends a pack its own BMS stops. Round 3's
  machinery transfers directly — new **part 12i**: six real `cycle()`s with the
  emptiest pack at every position of a three-pack roster and both positions of a
  two-pack one (the observable read off `PeakDecision.soc`, cut-outs held flat
  so the floor axis cannot be what moved the answers); computed slice and
  non-slice tables that go red naming any member nothing kills; a **length
  sweep** of 78 real rosters, lengths 1–12, closing head and tail windows of
  width 1–11 and recording width 12 as honestly unreached; a **none-guard
  sweep** over the same 78; and **three exact structural pins** — the `vals`
  comprehension must iterate the bare `self.inverters`, and both `any()` and
  `min()` must reduce the whole `vals` — which close every slice width at every
  roster length. No shipped behaviour changed; only what observes it.
* ⚠️ **SPIKE-04: production is DEGENERATE on the floor axis, so the
  differing-floor fixtures are the entire exposure.** Pack B is 4 × Dyness
  DL5.0C, DoD 90 → floor 10 % — **FROM THE DATASHEET AND THE SERIAL; inverter B
  does not exist and nothing about it is measured.** Pack A is DoD 90 → 10 % as
  well, and that one *is* measured. Both packs therefore floor at the same value
  in the field, `max(10, 10)` is 10 whatever it reduces, and a differing floor
  exists **only in the fixtures**. That is the argument for keeping them, not for
  simplifying them: anyone who collapses `A_PACK_FLOOR != B_PACK_FLOOR_SIM` to
  one number deletes the only data this repository has that can see the
  reduction at all, and part 12c2's precondition check goes red if they do. The
  **SoC** axis has no such degeneracy — two batteries at the same state of
  charge is a coincidence, not a configuration — which is what makes
  `pack_floor_soc` the half of the pair that will be live from the day inverter
  B is wired up.
* ⚠️ **The FAMILY enumerated rather than the member closed, and it found four
  more survivors in code this card never touched.** Every reduction over
  `self.inverters` / `self.packs` in `Controller` was found from the AST and
  mutated to a slice against the full suite. Eighteen were already graded
  (`site_batt_w`, `site_rating_w`, `site_max_w`, `site_dev_bound`,
  `site_circulating_w`, `inputs_ok`, `units`, `Controller.packs`, `maint_unit`,
  the maintenance-due minimum, the store-migration roster, `validate_unit_ranges`
  and the round-2/3 sites). **Four were not, and three of them are SAFETY-30's,
  all `any(i.sign_wd.tripped for i in self.inverters)`** — `cycle()`'s
  stale-input hold, `publish()`'s status vocabulary and `status_payload()`'s
  banner — each +4 bytes and green across all 19 files, for the reason every
  blocker on this card was invisible: **the only fixture that reached them was a
  roster of ONE, where `self.inverters[:1]` IS `self.inverters`.** The first has
  teeth: a latched unit must take the roster past that hold or its `failsafe`
  SET stops being re-filed the moment a poll is missed, which is main.py's own
  sentence for it. New **test_watchdog.py part 8** grades all three with the
  latch at EACH position plus a no-latch counterfactual, and new **part 12j**
  pins all 26 roster comprehensions in `Controller` to iterating the roster
  whole — so a reduction added over a slice anywhere in the class is red on the
  line it is written, with no fixture required. Two rows are reported open
  rather than closed: `_grade_store_migration()`'s second disjunct (now red on
  12j, and argued to be an equivalent mutant because a unit tripped at startup
  implies its namespaced key is set) and `cycle()`'s undeliverable-unit **log
  line**, where `sorted(...)[:1]` (+4 bytes) is still green — it truncates names
  in one event message; the refusal itself is a different expression and is
  graded, so the site still stops.
* ⚠️ **FOURTH FIX ROUND — the same family one MODULE over, and the pin widened
  until it discovers its own scope.** Round 4 enumerated the family and closed
  it inside the `Controller` class; the review found four more survivors
  immediately, in **validate.py**, on a 19/19 green suite: `min(caps)` →
  `min(caps[:1])` (+4) and `→ min(caps[-1:])` (+5) at `:779`, the caps
  comprehension over `caps_ix[:1]` (+4) at `:778`, and
  `min(caps_ix[-1:], key=…)` (+5) at `:893`. **None is an equivalent mutant.**
  With `max_w: 5000` on inverter-a beside `2000` on inverter-b and
  `sign_disagree_w: 3000`, the shipped validator refuses — *"the SAFETY-02
  sign-disagreement watchdog can never fire"* — and all four **accept**: the
  site starts with a watchdog reporting `armed` that cannot trip on pack B.
  Rule 5's own failure mode reached through rule 5; rule 6 loses its refusal the
  same way. **No shipped behaviour changed by the pin; what changed is that a
  hand-written scope stopped being trusted.**
  * **`test_peakguard.py` part 12j now parses every `.py` under `app/`** — no
    module is named in the check — and finds the roster by transitive closure
    from **one seed**, the `inverters` option key: through returns, through
    lists **accumulated in a loop over a roster** (which is what `caps`,
    `roster` and `problems` are, and what round 4's rule could not express),
    through attributes and properties, and **through call arguments**, which is
    how it walks into `peakguard`, `arbiter`, `allocator`, `circulation`,
    `maintenance` and `store` without being told they exist. A slice of a roster
    stays a roster so a mutation cannot cut the taint; an index is one unit and
    does not. **183 reductions across 9 modules** out of 23 files parsed — `main.py` 96,
    `allocator.py` 24, `validate.py` 19, `arbiter.py` 14, `circulation.py` 10,
    `maintenance.py` 10, `peakguard.py` 4, `inverter.py` 3, `store.py` 3.
  * **The hand-written module list would have been wrong in BOTH directions.**
    It named `watchdog.py`, which reduces over no roster at all — SAFETY-30's
    sign watchdogs are strictly per-unit — and omitted `maintenance.py` and
    `store.py`, which both do. A vacuous claim about one module and no claim
    about two others.
  * **The pin rediscovers all four blockers without being told where to look,**
    and finds a fifth beside them (`validate.py:783`, rule 5's message) that no
    review round named. Its counterfactual runs at two depths: every enumerated
    site's own expression mutated with each of the four cuts this card has been
    defeated by (732 spellings, all reported), and `[:1]` written into
    `app/validate.py` at each blocker position with the **whole** analysis
    re-run over the mutated source — a sliced site required at that exact line,
    where the clean package reports none.
  * **And the behavioural half, because a shape pin grades an identity and never
    a meaning.** `test_validate.py` grows a **roster SET** for the three rules
    that reduce over `unit_caps_indexed()` — widths 2, 3, 4, 5 with the one
    small cap at every position, 42 runs of the real validator — plus a
    **computed family table** of twelve candidate reductions each compared with
    what the shipped validator actually said on the same roster. Only `min`
    agrees everywhere; a member no roster kills is printed as a named survivor.
  * ⚠️ **A THIRD RULE WAS READING THE FLAT `max_w`, AND IT IS SAFETY-30's** —
    found by reading, **not** by the pin, which is the shape of what the pin
    cannot see. `unit_sign_disagree_w` was graded against
    `_number(opts, "max_w", 2000.0)` three lines under a branch that had already
    stopped doing that. **Both directions were measured:** two 2000 W units
    beside a leftover `max_w: 5000` and `unit_sign_disagree_w: 3000` produced
    **zero problems** — the site started with a per-unit watchdog that could not
    trip on either unit — and two 5000 W units with no flat key were **refused**
    against a 2000 W default the operator never wrote. It takes the weakest
    unit's own cap now and names that unit; at a roster of one it is the same
    comparison, boundary and sentence it always was.
  * ⚠️ **What the pin cannot see, with sizes**, in
    `DECISIONS/SAFETY-32-per-pack-bms-floor.md` §4d: semantics (`min` vs `max`
    is a zero-byte edit); direct index sites that are not reductions; a
    subscript by a *variable* (`weights[u]` is why); a slice through an
    intermediate call; anything outside `app/`, including `sim/inject.py`; and
    the taint's own seam at `*args` and unresolvable receivers, where everything
    downstream goes **unenumerated rather than reported unreachable**.
    `cycle()`'s undeliverable-unit log line is still green at +4 bytes and still
    filed open, unchanged from round 4.
  * **`tools/supervisor_shape_probe.py` re-run again (2026-09-05, 2026.08.0).**
    The rebase onto `67e649a` merged SPIKE-04a's `max_w: 4600` into the same
    §C/§D blocks this card had rewritten, and a transcript stitched by hand out
    of two runs is not a measurement. `all cases passed`, exit 0, with B's
    rating reading 4600 — its **datasheet** figure, still not a measurement of
    any unit.
  * ⚠️ **THE SWEEP, AND THE NUMBER THAT MUST NOT BE QUOTED ALONE.** All 183
    sites mutated to `[:1]`, one at a time, against the full 19-file suite in
    `goodwe-addon-ci` (`:ro`, `-B`, fresh rsync per mutant, 0 `.pyc` before and
    after): **183/183 RED** — which is tautological on its own, because the pin
    makes every slice red. The 51 whose first red was `test_peakguard.py` were
    re-run with the pin **neutralised**: **145 of 183 have a behavioural fixture
    that separates them from a slice of themselves; 38 have only the structural
    claim.** All 38 are listed by site, with what a `[:1]` there would cost, in
    §4d and in `test_peakguard.py` part 12j — 15 of them are rows and counts on
    the operator's page, 3 are `len(targets)` arithmetic beside a write path
    that IS graded, 6 are allocator internals ALLOC-01's fixtures reach but not
    with enough units to separate. **None of the 38 is closed behaviourally by
    this card, and none is counted as covered.**

⚠️ **No measurement of inverter B appears anywhere in this change.** It does not
exist yet. Every second-pack figure — in fixtures, in the probe, in
config.yaml's commented example — is **simulated and labelled**; only inverter
A's 10 % is read off hardware.

**At two inverters the sign watchdog stops comparing one unit against the whole
site meter, and becomes two checks instead of one (SAFETY-30).**
`disagrees(commanded_w, grid_w, ...)` compared the command **this controller put
on ONE actuator** against **what the whole site** is exchanging with the network.
At a roster of one those are two views of the same thing - §6.2's identity - and
that is what makes the check sound and what earned it the best record in the
safety table. At two they are not, and it fails **in both directions**: inverter
A commanded −1500 W while the meter reads +1500 W is evidence of nothing when
inverter B is discharging 3000 W exactly as told, and - worse - a unit genuinely
driving the wrong way is **MASKED** by a sibling that covers for it, because the
site figure nets. No threshold recovers that; the summation destroys the
evidence.

* **PER UNIT: each inverter's own share against its OWN measured power**
  (`Inverter.batt`, holding register 0x0518, per-instance since DEV-02).
  Conclusion: *the actuator I am driving is not doing what I told it* - evidence
  about one device, so the remedy stays the per-unit `failsafe` SET that
  `arbiter.CLAIM_SCOPE` has classified `SCOPE_UNIT` since it was written. That
  entry read *"the evidence is site-wide and the remedy is not"*; after this card
  the first half is simply true. Verified by making one unit reverse while its
  sibling covers it at the meter: the shipped comparison is false on **0 of 14**
  cycles while the new one convicts that unit on every one, and only that unit is
  parked.
* **SITE: the SUM of the standing commands against the meter**, as a new
  `sign_site` SET at `P_SAFETY` on the site actuator, latched under flat
  `site_sign_tripped` / `site_sign_reason`. This is the fault the 2026-08-20
  capture contains and **no per-unit check can see it**, because every inverter
  is faithfully executing its share - what is wrong is the number they were
  given. It stops every inverter and the banner names the site rather than an
  inverter, so nobody is sent to swap the wrong box.
* **⚠️ Different questions from SAFETY-31, and the PR and the decision record
  both say which.** `circulation.py` compares the units **against each other**
  and catches two inverters pushing *through* each other, which nets to zero at
  the meter and which no single command is evidence about. This compares each
  unit against **its own command**. They overlap in exactly one scenario - a unit
  reversing while its sibling pulls the other way - and `test_watchdog.py`
  asserts on that fixture that both fire and what each concludes. It is also not
  the saturation freeze: that is a magnitude test (`|command − measured| >
  500 W`) that shapes the loop and fires on a merely saturated unit; this is a
  direction test that latches and parks one.
* **⚠️ All four thresholds RE-DERIVED at two inverters - one carried with a
  justification, one refused.** 1890 closed-loop runs in which nothing is faulty,
  over the whole range `config.yaml` accepts for `gain` and `slew_w`, across load
  reversals with write staggers and flapping loads, **run at both rosters so the
  N=1 arm is the control.** `unit_sign_disagree_w: 250` is `saturation_w / 2`, so
  the smallest divergence this can see is exactly the gap at which the control
  law itself already says the inverter is not following; `unit_sign_disagree_s:
  85` is 8.5× the worst legitimate opposition those runs produce (10.0 s), 2.8×
  the firmware's own 30 s link veto and under half the site hold.
  `site_sign_disagree_w: 300` is **carried** - its argument is the meter's noise
  floor and the deadband, and the roster touches neither. **`site_sign_disagree_s`
  is NOT: 180 s would have shipped a false trip.** The worst legitimate crossing
  at 300 W is 110 s with one inverter and **215 s with two**, because the command
  must travel a rail that is the roster sum at a `slew_w` that is not - reachable
  at `gain: 0.05`, `slew_w: 50` and a 2.2 kW load reversal, all schema-legal.
  325 s is 1.51× that, inside the 900 s billing quarter and inside every episode
  of the fault day. See `DECISIONS/SAFETY-30-per-unit-sign.md`.
* **⚠️ At ONE inverter this changes nothing, and "nothing" is asserted rather
  than assumed.** The two questions are the same sentence at a roster of one, so
  only the detector that shipped is armed: same comparison against the meter,
  same 300 W / 180 s, same reason string byte for byte, same namespaced latch and
  flat rollback mirror, same `failsafe` claim, same banner, same single status
  row - and the site detector's window is never so much as opened.
  `unit_sign_disagree_*` and `site_sign_disagree_*` do nothing on a
  one-inverter install.
* **The threshold WIRING is pinned, not just the threshold values.** Review
  found that `arm(self.sign_cfg if one else self.unit_sign_cfg, ...)` could be
  collapsed to `arm(self.sign_cfg, ...)` — arming every unit with the
  meter-facing pair instead of the re-derived per-unit one — with all 19 test
  files still green, because nothing at a roster of two read `sign_wd.cfg` and
  the trips in the fixture happen under either pair. `test_watchdog.py` part 3
  now asserts the pair and the instrument on **every** detector against
  literals, and the fixture's three threshold pairs were made pairwise distinct
  in both coordinates so that a hold delivered from the wrong pair cannot
  compare equal to the right one. M14–M16 in
  `DECISIONS/SAFETY-30-per-unit-sign.md` §5.
* **Known and recorded rather than fixed:** the site comparison is defeated by a
  square-wave load whose half period is about one control cycle, which holds it
  opposed for ~1000 s at *both* rosters. The N=1 control arm is what says this is
  a property of comparing a lagged command against a meter and has nothing to do
  with the second inverter, so it is not acted on here - it wants an alternation
  test, not a number, and a card of its own.

**An operator can declare a second inverter, and the shape was measured against
the Supervisor rather than assumed (CFG-04).** DEV-02 made rating, SoC,
watchdogs, heartbeat and store keys per-instance, so the code no longer assumed
one inverter — but the options schema still described exactly one, and
`build_roster()` returned a list of length one and said so in its own docstring.
The roster could only ever be a roster of one. **Version bumped to 0.6.0,
because the Supervisor keys the built image by `version:` and a schema change
without a bump is a schema nobody reads.**

* **`inverters:` — a list of objects, defaulting to `[]`.** Empty is not
  "unconfigured", it is **the single-inverter configuration**: the flat
  `setpoint_entity` / `soc_entity` / `batt_entity` / `batt_invert` / `max_w`
  keys describe the one unit exactly as they always did. An install that
  upgrades reads identically — same entities, same rating, same behaviour, no
  operator action, and the whole existing check suite runs against it unchanged.
* **⚠️ The shape was ESTABLISHED, not assumed, and the empty default is the
  load-bearing part.** There was no list-of-objects anywhere in `config.yaml`
  and `static_profile` carries its structure as a JSON string in a `str?`
  precisely because a nested schema was doubted. So Supervisor **2026.08.0**'s
  own `AppOptions` validator — the version this site runs — was run against this
  repository's `config.yaml` out of the release image
  (`tools/supervisor_shape_probe.py`, committed; the transcript and the
  reasoning are in `DECISIONS/CFG-04-inverter-roster.md`). It accepts the list,
  enforces per-entry types, refuses a missing `name` / `setpoint_entity` /
  `max_w` **by name**, and renders the option in the UI as a repeatable section.
  And the shipped `options:` block validates as it stands while the **same
  block with only `inverters` removed** is answered with **`Missing option
  'inverters' in root`** — one key varied, so the refusal is about that key —
  which is every options blob written before this version, since the key did not
  exist when they were written. `App.options` is
  the shipped `options:` block deep-merged with the operator's stored one, so
  `inverters: []` in `config.yaml` is the difference between an upgrade nobody
  notices and an add-on that cannot start with a schema pinned to the version
  that would fix it. That is `docs/deploy-rollback.md` §4's trap in the forward
  direction, found before it was sprung rather than after.
* **Nothing is inherited by an entry that omits a key, and that is DEV-02
  restated as a config rule.** `name`, `setpoint_entity` and `max_w` are
  required of every entry; a missing `soc_entity` means *this unit has no SoC
  sensor*, never *use the other one's*. A three-line fallback to the flat keys
  would have rebuilt the sibling-grading defect DEV-02 spent a card removing, in
  YAML. Defaults still live in exactly one place (CFG-03): there is no
  per-entry default anywhere, because there is no optional per-entry number.
* **SAFETY-28's `actuator_writers()` refuses two inverters naming one
  `setpoint_entity` — the first configuration in this add-on's life in which
  that was expressible.** The census was already structural and case-folding, so
  it needed no change; what it needed was a config that could express the
  hazard, and now there is one. Its sibling refusal is new: a flat
  `setpoint_entity` left set beside a declared roster names an actuator nothing
  will write, and is refused rather than ignored.
* **⚠️ Two PERMISSIVE READERS in shipped code, found and fixed.** Both failed in
  the direction that hides a defect. (1) `validate.py` rules 5 and 6 compared
  `sign_disagree_w` and `maintenance_charge_w` against the **flat** `max_w`; at
  a roster of two 2000 W units with `max_w: 5000` left over from a
  single-inverter install, `sign_disagree_w: 3000` passed rule 5 while
  SAFETY-02's watchdog could fire on **neither** unit — rule 5's own failure
  mode, reached through rule 5. They now read the roster's caps
  (`unit_caps()`), which at one inverter is `[max_w]` exactly. (2)
  `run_control()` re-ran the device-rating check behind `if
  self.o.get("setpoint_entity")` — the flat key, which a roster config leaves
  blank — so on a two-inverter site **no** inverter's `max_w` was ever graded
  against the device it is written to. `validate_unit_ranges()` grades each unit
  against its own ESPHome and runs unconditionally.
* **⚠️ Three config.yaml READERS IN THE CHECKS took the first match on a
  stripped line, and one of them was already known.** `test_peakguard.py`'s
  `read_option()` and `_schema_bound()` and `test_inverter.py`'s
  `read_schema_max()` threw indentation away, so the nested per-inverter
  `max_w:` began outranking the flat one — **silently, and correctly**, because
  the two bounds are the same `int(100,5000)` today. A reader that is right by
  coincidence is the failure mode this project is named after. All three are
  anchored at exactly two spaces now, the halves of the file disambiguate
  `peak_cap_w` instead of first-match, and an ambiguous key raises rather than
  picking one. `test_tariff.py`'s copy of the idiom went with them.
* **The audit is a check, not a promise.** `test_inverter.py` part 11 finds
  every `config.yaml`-reading function in the suite — transitively, so a reader
  that never names the file because it delegates is still found — **executes**
  each one against a synthetic two-inverter `config.yaml` whose per-inverter
  numbers differ from its flat ones, and fails if any returns the per-inverter
  value. The table is not maintained by hand: the signature decides, so the
  eleventh reader is named on its first run. `config.yaml` declares `inverters:`
  **before** `max_w:` on purpose, and part 11 asserts that ordering, so the trap
  stays baited rather than sorted out by an ordering nobody would maintain.
* **Ten mutations run and observed red** (`tools/mutation_run.sh`), listed in
  `test_inverter.py` part 12 with the check that bites each. **Two survived the
  first run and the checks were changed, not the table:** the inheritance check
  had blank flat keys, so "inherited" and "not inherited" were the same empty
  string; and the `run_control()` check asserted a **source line** was present,
  which re-indenting it does not remove. The second is now driven through the
  shipped `run_control()` under a deadline, because that mutation's real symptom
  is a hang rather than a red.

**Circulating power - two inverters pushing against each other - becomes a
measured quantity, a bounded one, and an alarm that survives a restart
(SAFETY-31).** Inverter A charging at 3 kW while inverter B discharges at 3 kW
nets ~0 W at P1: the site reads idle while the packs are cycled through two
round-trip conversion losses continuously. **Every guard in this add-on is a
consumer of that net figure**, so none of them can see it - SAFETY-02's sign
watchdog gates on `abs(grid_w) > 300` and the gate never opens, and both the
peak guard and the capacity tariff bill grid exchange.

* **A site sensor, `Σ|Bᵢ| − |ΣBᵢ|`.** Zero when the units pull the same way,
  `2 × min(|B_a|,|B_b|)` when they oppose. Published as
  `sensor.goodwe_circulating_power` beside the grid figure, because *"grid 0 W,
  circulating 6000 W"* is the whole phenomenon in two dashboard rows and no
  arithmetic on the grid figure alone can produce the second one.
* **⚠️ Derived from MEASURED per-inverter power, never from the allocation
  record, and that is the whole card.** ALLOC-01's sign-coherence gate
  (`REFUSE_SIGN`, its AC 7) already refuses any allocation whose shares carry
  two signs, so a record-derived sensor reads zero **by construction** and could
  only ever confirm that the allocator agrees with itself. What this detects is
  circulation **nobody commanded**: a unit ignoring or lagging its share, a
  stale share still being executed, a maintenance or SAFETY-25 pin holding one
  unit while the loop drives the other, or HW-01's power stage running its last
  instruction for **6 h 25 m** with the comms dead and the allocation record
  impeccable throughout. Verified by making one unit ignore its share on the
  TEST-19 rig: the record stays `ok` with both shares at −1250 W - a
  record-derived sensor reads **0 W** - while the shipped sensor reads
  **2000 W**. `Inverter.batt` became per-instance in DEV-02; before that this
  sensor could not have been written.
* **The trip stops the site, and the decision record says why that is an action
  rather than theatre.** Sustained opposition files `Claim.set("circulation",
  P_SAFETY, 0.0)` on the **site** actuator - every inverter to 0 W, latched.
  HW-01 established that a wedged inverter cannot be commanded at all, so the
  remedy is not aimed at it: a circulating loop needs **two** participants and
  **one of them is this controller**, faithfully driving the listening sibling
  into the one that stopped listening. Stopping halves the loop, and - the more
  important half - **makes the residue visible at the meter**, where every
  existing guard can finally see it. What it cannot do is stop the wedged unit;
  the notification says so, and names the isolator as the last actuator.
* **The threshold and the window were measured, not chosen (AC 3).** 924
  closed-loop runs on the TEST-19 two-master rig in which *nothing* is faulty,
  swept over the whole range the schema accepts for `gain` and `slew_w`. Worst
  **continuous** legitimate opposition by threshold: **505.0 s at 100 W, 40.0 s
  at 250 W, 10.0 s at 300 W, 5.0 s at 400 W and above**. The knee is 400 W and
  it is the schema floor. `circulation_w: 700` is 1.75× it, the headroom bought
  for inverter B's ramp and lag being **placeholders until SPIKE-04 measures
  them**. `circulation_hold_s: 75` clears three bounds - 15× the worst
  legitimate episode (5.0 s), 2.5× the firmware's own `LINK_STALE_S` veto, and
  12× inside the 900 s capacity-tariff quarter. **A momentary opposition while
  one inverter slews through zero is legitimate** - the serialised write path
  lands B's number up to a control cycle after A's - and the guard does not fire
  on it, in any of the 924 runs.
* **The alarm survives a Home Assistant restart (SAFETY-34).**
  `persistent_notification` is in memory, which is exactly how the 2026-09-02
  outage went unseen for 6 h 25 m with the detector correct throughout. The
  latch is persisted to `/data` as flat, site-level keys - flat because the
  quantity is defined over the roster and is identically zero for any one unit,
  so there is no inverter to attribute it to - and the notification is
  **re-raised on the first cycle after every start**, through the same code path
  as a fresh trip.
* **At one inverter the sensor is identically zero and nothing changes (AC 5).**
  `|B| − |B| == 0` for every reading, including one nobody took; no claim is
  ever filed, the arbitration is byte-identical, and the status page says *n/a*
  rather than claiming to watch something that cannot happen.
* **Verified by breaking it.** 18 mutations against the sensor, the window, the
  latch, the claim, the store, the alarm, the routing table and both config
  copies - **all red**, each on a named check; one equivalent mutant argued
  rather than claimed; and one genuine **survivor found and closed** (`publish()`
  turning an unknown into a published `0 W`). `arbiter.CLAIM_SCOPE` gains its
  ninth source and `test_peakguard.py`'s 10f/10g tripwires were updated
  deliberately, as their own comments require. Full ledger in
  `DECISIONS/SAFETY-31-circulating-power.md`.

**Every per-device value stops being a Controller scalar: rating, SoC, battery
power, watchdogs, heartbeat, write bookkeeping and store keys are now
per-instance (DEV-02).** `app/inverter.py` introduces an `Inverter` object that
owns its entities, its rating, its device range, its `SignWatchdog`, its
`last_write` and its `last_beat`, under **store keys namespaced by unit name**.
The Controller holds a list of them and keeps only what is genuinely site-level:
the P1 reading, the site integrator, and the tariff quarter.

* **What the scalars were silently asserting.** One `dev_min`/`dev_max` pair was
  read from one `setpoint_entity`, so a second inverter's rating was **never
  read at all** and its clamp, its freeze rail and the peak guard's
  `achievable_w` were the first one's numbers. One `soc` fed the maintenance
  FSM, SAFETY-24's pack floor and arbitrage's bounds. One `SignWatchdog` - whose
  own docstring says *"one instance per actuator"* - latched under **flat** store
  keys, so a trip caused by B latched A across restarts and one operator press
  cleared both. One `last_beat` and one serialised `hass.call()` meant a hung
  write to A ate B's 30 s layer-1 budget.
* **`actual_w` became SITE level while the readings became per-instance, and
  those are opposite directions on purpose.** `compute()`'s anti-windup test is
  `abs(prev_w - actual_w) > saturation_w` with `prev_w` the **site** target, so
  feeding it one inverter's power compares a site number against a unit number,
  trips the saturation freeze permanently and leaves a standing grid error.
  Measured on `admin/goodwe`'s two-master rig, 2 kW load: with one unit's power
  the aggregate trace at the meter is split-dependent by **510 W at `gain: 0.2`,
  250 W at 0.4 and 50 W at 0.6**; with the site sum **every split is
  bit-identical at every gain the schema accepts**. That was ALLOC-01's
  disclosed finding, graded in `test_alloc.py` part 8 and now closed there.
* **The store migration is graded, not merely performed.** `sign_tripped` /
  `sign_reason` move to `unit/<name>/...` once, verified by **reading back**
  what was written, and the version marker is only set when nothing that carried
  state was lost. The flat keys are left in the file so a rollback still finds
  its latch. And `Controller.__init__` compares the pre-migration flat latch
  against what unit A actually came up with: if the latch did not survive it is
  **re-armed from the legacy value and shouted about**, because a latch that
  silently fails to migrate is a disarmed safety guard on an install whose state
  file says it is protected.
* **Heartbeats are independent, and the budget is derived.** Each inverter owns
  its own deadline, the refreshes are dispatched **concurrently**, and each write
  is bounded by `WRITE_BUDGET_S` (4 s). The inequality that fixes that number is
  graded rather than asserted: `4 + heartbeat_s`'s schema maximum of `25 < 30`,
  layer 1's `FIRMWARE_INGEST_VETO_S`. Verified by injecting a transport that
  accepts and never answers: B's write reaches the wire while A is still
  hanging, and the serialised shape it replaces is run beside it.
* **`live_units()` answers honestly.** A unit whose every write has been rejected
  for a whole layer-1 window is not commandable; the allocator drops it, the
  survivors carry its share, the action log names it, and the dropped unit is
  commanded **0 W** rather than left standing on its last share. A roster entry
  with **no** `setpoint_entity` is deliberately *not* handled there - that is a
  configuration that cannot be delivered, and it reaches main.py's refusal
  branch, which commands 0 W and names the inverter.
* **At N=1 nothing moves.** The removed scalars survive as read/write properties
  of the sole inverter, so every existing replay in this repository runs
  unmodified - and they **refuse at N>=2** rather than answering for the first
  inverter, which is the tripwire on a consumer that was never made
  per-instance. `test_inverter.py` is the new file (77 checks); `test_alloc.py`
  part 8, `test_peakguard.py` 10g/10h, `test_watchdog.py` and `test_clamp.py`
  were re-aimed rather than relaxed.
* **Inverter B still does not exist.** `build_roster()` returns exactly one
  inverter, from the flat options, and is the one function CFG-04 changes. Every
  two-unit figure above is architecture, not a measurement of a second device;
  the second rating in the checks is marked as a hypothesis.

**The write path is gated on the allocation record, not on how many config keys
exist (SAFETY-28).** specs §1.5 makes *"any configuration in which two writers
can reach one actuator"* a non-goal. Until now that non-goal was enforced by
`len(validate.ACTUATOR_WRITE_KEYS) == 1` - there is exactly one option key, so a
second writer is not expressible in the schema. True, and a statement about **the
shape of the options schema rather than about the writes**: add a per-inverter
config (CFG-04) and there are two real writers with that length still 1, so the
check stays green while the property it protects is gone. `sim/inject.py`'s
`two_writers` scenario asserted on the same tuple and went green with it.

* **The enforcement moved onto the writes.** `write_setpoint()` now refuses any
  value that is not in **this tick's** allocation record from ALLOC-01:
  `Controller.write_refusal()` asks `Allocation.authorises()` whether the single
  allocator placed exactly this number, for this unit, on this tick. A second
  writer with arithmetic of its own cannot satisfy that - its number appears in
  no record - so it has to file a `Claim` and go through `resolve()`, which is
  the one allocator. "Exactly one allocator per shared resource" stops being a
  convention and becomes something the write path checks.
* **A refusal is a refusal, not a log line.** The unauthorised value is replaced
  with **0 W and written**. Returning instead would leave the inverter latched
  on its last command with the ESP32 heartbeat still being fed - holding an
  unauthorised setpoint quietly, which is the outcome the gate exists to prevent.
* **A stale record cannot be replayed.** Every record now carries the `tick` it
  was made for, handed *in* by the caller so `allocator.plan()` stays the pure
  function `test_alloc.py` part 10 grades. The counter advances **once per
  resolution, not once per loop iteration**: a new allocation makes the previous
  record stale immediately, while the "inputs missing, not yet stale" hold -
  which returns before the allocator runs, so one dropped meter reading does not
  zero the inverter - leaves the standing authorisation intact for the heartbeat.
  An **unstamped** record authorises nothing; "I cannot say when I was made"
  fails closed.
* **⚠️ 0 W IS AUTHORISED UNCONDITIONALLY, SO THE FAILSAFE CANNOT DEADLOCK BEHIND
  THE GATE.** `write_setpoint()` is also how layer 2 commands zero, and every
  path that does so - `cycle()`'s two refusal branches, `shutdown()`, the
  operator's stop button - runs while the allocator has *refused*. A gate that
  asked the allocator's permission to command zero would block the safety write
  behind the very component whose failure the safety write answers. The
  exemption is on the **value**, not on a `failsafe=True` flag a caller asserts,
  and the argument is that the hazard §1.5 names is N-times loop gain from
  *broadcast*: at `S = 0` the broadcast and the correct allocation coincide -
  `a_i = 1` for every unit gives every unit 0 W, which is exactly its share - so
  there is no over-command to make at zero. Driven end to end through the real
  refusal branch in `test_clamp.py` part 7c.
* **`unmet_w` is still not `ok`.** The gate reads `ok` and the unit's share and
  never `unmet_w`, because an allocation that could not place every watt is
  still safe to write; fusing them would let one saturated inverter park the
  whole site. Graded on a record with 3800 W unmet.
* **`ACTUATOR_WRITE_KEYS` is replaced by `actuator_writers()`**, which counts
  **writers per actuator** rather than option keys, and finds them by walking
  the options tree rather than by looking for a container name CFG-04 has not
  chosen yet. Two inverters with two entities are accepted - that is what the
  per-inverter roster is *for*; two declarations aimed at one entity are refused,
  naming the actuator and every writer pointed at it. The tuple survives only as
  the enumeration a census needs, and nothing asserts on its length any more.
* **What neither check observes, said out loud.** The census sees *configured*
  writers, so a second write path added in **code** is invisible to it - exactly
  as it was invisible to the tuple. That half is closed at runtime by the record
  gate, which does not care what the options look like or who called the method.
* Verified by breaking it: seven mutations on a purged tree via
  `tools/mutation_run.sh` - removing the gate, making the zero exit an amnesty,
  dropping the tick comparison, deleting the census refusal, fusing `unmet_w`
  into `ok`, freezing the tick on the live path, and moving the gate below the
  clamp - each turn `test_clamp.py` or `test_validate.py` red.

**The site grid target is now ALLOCATED across N inverters, sum-invariant, and
never broadcast (ALLOC-01).** `ARB-03` left `allocate()` as a documented seam
that passed through at N<=1 and DECLINED at N>=2; this fills it. The new
`app/allocator.py` turns one site target `S` into per-inverter shares with one
invariant - **`sum(shares) + unmet_w == S`, in integer arithmetic** - and each
inverter is then told a function of **its own share**, never of the shared
measurement.

* **Why a broadcast is a fault and not merely an inefficiency.** The control
  primitive is meter emulation: the master reports a *fabricated grid reading*
  and the inverter regulates itself to it, so one number means "the battery
  moves X W" here and "the grid is at X W" there - an identity that holds only
  at N=1. Hand both inverters the same target and each supplies it in full, so
  the loop gain at the meter becomes N times the per-unit gain. Measured on
  `admin/goodwe`'s two-master rig (TEST-19): **-400 W export overshoot at
  `gain: 0.6`, and at `gain: 1.0` - which `config.yaml` accepts as its
  documented maximum - a permanent 4000 W peak-to-peak limit cycle with not one
  guard firing.** The sign watchdog is evaluated on every cycle and never trips,
  because the disagreement flips on alternate ticks and never holds its 180 s
  window.
* **The stability property is structural, not tuned.** With `sum(a_i) = 1` the
  gain at the meter is `sum(a_i * g_i)`, a **convex combination** of the
  per-unit gains, so it lies in `[min g_i, max g_i]` **for any split**. That is
  what makes the weights a comfort choice (SoC, thermals, wear) rather than a
  tuning parameter, and it means a bad split cannot destabilise the site.
  Graded directly in `test_alloc.py` part 9, and measured end-to-end: a 90/10,
  a 50/50 and a 100/0 split produce a **bit-identical aggregate trace at the
  meter**, and a gain sweep to 1.0 at every split shows a flat tail at all 84
  combinations.
* **Rounding is a safety property here.** The register is 1 W, and a per-unit
  `round()` biases the sum - measured at **-125 W over 50 000 ticks at three
  inverters**. At 1-2 Hz that is a permanent standing error the site integrator
  chases and never reaches. So the target is rounded **once**, at the top, and
  the fractional parts are settled by **largest remainder** with a deterministic
  tie-break; every sum below that line is integer arithmetic and cannot drift.
  Verified over 250 000 randomised ticks: **0 sum failures, 0 envelope escapes**,
  and the one rounding that is left - the single site-level `quantisation_w`,
  never distributed into a per-unit bias - totals **-39.8 W over 137 964
  allocations, i.e. -0.0003 W per tick.**
* **Each share is clipped to that unit's interval, and THE TWO ENDS OF THE
  INTERVAL ARE APPLIED INDEPENDENTLY.** `UnitBounds()` defaults to
  `(-inf, +inf)`, so a one-sided envelope is the ordinary case at N>=2 rather
  than an edge. The first cut of this work folded both ends into one
  `min(max(v, lo), hi)` and substituted `v` itself for a missing bound, which
  reads as "an absent bound cannot bind" but is not what it does: with
  `hi == inf` the expression collapses to `min(max(v, lo), v) == v`, so **a
  missing ceiling silently discarded the floor.** `_clamp_int(0, 2142, inf)`
  returned 0. The worked failure it caused: a unit with a 2000 W floor and no
  ceiling, beside a symmetric unit, against a +1000 W site target, was allocated
  **+500 W - a quarter of its own floor** - and the sign-coherence refusal that
  the same case with a finite ceiling correctly raises was **suppressed** with
  it. Fixed by clamping each end on its own, and `UnitPlan.explain()` can no
  longer print a share as being inside an interval that excludes it: that line
  now reads `OUTSIDE ... - ENVELOPE VIOLATED, DO NOT WRITE`.
* **The envelope is read per tick, never cached.** `plan()` is a pure function
  of that tick's intervals and holds no state, because a unit's range moves for
  reasons unrelated to the allocator - a stale link, a pack at its floor, a
  device rating that failed to read.
* **A dead link is announced, not absorbed.** A unit whose master is stale is
  dropped from the envelope set; the survivors carry its share, and the
  allocation record names the unit and the reason so `main.py` logs it. A site
  running at half its inverters and reporting nothing reads as healthy on every
  dashboard.
* **Opposite directions are refused, not averaged.** One pack discharging while
  another charges nets to exactly the right number at the meter, so no existing
  guard can see it. A sign-coherence gate makes it inexpressible.
* **`scope_violations()` is now consumed, and it is consumed as a refusal.**
  ARB-03 left it logging without changing the resolution. A mis-routed claim
  means the number about to be divided may itself be wrong, and dividing a
  number that may be wrong is worse than not dividing it - the same answer
  clause 3 already gives a contradiction. It is *not* promoted into an envelope:
  a unit-scoped claim that landed on the site names no unit, so promoting it
  would mean guessing one.
* **At N<=1 it is a pass-through, byte-identical.** The share is the site
  target, as a float, unrounded and unclamped. Every one of the existing suites
  replays unedited and green.
* **`main.py` still refuses to act at N>=2, for a new and narrower reason.**
  There is one `setpoint_entity`, so this controller can command one inverter; a
  valid two-way split names an inverter it has no writer for, and commanding a
  partial site is a worse instruction than commanding nothing. The old
  "no allocator" refusal and its test (`test_peakguard.py` part 10h) were
  **re-aimed, not deleted** - `allocated=False` is now reachable on a mis-routed
  claim, an empty envelope set, or sign incoherence, and all three are graded on
  real cycles.
* **And the replay that missed it was widened, so the fix is graded rather than
  merely applied.** The 250 000-tick replay drew every envelope as
  `B(-uniform(0, rating), uniform(0, rating))` - **finite on both ends and
  containing zero, every time** - while the check it fed claimed "no share ever
  leaves the envelope it was clamped into". The data could only exercise a
  symmetric closed interval around zero, so it reported 0 escapes against a live
  defect. The generator now draws from **every shape `UnitBounds` can express**:
  half-open above, half-open below, intervals that exclude zero, and bare
  `UnitBounds()`. Measured A/B against the defective clamp on the same seed and
  the same 250 000 ticks: **the old generator sees 0 escapes, the new one sees
  39 991.** Reverting the clamp fix under the widened generator takes
  `test_alloc.py` red on four checks; re-applying it takes it green. The
  collapse is now a row in part 12's mutation table.
* **Known and not closed here:** at a roster of two the loop's `actual_w` is one
  inverter's power used as a site measurement, because the add-on binds a single
  `inverter_power` sensor. That is the *sensing* half of the same N=1 identity
  this card removes from the *commanding* half, and it makes the aggregate trace
  split-dependent below `gain: 1.0`. Measured, with the counterfactual, in
  `test_alloc.py` part 8 and `tools/alloc_sim.py --splits`. It belongs to
  `DEV-02`.

**`Claim` carries the `actuator` specs.md always said it did, and `resolve()`
is three-stage (ARB-03).** `specs.md` §6.1 declares `actuator: str` as the first
field of `Claim`; the shipped `app/arbiter.py` had no such field, so every claim
was implicitly about "the one and only actuator", `resolve()` returned one
scalar, and `main.py` assigned it to one `self.target`. **There was no seam an
allocator could occupy** - and `ARCH-01` decided the site grid target is
ALLOCATED across several inverters and never broadcast. This is a regression fix
against the written design, and it creates that seam. **At N=1 it is a pure
refactor: every one of the 17 suites replays unedited and green, on the same
numbers.**

* **The split, verified at each construction site rather than assumed.** Four
  sources are about the **grid connection** and are SITE claims - `loop`
  (regulates `grid_w`, the P1 measurement), `peak` (the quarter-hour average
  offtake the capacity tariff bills), `arbitrage` (buys and sells at the grid
  against a day-ahead price) and `safety` (`stopped` is the site's switch,
  `inputs missing` is the site's own meter). Four are about **a pack, a device
  or one write** and are PER-INVERTER - `device` (one inverter's own reported
  rating), `supervised` (`max_w` caps what may be WRITTEN to one setpoint
  entity), `failsafe` (**site-wide evidence, per-unit remedy**: SAFETY-02 grades
  the command THIS controller put on THIS actuator, and at N>=2 that evidence
  cannot be pinned on a sibling nobody here commands) and `maintenance` (§12's
  recalibration is a pack procedure). The reasoning is recorded **at the line**,
  in `arbiter.CLAIM_SCOPE` and again at each `claims.append` in `main.py`.
* **Getting it wrong is the parked-battery bug, so it is routed rather than
  written.** A per-inverter reserve filed as a site `hi` against the peak guard's
  site `lo` is `lo > hi`, arbiter clause 3, **0 W in the billed quarter** -
  SAFETY-25 §0a's fatal case. `actuator_for()` is the one routing point and it
  **raises `KeyError` on an unclassified source** rather than defaulting to the
  site, because defaulting to the site is the failure itself.
* **Three stages, each a function a test can construct a case at.**
  `pre_resolve()` turns one inverter's own claims into that unit's interval;
  `site_resolve()` turns the site claims into one grid target and **is the
  shipped arbiter, unchanged**; `allocate()` is **the seam, and deliberately not
  an allocator** - `ALLOC-01` builds it. With at most one unit it passes the
  target through (§6.2's N=1 identity); with two or more **it declines, no
  shares, `allocated=False`**, because the obvious placeholder - hand the target
  to each - IS the fault `ARCH-01` was filed about. **What `main.py` then does
  is SUBSTITUTE, not abstain**: it replaces the resolution with a 0 W `failsafe`,
  logs "refusing to broadcast the site target", and writes that 0 W through
  `write_setpoint()` like any other cycle. The refusal is to BROADCAST; the
  write still happens, in the fail-toward-inaction direction. Graded end to end
  on a real two-inverter cycle by new **part 10h** - the branch is unreachable
  on the shipped roster, and replacing it with `if False:` had left all 17
  suites green.
* **`resolve()` HONOURS the roster it is handed, and that is now graded too.**
  Left `None` the roster is derived from the claims, which is the documented
  default; passed explicitly it is used. Deriving it *always* - ignoring the
  argument - makes a unit that filed nothing this cycle cease to exist, so a
  site of two reads as a site of one and the seam passes the whole grid target
  through to the inverter that spoke: `ARCH-01`'s fault with its own evidence
  removed. `test_arbiter.py` part 7 grades it on the one case that separates the
  two, a **silent unit**.
* **The contradiction is no longer guarded against, it CANNOT BE WRITTEN DOWN.**
  The partition is by `actuator` and it is a partition: a unit's `hi` and the
  site's `lo` are never both terms of one `lo > hi` test. Constructed and swept -
  40 filings (every per-inverter source x unit x bound, a unit contradicting
  ITSELF included) against a live site floor, and not one moves the site target
  or raises a contradiction. A unit's own impossible pair parks that unit and
  says so on its own bounds; it does not park the site. That is SAFETY-25 §0c's
  table made mechanical.
* **At N=1 nothing splits, and that is §6.2's identity rather than the
  classification being ignored.** With one inverter the fabricated grid reading
  and the battery setpoint are the same number, so the grid connection and the
  only unit ARE one actuator and every source routes to the site. **Measured:
  dropping that identity - splitting at N=1 - takes `test_peakguard.py`,
  `test_freeze.py` and part 10g red**, because the `failsafe` and `maintenance`
  SET claims would stop competing for the winner, `explain()` would name a
  different source and `loop_is_following()` reads that name.
* **SAFETY-24's floor asymmetry survives, and now has a second half.** The
  combination sweep is re-run and still holds - no negative floor, `hi` still
  `+inf` on every claim the guard files - and it now also asserts every one of
  them is on the **site** actuator, because a floor bound to one inverter's share
  would leave the site, the thing the capacity tariff meters, with no floor.
* **The part 10f tripwire is kept and widened, not adapted around.** Its static
  AST scan and its dynamic SoC sweep are untouched in substance - `source` is
  still `Claim.set()`'s first positional argument precisely so the scan keeps
  reading it, which is why `actuator` is the LAST field of `Claim` and not the
  first. Its spy gained `**kw` so the roster passes through unread rather than
  being silently dropped. New **part 10g** pins `CLAIM_SCOPE` to the same
  eight-source set, so a ninth source now goes red twice. Verified by breaking
  it: building SAFETY-25's Option B as a ninth source still takes 10f red on the
  line it is written.
* **And the CALL SITES are pinned statically, at all eight of them.** 10g's
  first cut graded routing only through a **real two-inverter cycle** and called
  that "the line a mis-typed `actuator_of(...)` at a call site goes red on".
  **It was not.** That cycle reaches `device`, `supervised` and `loop` and
  nothing else - three of the eight direct appends - so a mis-type at the other
  five survived every suite in this repository. Proved on the shipped tree by PR
  #40's reviewer: `actuator_of("maintenance")` -> `actuator_of("peak")` at the
  maintenance-shaping append left `test_arbiter.py`, `test_peakguard.py` and
  `test_arbitrage.py` all green while, at a roster of two, it built a site loop
  of -1500 W against the guard's site `lo` of 1500 W and the now-misrouted
  charge-only `hi = 0` - `lo > hi`, arbiter clause 3, **0 W in the billed
  quarter**, SAFETY-25 §0a's fatal case from a one-word typo. That is this
  repository's named failure mode - **a check observing something other than
  what it claimed** - reproduced inside the check written to prevent it, so it
  is closed the way 10f closes it: over the **same AST scan**. All eight direct
  appends expose their actuator as a literal `actuator=self.actuator_of('...')`,
  so the literal is required to EQUAL the source the claim is filed under - all
  eight, on the line the typo is written, in every configuration and at every
  roster, whether or not any fixture reaches the state that files them.
  **Measured: each of the eight mutated to a wrong actuator in turn, each one
  red.** The two INDIRECT appends (`peak_claim`, `arb_claim`) carry no
  `actuator=` at the call site to mis-type, and are pinned where their routing
  is actually chosen - `peakguard.claim()` by the `SWEEP_SITE` sweep,
  `arbitrage.claim()` by `test_arbitrage.py`. 10g's dynamic cycle is kept and
  now says which three of the eight it sees instead of claiming the class.

**SAFETY-25's open question is ANSWERED: the BMS cut-out is the DELIBERATE and
only backstop for demand-driven discharge, and a loop-side reserve floor is
DEFERRED to `ALLOC-01` rather than rejected (SAFETY-25).** Decided by the
product owner on 2026-09-02. **Option A is the shipped behaviour, so `app/`
changed only where a comment claimed the question was still open** - no
setpoint, no claim, no option, no default moved, and the 17-suite run is green
on the same numbers as before.

* **Why deferring is the safe direction, and it is a property of the shape
  rather than a preference.** Option B's proof - `bms_soc_floor <
  loop_soc_floor <= peak_soc_floor`, verified through the real arbiter in
  `test_peakguard.py` part 10d - has an unstated premise: the reserve predicate
  and `discharge_permitted()` read the **same** state of charge. With two packs
  they read different numbers. The reserve binds on the emptier pack and files
  `hi = 0` while the peak guard, reading the other pack or an average, files
  `lo = excess > 0`; `lo > hi` is arbiter clause 3 and the answer is **0 W, the
  battery parked in the billed quarter** - the failure the memo itself marks
  fatal. **No config-ordering validator can catch it**, because it is a runtime
  divergence between two readings and not a config ordering, so CFG-01's shape
  of check would pass on the exact tree that parks the battery. Option A's cost
  is energy; Option B's cost built too early is a parked battery in the one
  quarter of the month that sets the bill.
* **The revisit trigger is recorded explicitly, and it is `ALLOC-01`.** Under
  `ARCH-01`'s host-side allocator the two claims stop sharing a list: the loop
  floor becomes a **per-inverter `hi` consumed as an interval at stage A**,
  while the peak floor stays a **site `lo` at stage B**. Different claim lists,
  different stages, **structurally unable to cross** - which is a stronger
  argument than the ordering rule, because it makes the crossing unreachable
  rather than harmless. That is the shape in which Option B becomes safe to
  build, and the memo says so at the point a future reader will look.
* **What will have to be RE-EARNED at N=2 is stated plainly rather than left
  for someone to discover.** `SAFETY-27` now names **five floors across two
  packs**, so "the SoC" stops being a scalar the config can bound. And
  `test_peakguard.py` **part 10f - the pin that made SAFETY-25 mergeable** -
  **cannot express two packs**: its dynamic half is a scalar 0-100 % SoC sweep
  against a single-series capture, and a two-pack exposure lives in the
  *difference* between two readings, which that sweep visits at no resolution.
  Its static half (the AST scan pinning the eight claim sources) still bites at
  N=2; the dynamic half goes silent. Merging a two-pack reserve on 10f's
  current green would be merging on a check that cannot see the failure.
* **The options analysis is kept intact on purpose.** A memo that recorded the
  verdict and deleted the reasoning is the failure mode this `DECISIONS/`
  directory exists to prevent, so §1-§7 stand exactly as written - including
  §5's *"(recommended)"* on Option B and §6's recommendation - with the new §0
  saying what was decided and §6 marked as answered. Part 10e's check moved
  with the memo: it used to pin the honest `Status: OPEN` line and now pins
  that the memo names the option, the date and the trigger **and still carries
  both options and the recommendation**.
* **The tripwire stays, and "deferred" is why it stays.** Part 10 and mutation
  13 are unchanged in substance: a ninth claim source in `Controller.cycle()`
  still takes 10f red on the line it is written. Deferred, not rejected, is
  precisely the state in which someone builds Option B by accident.

Operator-facing surfaces follow the decision and nothing else changed on them:
`DOCS.md`'s SAFETY-25 section now says a choice was made and points at the
trigger, and `app/freeze.py`'s finding and `app/peakguard.py`'s point 2 say
"decided, on purpose" where they said "open".

**`PeakStore.record_block()` now has its POLICY ARGUMENT graded, and the
SAFETY-20 import check now looks at all three ways of writing an import
(TEST-13).** Two survivors from TARIFF-01's review, both closed in
`test_tariff.py` only - `app/` is untouched and no shipped behaviour changed.

* **The plumbing of the per-phase reduction was unobservable.** Mutating
  `record_block()`'s `billed_w(block, policy)` to `billed_w(block, POLICY_NET)`
  left all 17 suites green (+4 B; recorded by TARIFF-01's reviewer, re-run
  through the purge harness for TEST-15, and re-run again here). The cause is
  the third instance of the same shape as WITNESS-02's span normalisation and
  TARIFF-01's own DST window: **a stimulus evaluated where the quantity it
  grades is degenerate.** `billed_w()` short-circuits to `block.offtake_avg_w`
  whenever `per_phase_offtake_avg_w` is absent, and every block that reached a
  store anywhere in the suite was single-phase or built with
  `per_phase_offtake_avg_w=None` - so NET, SUM and MAX collapsed onto one
  number and the argument could not be seen. Section 7 already owns an
  UNBALANCED fixture, `(3000, 500, -200) W`, whose three reductions are 3300 /
  3500 / 3000 W; that block is now round-tripped **through the store** under
  each policy and read back from both tables. NET is the smallest of the three
  here, so a store that quietly billed NET while asked for MAX would
  under-record the month's peak - maximum headroom for SAFETY-07, the same
  under-guarding direction as the `getattr(..., 0.0)` default SAFETY-07 already
  removed from `project()`.
* **The import half of the SAFETY-20 grader made a claim it did not check.** It
  collected `ImportFrom.module` only, which is blind to two of Python's three
  import spellings: `from . import arbiter as _ab` is an `ImportFrom` whose
  `.module` is **None** (the module name is in `.names`), and `import
  app.arbiter` is an `ast.Import` node the walk never visited. Both were added
  to `app/tariff.py` and the shipped suite stayed **green**. Never a live
  exposure - the `_names` half still catches any USE of either alias - but the
  printed sentence "it imports nothing from arbiter, clamp or control" was
  false for two forms out of three, which is worse than claiming less. The
  grader now walks `Import` as well, joins `from . import X` onto its base, and
  matches per dotted SEGMENT so `app.arbiter` cannot hide behind a package
  prefix while `clamps` / `arbiters` / `controller` are not falsely flagged.
  Both graders are also now shown non-vacuous against `main.py`, which really
  does import all three modules.
* **Two items on the card were already closed and are struck, not
  re-implemented.** `PeakReading.is_observed` is `self.state == OBSERVED`
  (`app/tariff.py:266`), already false for `NO_RECORD`; and `project()`'s
  `held` read is `max(averager.held_net_w, 0.0)` with **no** `getattr` default,
  fixed by SAFETY-07 in 616b62d with a comment naming TEST-13 item 4. Restoring
  a default there would re-open a money guard. Removing `clear_partial()` also
  stays green and that remains a **PASS**: the `quarter_start` PRIMARY KEY plus
  `ON CONFLICT ... WHERE excluded > existing` makes a replay a structural
  no-op, which is TARIFF-01's claim rather than a hole in it.

Verified through `tools/mutation_run.sh` (purge, `python3 -B`, tree re-read from
disk, mutated line grepped off disk, byte delta recorded): 7 mutations, all red
on the named checks, plus green baselines before and after and a control run
showing the shipped suite passing every one of them.

**The two `arbitrage()` rungs SAFETY-26's AC4 sweep found pinned from ONE SIDE
ONLY are now pinned from both - and, the actual finding, are pinned so they stop
MASKING EACH OTHER (TEST-16).** `spread == 0.0` weakened to `<= 0.001` survived
the whole suite at +2 B, and `len(firm) < 2` weakened to `< 1` survived at
+0 B - the delta MUTATION.md names as the risky one. Neither was individually
live, and that was the problem: a single firm hour *also* has a spread of zero,
so each rung's failure was absorbed by the other and both could be weakened
together with nothing reporting it. That is invisible to single-mutation
testing, which is how the pair survived four rounds of review on TARIFF-03 and
FCST-03. `test_static_tariff.py` gains section 9, which grades WHICH RUNG
answered each stimulus rather than only whether it refused - the masked case
refuses correctly while reporting the false sentence "the tariff is flat" about
a series that is not flat, so a check asserting only "refused" passes on a
broken tree. The flat rung is swept with small-but-REAL spreads down to
0.000001 EUR/kWh at `min_spread_eur_kwh=0.0` (a value `arbitrage()`'s docstring
documents and its validator accepts on purpose), the firm-pair rung is probed
with a ONE-FIRM-HOUR series and graded on §8.3's sentence coming back rather
than the flat rung's, and a five-row table denies the masking on both sides.
Verified through TEST-15's `tools/mutation_run.sh` unchanged - recursive
`__pycache__` purge, `python3 -B`, tree re-copied from disk between rounds,
mutated line grepped off disk, byte delta printed: 7 mutations plus a green
baseline, 4 of them at +0 B, ALL RED, including both rungs weakened AT ONCE
(5 checks red). **No shipped behaviour changed - both rungs were correct all
along**; `app/` is untouched. `MUTATION.md` gains a section on masking as a
second way a survivor lies, and `test_arbitrage.py`'s SAFETY-26 AC4 note is
marked closed rather than left reading as an open finding.

**`check_output_clamp()` could be moved above `write_setpoint()` in `cycle()`
with test_clamp.py, test_freeze.py and test_loop.py all staying GREEN — the
guard's write-path evidence, `self.clamp_last`, had nothing watching it.**
Measured before the fix, on the move itself: all three suites green. The guard
has two evidence sources and only one of them was covered. `self.loop_clamped`
is set from the `Decision` well above the write, so it survives the move — and
on the §10.3 runaway replay it is the source of every engagement the alarm
reports (0 write-path engagements against 2 loop-side ones), which is precisely
what made the move nearly invisible. `self.clamp_last` is assigned by
`write_setpoint()` and by nothing else (`main.py`), is `None` until it has run,
and is what `engaged_write` reads; moved above the write, that consumer goes
quietly false and a DEVICE-bound clamp — `max_w` above the inverter's own
rating, SAFETY-20's install — stops being reported by anything.
test_clamp.py gains part 6, and it grades the CONSEQUENCE rather than the
position: a spy subclassed from the shipped `ClampWatchdog` records what
`self.clamp_last` held each time the guard consulted it, over a slew-limited
ramp where every cycle's command is a different number, and asserts the
evidence is always the write of THAT cycle. Hooking `clamp_wd.note()` rather
than the guard's name is what keeps a rename legitimate; the same assertions
are re-run through a controller whose guard is renamed and whose write lives in
a helper, and were verified against `main.py` itself — renaming the method
throughout and moving both calls into a new `_emit_and_guard()` helper keeps
part 6 green, while swapping the two calls turns it red on a line that names
`clamp_last`. AC 3 needed a stimulus of its own: through the real `cycle()` the
write-path half can never engage, because the arbiter files the device range
and `max_w` as P_SAFETY LIMIT claims and `resolution.target_w` is already
inside both bounds before `write_setpoint()` re-applies them — that duplication
is deliberate defence in depth, and part 6 now measures the 0 engagements
rather than leaving it as a claim in a comment. So the write path is driven
directly at `max_w` 5000 against a ±2500 W device: the wire carries 2500 W,
every loop-side flag stays clean, and the alarm fires on `clamp_last` alone —
with the two calls swapped, the same stimulus clamps the same value and the
alarm is silent. `main.py` gains one comment sentence pointing at the check,
mirroring the one SAFETY-18's block already carries for test_freeze.py part 5.

**`max_w` and the `LIVE_MAX_W` literal beside it were unpinned in test_freeze.py
and test_tariff.py, the same gap TEST-06 closed one key over for
`meter_freeze_s` (TEST-10).** Measured before the fix: `max_w: 2000 -> 4500`
in `config.yaml` left every check in both files GREEN, and `LIVE_MAX_W = 5000.0`
was a bare literal that silently disagreed with the shipped default (2000) with
nothing saying so — for contrast, `deadband_w: 15 -> 5` already went RED 2 + 1,
so the pinning technique itself was known to work, just not applied here.
Fixed three ways. First, `test_freeze.py` gained a check comparing `Tuning()`'s
code default (`control.py`, 2000.0) against `SHIPPED_MAX_W`, read from
`config.yaml` by extending the file's existing `_cfg_default()` regex reader
rather than reimplementing test_arbitrage.py's `_shipped`. Second, the literal
itself is renamed `LIVE_INSTALL_MAX_W` in both files, with a comment stating
plainly that 5000 is the live install's own override and not config.yaml's
shipped `max_w` — the defect was the silent literal, not the number. Third,
`test_freeze.py` now states `PINNED_KEYS = ("meter_freeze_s", "deadband_w",
"max_w")` in one place as the complete set of config keys this suite pins,
deliberately excluding `maintenance_*` (CFG-03 gave those a separate source of
truth in `app/maintenance.py`). Verified by breaking what the new check
watches: `max_w: 2000 -> 4500` in `config.yaml` now fails exactly
`"the code default and the shipped default agree on max_w too"` and nothing
else, and reverting returns the file to all-green.

**Maintenance option defaults lived in three places that had already disagreed,
and MAINT-04's tripwire was holding them apart rather than closing them
(CFG-03).** `config.yaml` shipped `maintenance_charge_w: 2000`, `main.py` carried
two `opts.get("maintenance_charge_w", 2500)` fallbacks, and `MaintConfig`'s
dataclass carried a third `2500.0`. An options dict with the key **absent** was
therefore accepted by `validate.py`'s rule 6 — which read 2000 — and then built
by `main.py` at 2500, against a 2000 W ceiling. Demonstrated through the shipped
code:

    options dict with maintenance_charge_w ABSENT, max_w=2000
      validate_startup_config  ->  [] (ACCEPTED, no problems)
      MaintConfig().charge_w   ->  2500.0
      arbiter.resolve()        ->  0.0 W, contradiction=True

That is the bug rule 6 exists to refuse, **passing rule 6**. Unreachable through
the Supervisor, which always supplies every key from `config.yaml`; fully
reachable from `sim/inject.py` and from the test suite, which is where anyone
debugging maintenance is.

**`maintenance_discharge_w` had the identical divergence** (`config.yaml` 2000,
`MaintConfig` 2500) and the card named only the charge side. Closing one key and
leaving its twin would have reproduced the defect one line down, so both are
closed.

**The single source of truth is `config.yaml`'s `options:` block** — the file the
Supervisor reads and the operator edits. There is now exactly **one** in-code
copy, `maintenance.MAINT_OPTION_DEFAULTS`, which is TARIFF-02's
`PRICE_OPTION_DEFAULTS` and SAFETY-24's `PEAK_OPTION_DEFAULTS` pattern applied to
the family that named the card. A mirror rather than a runtime read because the
`Dockerfile` copies `app/` and `run.sh` into the image and **not** `config.yaml`:
the running add-on cannot open the file it is configured by. `test_maintenance.py`
pins every entry by **reading** the shipped file, so a drift is a CI failure
rather than a comment. Every other declaration site now derives from it or has
ceased to exist:

- `MaintConfig`'s eight dataclass defaults are indexed out of the mirror — no
  numeric literal survives in the class that held the third opinion;
- `MaintConfig.from_options()` is the **only** place in the add-on that reads a
  `maintenance_*` option. `main.py`'s eight inline `opts.get(key, <literal>)`
  calls are gone;
- `main.py`'s second fallback site (the cheap-window charge floor) reads
  `self.maint.cfg.charge_w` instead of re-opening the options dict, so the rate
  the shaping asks for and the rate the state machine drains at cannot be two
  numbers from one install;
- `validate.py`'s rule 6 and its `maintenance_discharge_w` warning take their
  fallbacks from the mirror.

**Where the RANGE lives is now stated and checked, so the next person does not
add a fifth site.** `maintenance_charge_w: int(500,5000)` in `config.yaml`'s
`schema:` block is a fourth declaration of a different kind — a bound, not a
default. Bounds live in the schema block **only**: the Supervisor enforces them
and reads only that file. `test_maintenance.py` asserts that every mirrored key
has a schema entry (or the Supervisor silently drops the option and the add-on
runs on the mirror instead of the operator's value) and that every shipped
default lies **inside** its own range. What the schema grammar cannot express is
a relation *between* keys — `maintenance_charge_w` against `max_w` — which is
why rule 6 is written in Python and does not belong in the mirror either.

**MAINT-04's tripwire in `test_validate.py` is DELETED, not left passing**, as it
and `validate.py`'s ⚠️ block both instructed each other to do. Its passing
condition *was* the bug. In its place, the same four consumers exercised the same
way and now agreeing: the validator accepts a dict with every `maintenance_*` key
missing, a **real `Controller`** built from that dict holds 2000 W, `MaintConfig()`
answers 2000 W with no dict at all, and the real arbiter emits `-2000 W, limited
by supervised(max_w)` instead of `0 W, contradiction=True`.

**Checked by breaking it, five ways, each observed red before being reverted:**
move `config.yaml` alone → the mirror pin fails; move `config.yaml` **and** the
mirror to 1500 → every consumer follows together (`MaintConfig()` 1500 W, the
`Controller` 1500 W, the arbiter `-1500 W, limited by maintenance(cheap-window
charge)`) and both suites go green again; retype `charge_w: float = 2500.0` into
the dataclass → red; restore `opts.get("maintenance_charge_w", 2500)` in
`main.py` → red; delete a key from `config.yaml`'s options block → red.

Stale prose that asserted the closed divergence as current is corrected in
`app/watchdog.py`, `test_watchdog.py` and `test_peakguard.py`.

**`peak_soc_floor` reads like a pack-wide reserve and is not one, and on a
sustained-load evening the grid-following loop discharges straight past it to
the BMS cut-out (SAFETY-25).** The gap has been written down four times without
being filed — SAFETY-18's finding (`freeze.py`), SAFETY-24's module note, its PR
body, and its honestly-uncaught 13th mutation — and never once measured. It is
measured now, on the committed 2026-08-20 capture, driving the shipped
`control.compute()`, `peakguard.decide()`, `peakguard.claim()` and the real
`arbiter.resolve()`:

    19:00-22:00 UTC, 830 W mean, minimum 701 W, not one watt of export
    reference pack (300 Ah x 51.2 V x DoD 90) started at 21 %, one point
    above peak_soc_floor

      ends at 10.0 %  -  the bms_soc_floor cut-out
      1525 Wh delivered by the LOOP over 587 cycles
      549 cycles below peak_soc_floor, 343 of them still discharging
      0 cycles on which the peak guard filed any floor at all

**And it is a behaviour, not a config value.** The same evening ends at
`9.9715 %` at `peak_soc_floor` = 5, 11, 20 **and** 60 — the shipped value and
both ends of the schema range — identical to four decimal places. The option an
operator would reach for cannot move the number by any amount. On the same day's
high-load block the two states coexist: 49 cycles reporting `PEAK GUARD
FLOOR-LIMITED` and, on **all 49**, the arbiter simultaneously granting the loop
up to 2000 W of discharge.

**Why SAFETY-24 could not close it.** The peak guard files a `lo` limit — a
*lower* bound. That asymmetry is what makes it provably unable to lower a target,
cap a discharge or contradict the arbiter, and it is therefore equally unable to
stop a discharge the loop itself wants. A floor on the guard cannot become a
floor on the pack by being raised.

**What changed is the honesty, on four operator-facing surfaces**, because that
half is required whichever way the open question is answered and it is the actual
misstatement on the live system today:

- `config.yaml` says next to the option that it **bounds the peak guard only** and
  is **not a pack-wide reserve**, with the replayed evening and the sweep;
- `DOCS.md` says it in the options table and in a new section carrying the
  figures above;
- the **status-page row** for a floor-limited guard carries the scope sentence —
  it is the one line a human reads at the moment the guard stops, and read alone
  it implies a reserve the replay shows does not exist;
- the **persistent notification** says what has stopped (the guard buying a lower
  bill) and what has not (the pack being spent).

`peakguard.py` and `freeze.py` carry the measurement beside the findings they
already recorded.

**The product decision is NOT taken here.** Whether the loop gets a reserve floor
of its own moves real energy on this house and belongs to the owner.
`DECISIONS/SAFETY-25-loop-soc-floor.md` records the options, what each costs and
protects, weighed against MAINT-01 and against the fact that this would be the
**fifth** SoC key (OPT-04's `arbitrage_soc_floor` fired SAFETY-24's "revisit if a
third floor appears" trigger unremarked), and a recommendation. It also records a
trap found while writing it and now checked through the real arbiter: the obvious
shape — a `hi` limit on the loop — crosses the peak guard's `lo` and resolves to a
**CONTRADICTION at 0 W**, the battery parked in the quarter that sets the bill,
unless `bms_soc_floor < loop_soc_floor <= peak_soc_floor` is enforced at startup.

**`test_peakguard.py` part 10 is a tripwire, not a guarantee, and part 10f is
what makes it one.** The replay half records the defect; on its own it could not
see the defect being FIXED, because `replay_loop()` hand-builds main.py's claim
list and is therefore blind to any claim main.py adds — which is exactly where a
loop-side floor would be built. PR #31's reviewer proved that by building the
memo's own Option B in `main.py` and watching the whole suite stay green. **10f
pins the claim-source set:** every `claims.append(...)` in `Controller.cycle()`
is read off its AST and must be one of eight named sources, and real
`Controller.cycle()`s — spied at main.py's own `resolve`, swept over the whole
0–100 % state-of-charge range in both the idle and the floor-limited state — must
hand the arbiter exactly `{device, supervised, loop}` and
`{device, supervised, loop, peak}`. A ninth source, at any state of charge, in
any configuration, takes part 10 red on the line it is written. A floor built
inside `control.compute()` or the guard instead moves 10a–10c's replayed numbers,
which the replay drives directly. So the exposure cannot change size under a
green suite.

**Nine mutations, re-run against the fixed tree**, each against a freshly copied
tree with no `__pycache__` present at all, mounted `:ro` — the same structural
immunity `MUTATION.md` records for CI. Six are the AC 3 and behaviour mutations
already on this ticket's record (the status row's scope sentence, the
notification's, `config.yaml`'s, the memo deleted, the loop's gain zeroed, the
guard's claim given a `hi`); two are the reviewer's, and are the ones that
motivated 10f — the memo's own Option B at its recommended default, **now three
checks red where it was fully green**, and an absurd always-binding 95 % reserve,
**six red**; one is new, against the banner short form below. All nine are red,
no survivors, and no byte delta is `+0 B`.

**The gate has never once graded the content a pull request would land, and a
run that never started reported as a test failure (CI-02).** Both halves were
measured on a deliberately conflicted throwaway pull request rather than read
off the workflow, because the Actions log endpoints 404 on this Gitea and a
step's own conclusion is the only channel out of a job.

**Case 1, and the ticket's mechanism was wrong in a way that made it worse.**
CI-02 was filed saying a conflicted pull request loses its merge ref, so that
`git fetch $GITHUB_REF || git fetch $GITHUB_SHA` falls through and grades the
bare branch head. The symptom was real — PR #14 reported `checks / unit ->
success` while `mergeable: false` — but the mechanism is not. On this server
`GITHUB_REF` **is** `refs/pull/<n>/head`, `GITHUB_SHA` is the same commit, and
`git ls-remote origin 'refs/pull/*'` advertises no `/merge` refs at all, for
open and closed pull requests alike. The first fetch has always succeeded; **the
fallback has never fired once.** So the gate did not lose merged content on
conflicted pull requests — it has never had merged content, on any pull request
in this repository's history. Green has meant "this branch passed on its own",
which on a base that has moved is not what lands, and `block_on_outdated_branch`
is off on release/1.0 so the base usually has moved.

**The checkout now fetches the base tip and the pull request head, merges them,
and refuses if the merge does not apply** — as its own named step, so
`/actions/runs/<id>/jobs` distinguishes "nothing was tested" from "the tests
spoke". Verified by producing it: with the fix on a conflicted pull request the
required context went `failure` in 2 s with the two test steps `skipped`. The
`--depth 1` is gone (a merge needs a merge base; a full fetch of this repository
is 6.8 MB and 1.1 s), and the fallback survives only on the `workflow_dispatch`
path, which has no pull request to merge and now says so in its output.

**Case 2: `concurrency:` is gone, and `cancel-in-progress` was never the knob.**
Runs 531 and 533 on PR #14 were cancelled at 0 s each at the exact second the
next run was created, having never started; a cancelled run posts `failure` on
the required context, described `Canceled after 0s`, and branch protection
compares state and not description — so "nothing ran" and "your tests are
broken" arrive at the gate as the same red, wanting opposite responses. Three
settings were measured with one run holding the single runner slot and two
pushes behind it: with **no block at all** both queued runs survived and
executed; with the group kept and **`cancel-in-progress: false`** the queued run
was still cancelled at 0 s. A shared group cancels a run that has not started
regardless of that flag. Having no group is the only configuration measured to
leave a queued run alive, and it is what ships. The cost — a developer who
pushes four times occupies the shared runner four times — is CI-01's own trade
written the other way round: a delayed verdict, never a destroyed one.

**`test_ci_gate.py` grades the gate by running it, not by reading it.** It
extracts the checkout and merge steps out of `checks.yml` and executes the
shipped commands against two throwaway local repositories it builds — one whose
pull request merges, one whose does not — and pins that the old one-liner exits
0 on the unmergeable one while the new steps refuse. It also pins the two tokens
that assemble `checks / unit (pull_request)`: renaming `name:` or the `unit:`
job key would leave every open and future pull request waiting forever on a
context that never arrives, with no error anywhere in that sequence. Seven
mutations of `checks.yml` were each confirmed red. `.gitea/ci/Dockerfile` gains
`git` for it.

**What was NOT established, stated so nobody reads more into this.** The
original report that "no run was ever created for `0942228d`" was already
corrected on the card — run 533 was created and cancelled — and this work
confirms that from the runs list and the per-commit status history. A commit
that never had a run carries an empty `/statuses/<sha>` list; a queued one
carries `pending / Waiting to run` within seconds, and a concurrency-held one
`pending / Blocked by required conditions`. Those are distinguishable at the
API. `/actions/tasks`, which is what was polled originally, lists only tasks
that started, and cannot answer the question at all.

**The NOT READY banner no longer repeats the row underneath it.**
`status_payload()` renders the first hard-failing check as the page headline, so
the scope sentence above turned the loudest element on the page into a
550-character paragraph that the checks list then repeated verbatim. A row may
now carry an optional short `banner` form; the floor-limited row does, and its
headline is the guard's own line plus a 68-character scope tag. **The row itself
is unchanged** — the checks list is where an operator reads what the floor does
and does not bound, and shortening that would have been weakening AC 3. Rows
without the key are unaffected, and the page still renders with `textContent`.

**A setpoint device smaller than `max_w` silently disarmed the frozen-meter
detector, and the status page reported it as armed (SAFETY-20).** SAFETY-18's
conjunction needs `abs(commanded_w) >= 0.95 * max_w`, but the emitted command is
clipped to the setpoint device's own `[dev_min, dev_max]` range as well as to
`max_w` — in `clamp_to_rating()` and again as the arbiter's `device` P_SAFETY
limit claim. Measured on the shipped predicate at the ticket's numbers,
`max_w: 5000` on a ±4700 W inverter:

    looks_frozen(unchanged_s=3600, commanded_w=4700, max_w=5000)  ->  False

The command is pinned at the hardware's rail, the meter has been frozen for an
hour, and term 2 is unsatisfiable by construction because 4700 < 4750. Any
rating at or below `0.95 * max_w` disarms the conjunction outright. The page
then printed `armed (… command is over 4750 W)` — a guard that cannot fire,
advertising that it can, which is the failure SAFETY-15, SAFETY-19 and
validate.py rule 5 all exist to prevent.

**The rail is now measured against `min(max_w, |dev_min|, |dev_max|)`** —
`freeze.rail_ceiling_w()`, threaded through `looks_frozen()`,
`FreezeWatchdog.evaluate()` and both call sites in `main.py`. An unread rating
(`hass.limits()` returning `None` when ESPHome is not up) leaves the old
behaviour exactly, and correctly: on that path no device clamp is applied
either, so `max_w` really is the ceiling. A rating *larger* than `max_w` cannot
raise the rail — this is a `min()`, not a swap. The status row prints the rail
the predicate actually uses and names the device when the device is what capped
it; the log line and the HA notification do the same, since they outlive the
page.

**The smaller of the two device bounds, not a per-sign pair.** A per-sign rail
is tighter and is not wrong, but it is two rails needing two coverage verdicts
on a status row that carries one number, and the only configuration that
distinguishes them is refused at startup anyway. The symmetric ceiling can only
*lower* the rail, which widens term 2 rather than disarming it; term 1 carries
the no-false-positive argument and is untouched.

**A ceiling of 0 W is NOT COVERED, not "armed at 0 W".** A device that can
command nothing in one direction would make `abs(command) >= 0` true of every
command including a parked one — term 2 degenerately *true*, collapsing the
conjunction into the bare `unchanged_s > N` threshold TEL-05 refused for
false-tripping at this controller's own operating point. `rail_unreachable()`
returns the reason, `looks_frozen()` stays silent, and the page uses the NOT
COVERED row the unsupported-transport case already has — a second reason to use
that surface, not a new mechanism.

**AC4, decided rather than assumed: no new `validate.py` rule.** The
configuration is `max_w` above the device rating, which rule 2 (CFG-01) already
refuses on the identical predicate, re-run by `run_control()` once the real
rating is known. A rail-specific rule would fire on a strict subset of what rule
2 refuses (`0.95 * max_w > dev_max` implies `max_w > dev_max`) and could never
be the rule that stopped a start. The residue is the state where `hass.limits()`
answered `(None, None)` and the range checks were skipped — there is nothing to
refuse there, by construction, which is why the answer is a runtime row. Written
into validate.py's "deliberately not refused" list, and *pinned* by four checks
in `test_validate.py` so the decision's premise goes red if rule 2 is ever
softened back to a warning.

**And it was verified by breaking it.** Reverting `rail_w()` to bare `max_w`
fails nine checks; deleting the NOT COVERED handling fails five; and — the one
that mattered — shipping the corrected predicate while *not* passing the rating
from `check_meter_frozen()` originally left the whole suite green, because
nothing graded the caller. `test_freeze.py` part 5b now drives the real
`cycle()` on a ±4700 W device and asserts on the notification that leaves the
process; the unwiring mutation fails three checks there. The peak-guard call
site's wiring is still ungraded — it sits behind `p1_enabled` and needs a fed
averager to have any observable consequence — and that gap is named in the file
rather than implied.
**A mutation run could grade code that was not there, and the fix is a purge
rather than a flag (TEST-15).** FCST-03's developer reverted `self.provisional`
to `here.provisional`, re-ran, and reported a survivor; disassembly showed the
source read `here` while the running code loaded `self`. CPython validates a
`.pyc` on (source size, source mtime truncated to whole seconds), and a
size-preserving edit reverted inside one second moves neither number, so the
stale bytecode was reused. That does not make past results wrong - it made them
unverified, and the distinction is the whole card.

Reproduced rather than reasoned about. `test_bytecode.py` builds a real package
whose two return values are the same length, imports it in a real subprocess,
rewrites it inside the same second and reads back which value ran. Part 1 shows
the interpreter running `SELF` off a file that says `HERE`.

**And it shows the two obvious mitigations do not work.** `python3 -B` and
`PYTHONDONTWRITEBYTECODE=1` are write-side only: neither stops an EXISTING stale
`.pyc` being read, and a mutation run always has one because the previous run
wrote it. The advice already in `test_control.py`'s coverage audit led with `-B`
and is corrected here - the purge is the half that works, and `-B` earns its
place by protecting the NEXT edit rather than this one. No `sleep` anywhere:
one-second mtime resolution is the mechanism, and waiting it out would slow every
run down while still only covering the edit somebody remembered to slow.

`goodwe_controller/tools/mutation_run.sh` is the harness the house adopts: it
purges, prints a sha256/byte fingerprint of every `app/*.py` it is about to
grade - so a mutation report can say WHICH source produced it - runs each suite
under `python3 -B`, refuses a tree with no `test_*.py` (CI-01's rule), and fails
if a cache appears during the run. It takes the tree as an argument so the infra
repo's replay harnesses can call it.

`MUTATION.md` states the guarantee where a mutation gets written, and every one
of the 16 suites now points at it from its docstring - a list `test_bytecode.py`
derives from the tree, so a new suite joining without the pointer is red.

CI was already immune and is described that way rather than as a hole being
closed: `checks.yml` mounts the checkout `-v "$WS:/w:ro"`, and a read-only tree
cannot take a `__pycache__` at all - measured, and it fails silently, exit
status 0 - while each job `rm -rf`s its workspace and clones afresh.
`PYTHONDONTWRITEBYTECODE=1` is added as belt and braces, because the `:ro` is one
character wide. The shipped add-on was never affected: `Dockerfile` has set
`PYTHONDONTWRITEBYTECODE=1` since it was written.

**The survey of this week's mutation claims is in `MUTATION.md`, scoped by the
property rather than by re-running everything.** The defect is one-directional -
stale bytecode turns a red into a green survivor and never the reverse - so every
"all red, no survivors" result in the window is not at risk, and the exposure is
the recorded SURVIVORS from source-edit runs made before the purge discipline
landed on 2026-08-31. Two of them were re-run here under the harness and both
still stand (TARIFF-01's `record_block()` ignoring its `policy` argument;
SAFETY-03's `abs()` on the clamp limit). The rest are labelled still-unverified
rather than quietly assumed or noisily re-run.

Out of scope and recorded: the infra repo's `sim/` replay harnesses import the
add-on and write bytecode into both trees. Anyone mutating a module graded by a
replay must purge both by hand until that repo adopts the harness.

**`release/1.0` went red at midnight on a commit nobody had touched (TEST-17).**
`Controller.checks()` read the capacity-tariff peak with `self.peak.peak()` - no
argument - so `tariff.py` resolved the month from `datetime.now()`, while
`test_tariff.py` section 10 banks its quarter at the fixed fixture
`2026-08-30`. From 1 September the page was correctly asked for September,
correctly answered `NO_RECORD`, and the row read *"no completed quarter recorded
yet"* - a string containing no `"peak"`, which the section's filter required.
Four checks failed. The suite had been a latent time bomb since TARIFF-01
merged on 2026-08-30 and every PR that ran it ran in August.

Isolated by holding the commit and the image constant and moving only the clock
under `libfaketime`: green at 2026-08-30, four failures at 2026-09-01, identical
at `f4914a1`, `04292c4~1` and `04292c4`. The calendar, not a regression.

**The product was right and the test was wrong**, so the fix does not change what
the live page says. `checks()` takes `when: datetime | None = None` and threads
it into `peak()` and `warning_frac()` - the convention every other peak query in
`tariff.py` already follows. The single production caller passes nothing and
still means today. Section 10 now grades at its own fixture's month.

Three new checks make the defect detectable rather than seasonal: the same
controller read at two different reference months must produce two different
rows, which no clock can satisfy; the later month must read `NO_RECORD` and stay
distinguishable from both an observed peak and an unreadable store; and calling
`checks()` with no argument must still mean today. Verified by mutation - a
`checks()` that drops `when` is now red at every clock tried, where the old
section went green in August. Full suite green, 15 files, at the real clock and
at faked clocks in four other months.

One thing left alone deliberately: only the UNREADABLE branch prefixes its row
with `capacity-tariff peak:`. The `NO_RECORD` row carries neither that phrase
nor the word `peak`, which is what made the original filter brittle and what
broke the first draft of the new checks too. It is product text with other
suites reading it, so it is noted here rather than changed.

**Thirty days of the real meter, and the acceptance criterion three cards
deferred is now a number (DATA-01).** TARIFF-01 and SAFETY-07 both merged with
AC 3 - *"monthly peak held under the soft limit in >= 95 % of intervals over a
30-day simulation replay"* - honestly unmet, on the belief that only two days of
capture existed and nothing longer could be produced. That belief came from a
true statement about the past being read as one about the future: HA's recorder
still held the whole month. `scenarios/p1-net-30d-2026-08/` is 2026-08-01 to
2026-08-30, **485,841 samples, 30/30 days read, 0 failed, 4.13 MB gzipped**, in
the same schema as the two days that were already there.

`tools/export_p1_history.py` is the exporter, committed and runnable - the
2026-08-25 one-off was not, and that is exactly why the data was thought
unobtainable. It reads day-at-a-time, retries, and **never lets a failed read
look like an empty hour**: a day whose read fails writes no file, is recorded as
`"read": "FAILED"`, and the run exits non-zero; the entity is proved to exist
before any empty day is believed. Nothing is interpolated, so the month's 13
gaps (1291 s, 99.95 % covered) and its 3 `unavailable` samples stay visible.
`--verify` re-fetches the two committed days by the same route and reproduces
them **byte for byte** after decompression.

`tools/replay_ac3.py` runs the month. Over 30 days the criterion measures
**98.89 % uncontrolled and 98.99 % guarded**, so it is met on that reading - and
the uncontrolled figure is the finding: a month with no controller at all passes
it. On the other reading of the same sentence it is not met - the guard takes
August's peak from **5029.8 W to 4107.3 W**, still over the 3500 W soft limit,
because once a month peak is set `CEILING_MONTH_PEAK` defends that peak rather
than the cap. Neither number is hidden; both are in that directory's README.

The size decision is committed and measured, not absorbed: a `git bundle` of
the whole history goes from **1.58 MB to 5.28 MB** (+3.70 MB, x3.3), which a
Home Assistant supervisor clones on every install. Downsampling was not
available - `meter_max_age_s` is 30 s, so any coarser resample converts real
coverage into modelled meter-gap and moves the number being measured.

The thirty files sit in a SUBDIRECTORY on purpose: `test_tariff.py` part 9 and
`test_peakguard.py` part 6 glob `scenarios/p1-net-*.jsonl.gz` non-recursively,
so CI keeps replaying the two small days and does not grow ~50 s per PR. Both
files' honesty notes are corrected to point at the month.

**OPT-04's AC4 is met: twelve months of real EPEX BE day-ahead prices are now in
the repository, and the strategy has been replayed against them (DATA-02).**
AC4 asked for "net-positive across 12 months of replayed prices". OPT-04 shipped
it as NOT MET and replayed a synthetic year instead, because `scenarios/` held
only P1 meter captures. That was the right call at the time and the wrong
diagnosis of why: the criterion was being carried alongside a 30-day *meter*
criterion as though both needed waiting, and EPEX BE day-ahead is **public data,
not a capture**. It needed a fetch, not a month.

`scenarios/epex-be-dayahead-2025-09-01_2026-09-01.json` — **8760 hourly points**,
2025-09-01 to 2026-09-01 Brussels-local, in §9.3's canonical three-field shape.
`scenarios/fetch_epex_be.py` produced it and `--verify` re-derives every claim it
makes about itself without touching the network.

**Licensing was checked before anything was committed.** The source is
Fraunhofer ISE's Energy-Charts API, which publishes the **BE** zone under
**CC BY 4.0 from Bundesnetzagentur | SMARD.de** and other zones under
private-use-only terms — so this zone is redistributable with attribution and the
data is committed rather than left as a script plus a checksum. The fetcher
**refuses to write a file** whose per-response `license_info` stops saying
CC BY 4.0. Cross-checked against the upstream: SMARD's own `4996/DE` series
matched this feed on all 96 quarter-hours of 2026-04-10, maximum difference
0.00 EUR/MWh.

**The figures are raw wholesale and the file says so.** `"all_in": false`, with
the markup/tax assumption carried in `provenance.all_in_assumption` as *numbers
to apply* rather than numbers already applied — TARIFF-02's double-counting trap
in data form. `test_arbitrage.py` section 7B applies it in one visible place and
**asserts arithmetically** that its two additive terms cancel out of every daily
spread, so the trade counts depend on the VAT multiplier alone and not on the
guessed markup. The same replay is run on raw wholesale as well, so how much of
the answer rests on the guess is reported (110 trades all-in, 88 raw).

**DST is handled by construction, not by a special case.** The fetcher iterates
in UTC and renders through `price.brussels()`, so **2025-10-26 comes out 25
points and 2026-03-29 comes out 23** — verified by counting per Brussels day
rather than asserting 24, and by checking that October's 02:00 appears twice at
`+02:00` and `+01:00` with two different prices while March's 02:00 does not
appear at all. The window is bounded at *local* midnight: bounding it at 00:00
UTC produced a 22-hour first day, and a 24-hour day silently missing from the
end, both of which the point-and-gap accounting caught.

**The result, at the shipped default, whatever it is.** At
`cycle_cost_threshold_eur` **0.10** widened by FCST-03's no-weather **1.5×** to a
required 0.1500 EUR/kWh, the real year completes **110 round trips** and is
**net-positive at +24.88 EUR** of benefit net of the cycle cost (2 kWh a trade).
Unwidened it completes 173. **Nothing was tuned:** the shipped default and the
2.5/1.5 factors are untouched, and every figure comes out of the same `decide()`.

**What the product owner's open decision now has under it.** The four
independent reports that a static day/night spread of ~0.0840 does not clear
0.10 **all stand** — a static-tariff site still never trades. What was missing
was the other half: a *dynamic* site's daily spread clears the widened
requirement on **121 of 365 days**, median daily all-in spread **0.1130**. The
threshold at which a median day would clear is **0.0754**, and at that value the
same year completes 156 trips for +35.42 EUR. Both numbers are reported; neither
is applied.

**And the synthetic year stays, because the real one cannot grade §17.** Not one
of the 365 real days has a negative *dearest* hour — 372 negative wholesale hours
and 0 days where the strategy would have sold into one — so "0 discharges into a
negative-price hour" over real prices is a count of something that could not
happen, and deleting the guard would not change it. Section 7B asserts that
degeneracy instead of quietly reporting the zero; section 7's generator is still
the only thing here that puts a negative price where the strategy would sell.

**Size, measured rather than assumed.** Plain JSON: 893 KB in the working tree
but **74.6 KB as a git object**, against 56.5 KB for a committed `.json.gz`. Git
already compresses blobs, so gzipping saves 18 KB and costs a reviewer the
ability to read the provenance block — including the licence claim they are
meant to check — in the diff. Plain, and `load()` still accepts `.gz` for a
future dataset that earns it.

**Nothing is wired into the add-on.** `generic_dayahead` remains the live fetch
path; nothing under `app/` imports the script, and the dataset is for replay.

**The capacity-peak guard's forced discharge now has an EMS-side energy floor
(SAFETY-24).** SAFETY-07 can command a discharge to hold a billed quarter down,
and nothing below it consulted the state of charge — `soc_floor` belongs to the
maintenance cycle, and freeze.py already recorded that for the grid-following
loop "the BMS is the sole energy backstop". On the 2026-08-20 replay the guard
asks for **878 Wh in a day**. Relying on the BMS is not absurd — it is a real
floor and it works — but it is the pack's protection rather than the operator's,
and a month of aggressive quarters leaves the battery repeatedly at that cut-out
with no EMS-side record of why.

`decide()` now takes `soc` and `soc_floor`, both **required** and neither
defaulted. Above the floor nothing changes. At or below it the guard files a
**0 W** floor instead of a sized one — one-directional, so charging stays blocked
and no discharge is capped — and reports `floor_limited` with the reason it
bound. It is not a trip: every stop in §11.2's table still outranks the claim,
and the floor can never be negative, so §9.1's "discharge and export are never
capped by the peak guard" survives as a property of the object (swept: 0 negative
floors across 1296 combinations of load, elapsed, battery, SoC and floor).

**Failing to shave is not silent.** A floor-limited cycle raises one notification
per episode (`goodwe_peak_soc_floor`), distinct from the saturation alarm because
the two ask the operator for opposite things, and renders its own status row.
An **unreadable SoC is floor-limited, not a full battery** — a full battery is
exactly the value that would disarm the floor, and this project has shipped that
defaulting once before (`| float(100)` in the stage-5 loop).

**Verified by replay, not by argument.** The 2026-08-20 capture is driven against
a modelled pack sized from the day's own measured appetite, so it genuinely
exhausts: without the floor it reaches the BMS cut-out and sits there for 1227
guard cycles; with it the pack stops at 20.0 % and never touches the BMS. The
energy withheld equals the energy between the two floors to within one cycle at
the rail. The honest half is reported too — the worst quarter goes from 4338 W to
5030 W, because a floor-limited guard shaves less on purpose.

**Two new options, mirrored once.** `peak_soc_floor` (20) and `bms_soc_floor`
(10), whose defaults live in `peakguard.PEAK_OPTION_DEFAULTS` — TARIFF-02's
pattern — imported by `validate.py` and `main.py` rather than typed as literals,
and pinned to `config.yaml` by a check that reads the file. CFG-03 still tracks
three places. `validate.py` rule 8 refuses `peak_soc_floor <= bms_soc_floor` at
startup, in CFG-01/MAINT-04's manner. `maintenance_soc_floor` is deliberately
**not** reused and deliberately **not** renamed: it is a drain target, not a
spend limit, and the rename is recorded as a decision to revisit if a third
floor ever appears.

**Price arbitrage - the first strategy that trades on a price (OPT-04).**
Four cards built the price stack and nothing traded on it. `app/arbitrage.py`
is §10.1's priority 3: charge in the cheapest published hour, discharge in the
dearest, **and only when the spread covers `cycle_cost_threshold_eur`**. It is
off by default (`arbitrage_enabled: false`) - it is the first strategy in this
add-on that commands the battery against the meter on purpose.

**It consumes the four cards rather than re-deciding anything.** Whether a
trade is permitted at all, on which two hours, and against which widened
threshold is `TariffPlan.arbitrage()`'s answer (TARIFF-03), and that answer
cannot be obtained without first stating whether the peak guard is holding. So
§9.1's "give peak shaving priority over arbitrage" is honoured
**structurally**: there is no code path to a tradeable spread that skips the
question. And because the claim is filed at `P_ARBITRAGE` (30) while peakguard
files its floor at `P_PEAK` (50), the arbiter clamps an arbitrage charge even
if that interlock were bypassed. Two independent mechanisms, neither
re-implemented here.

**A guard that never ran is not a guard that is clear.** On `meter_source: off`
there is no telegram stream, so the peak guard is never evaluated. That third
state is carried as `None` and it **refuses**: charging at the rail creates the
billed quantity while nothing measures the quarter it lands in. `validate.py`
refuses that configuration at startup too, along with arbitrage armed on a site
with no price source at all, and a SoC floor at or above the ceiling.

**§17: never export into a negative price.** The guard is on the hour being
sold into, not on the spread - a day priced -0.30 ... -0.05 has a 0.25 EUR/kWh
spread that clears any shipped threshold and a dearest hour that is still
negative. Buying in a negative hour stays allowed; being paid to charge is the
right trade.

**Prices are fetched in the background, never on the tick.** `resolve_tariff()`
may open an HTTPS connection, and the control tick is what feeds the ESP32
heartbeat - a supplier's TLS timeout there is a supplier able to push the
add-on towards layer 1's 30 s veto. A separate task refreshes every 30 minutes;
the tick reads an attribute and an age bound, and a plan whose refresh has
stopped answering retires rather than standing as if it were current.

**What is honestly not measured.** The acceptance criterion asked for
net-positivity across 12 months of *replayed* prices. This repository holds no
price history - `scenarios/` is P1 meter data - so that criterion is **not met,
and is not restated**. What `test_arbitrage.py` section 7 does instead is
replay a synthetic year and assert the property that does not need real prices:
every executed round trip cleared its own widened requirement, on all of them,
across a year containing 522 hours priced below zero and zero discharges into
one. And at the shipped 0.10 EUR/kWh threshold a realistic Belgian day/night
site **does not trade at all** - reported, not tuned. That number, and the
widening factors beside it, are one product decision and it is open.

**Provisional prices, and what may not rest on one (FCST-03).**
Day-ahead prices arrive once a day, around 14–16 h CET. Until they do, §8.3 says
the unknown hours use yesterday's profile and are flagged `provisional: true` —
and **no irreversible action may be taken on a provisional price.**
`app/price_forecast.py` does the fill, and the whole module is shaped by one
defect: a carried-forward point and a published one are the same three fields
and a bool, so the natural line `PricePoint(..., provisional=y.provisional)`
copies yesterday's `False` onto today's guess and produces stale data presented
as current. That is the meter-side SAFETY-18 failure one layer up — the reading
arrives on time, looks fresh, and is wrong.

**So the flag is never passed through.** Carried points are built by a
constructor that has no `provisional` parameter at all (the literal `True` lives
in one place, and there is nothing for a caller to forget or copy); published
points are passed through **by identity**, so there is no copy that could lose a
flag; and an audit re-derives the expected flag for every emitted point from the
set of hours the feed actually published and *raises* rather than repairing one.
A published point may itself be provisional — `eneco_eddy`'s current-quarter
imbalance print is — so the audit's rule for a published hour is "whatever the
feed said", never "False".

**The fill keys on the Brussels wall clock, not on `t − 24 h`.** A 23-hour day
has no 02:00 and a 25-hour day has two, so a fixed 24-hour subtraction shifts
every carried hour by one across a transition and misprices a whole day with
nothing to show for it. Re-running the fill after a late publication replaces
the carried hours with the real print: the number and the flag move together
because the series is rebuilt, never patched.

**Arbitrage now drops provisional hours instead of refusing the whole day.**
The refusal that read "the plan is provisional" would have switched arbitrage off
every morning on every dynamic site once carry-forward filling landed — a
part-firm series is the *normal* state before the print. What §8.3 forbids is
acting **on** a provisional price, so the provisional hours are dropped and the
spread, the cheapest hour and the dearest hour are computed over the published
ones only; fewer than two firm hours is still a refusal, which is where the
§11.4 fallback (wholly provisional) lands. The permitting variant now refuses to
be *constructed* around a provisional point at all. The cheap-window flag became
per-hour for the same reason, so a site's maintenance charge floor does not go
dark every morning.

**Real-time / imbalance products get measurably wider thresholds.** One widening
path, driven by a set of active uncertainty sources: the product's own (§9.2 row
2, derived from the declared `PriceProduct` and never from a plugin id) and
§11.4's "no weather data" row, which the spec widens by a different trigger and
which is the same mechanism. The combination rule is `max`, not a product — two
independent doubts must not compound into a threshold nothing can clear. The
caller supplies the base round-trip cost threshold, so OPT-04 still owns what
the number *is*; this owns only how much wider it gets and why, and the allowed
verdict reports the threshold it cleared.

**An irreversible action is a value, not a judgement call at each call site.**
`Irreversible` carries the hours an action commits — the maintenance drain's
window is derived from `maintenance.DRAIN_TIMEOUT_MIN` rather than typed twice —
and the gate looks at exactly those hours, blocking on a provisional one, on an
hour the series does not cover, and on no price data at all: three different
reasons, none of them an answer. ⚠️ It is the entry point for a start being
*justified by a price*; the calendar path main.py runs today is untouched,
because a gate that blocked on `no_data` would stop the monthly cycle on every
default install — §12's opening sentence.

**Zero new options.** No `config.yaml` key, no new defaults dict, no `o.get(k,
X)` in main.py: the module reads no options at all, so CFG-03 did not widen.

**The grid-following loop is now capped against the month's billed capacity peak
(SAFETY-07).** New `app/peakguard.py`. TARIFF-01 measured the quarter; this is the
consumer it was built for. Every cycle the guard projects the open quarter,
compares it to a ceiling and files ONE arbiter claim: a lower bound on the
setpoint at `P_PEAK`. That shape is both behaviours §10.1 asks for — a hard clamp
on charging and pre-emptive discharge — and it cannot cap discharge or export,
because a floor cannot. `excess = 900 × (projected − ceiling) ÷ remaining`, and
the floor is the MEASURED battery power plus that: a floor built on the standing
command winds itself to the rail whenever the inverter fails to follow, and one
built on delivered power cannot.

**It trips where SAFETY-18 only notifies, and the captures decided that.** On the
healthy reference day it engages on zero of 95 quarters; the action is
level-triggered, unlatched and re-derived every cycle, so a false engagement costs
one cycle of discharge rather than a parked battery. Replayed on 2026-08-20 the
billed quarter goes from **5029.8 W — above the inverter's own 5000 W rating — to
3485 W**, under the soft limit and without saturating. Quarters under the limit
across both committed days: 189/190 → 190/190. Two days is not thirty; AC 3 is
still TARIFF-01's open item and this does not claim it.

**Blindness forbids importing, and it is checked before the projection is
graded.** A fresh averager projects a confident 0 W, which is under every ceiling
there is; a meter gap longer than `meter_max_age_s` makes the projection read LOW;
and a frozen meter reading export projects 0 W at full coverage. All three become
a floor of 0 W, which is one-directional and cannot cost pack energy. That is the
backstop SAFETY-18 could not be — its alarm arrives at 300 s of a 900 s quarter
and commands nothing. The frozen term is `freeze.looks_frozen()` itself, evaluated
on this cycle's standing command, not a second copy of the conjunction.

**The ceiling is `peak_cap_w`, raised only by an OBSERVED reading.** `NO_RECORD`
and `NOT_OBSERVED` both leave it at the cap, so an unreadable peak store makes the
guard STRICTER. The gate is the reading's STATE, not whether a number happens to
be attached to it: today a `peak_w is not None` test behaves identically, and
depending on that coincidence is how a future last-known-peak field would silently
raise the ceiling off a read that failed. `headroom_w()` is deliberately not
consumed — `cap − peak` turns a zero peak into FULL freedom.

**Two defects found building it, both in the interaction rather than the
arithmetic.** A LIMIT does not out-vote another LIMIT — `arbiter.resolve()` applies
every limit at or above the winner's priority and takes the most restrictive — so
the peak floor (`lo ≥ 0`) and the maintenance charge-only shaping (`hi ≤ 0`)
resolved to a CONTRADICTION and 0 W, parking the battery in the exact quarter it
should shave. Fixed by suppression rather than priority: main.py drops the
maintenance shaping while the guard is engaged, which is what that branch already
did for the optional `peak_forecast_entity` — a key that is empty on a default
install, so until now the suspension could not fire on the site that needed it.
And thresholding `excess > 0` released the claim the instant it started working,
producing seventeen engage/release cycles in one 900 s quarter (§10.2's
oscillation) and leaving that quarter 49 W over; the threshold is now the floor
itself, which is the same arithmetic read as a level rather than as a switch.

**`PeakTracker.project()` no longer reads a private name with a default.**
`getattr(averager, "_last_net", 0.0)` failed quiet AND in the disarming direction:
a rename credits 0 W held, the projection reads low, and the money guard does not
engage. `QuarterAverager.held_net_w` is now a documented public property, read
without a default, and `Projection` carries it. TEST-13 item 4.

**No new option, and no fourth defaults site.** The ceiling is `peak_cap_w`, the
coverage threshold is `meter_max_age_s`, the freeze predicate is `meter_freeze_s`,
and the achievable discharge comes from `max_w` and the device rating.
`config.yaml` and `validate.py` are untouched; §14's `capacity_peak_limit_kw`
stays declined as the kW/W mismatch trap it is. On `meter_source: off` there are
no telegrams, so the guard does not run at all and the status page says NOT
COVERED — running it there would have blinded it permanently and stopped every
legacy install from charging, which is what the first cut did until
`test_clamp.py`'s export case caught it.

**Static day/night and fully-fixed tariffs, first-class (TARIFF-03).**
§9.2 puts a large share of the Belgian installed base on a static profile, and
an EMS that only works under dynamic pricing is an EMS for a minority. A new
`static_profile` option describes a two-period day/night tariff — or a single
fixed price — as JSON, and `app/static_tariff.py` turns it into the same
canonical `{start, eur_per_kwh, provisional}` series the fetched providers emit,
so nothing downstream learns that a tariff can be static.

**It is not a sixth provider plugin, and the reason is mechanical.**
`PriceProvider.capabilities()` is a classmethod, so a plugin's `PriceProduct` is
fixed when the class is written — but §9.2 makes STATIC and FIXED two different
products with two different required behaviours, and which one a site is on is a
property of the operator's data, not of any class. `StaticProfile.product`
derives it from the number of distinct prices instead. The second reason: the
plugin framework exists to stop a *fetched* figure being marked up twice, and a
number the operator typed as all-in has no publisher to disagree with.

**So the profile sits where §11.4 already says it does — under the providers, as
the fallback.** The configured-product path and the feed-is-down path are the
same generator differing in one flag: `provisional`. A static-tariff site runs
that path every day, so the degradation path for a dynamic site cannot rot
unused.

**"No arbitrage" is a type, not a boolean somebody has to remember to check.**
`TariffPlan.arbitrage()` returns a variant: the *allowed* arm carries the
spread and the cheapest and dearest hours, and the *refused* arm carries a
reason and nothing else — `.spread_eur_kwh` on a refusal raises rather than
answering 0.0. An optimiser cannot compute a trade from a refusal because it
cannot get the inputs out of one. Four refusals, all derived: the peak guard is
holding, there is no price data, the plan is provisional, or **the series is
flat**. That last one is AC2 — a fully-fixed tariff disables arbitrage by
arithmetic rather than by matching on an enum, and a dynamic feed that happens
to print one price all day is refused for the same true reason.

**Peak shaving outranks arbitrage, expressed so OPT-04 cannot skip it.**
`arbitrage()` takes `peak_guard_active` as a required keyword argument, so the
spread is not obtainable without first answering whether §10.1's priority 2 is
holding. §9.1's capacity tariff is a network cost charged whatever the energy
product is.

**Time-window shifting reaches the one shiftable load that exists.** main.py's
cheap-window flag — previously fed only by the `price_now_entity` /
`price_avg_entity` pair — now falls back to the static profile, and the Web UI
gained a **Tariff** row naming the window, the price, when it next changes and
why it is or is not cheap. The rule is *strictly* below the seven-day mean:
under a flat tariff `<=` would call every hour cheap and a fixed-tariff site
would top up from the grid at any hour, which is the opposite of
self-consumption only. Seven days rather than 48 hours because a 48 h mean taken
on a Friday is two-thirds weekend, and the same Tuesday 09:00 would come out
cheap or dear depending on when you asked.

**A profile must tile every day type exactly once, checked at startup.** A gap
is refused with the uncovered range named rather than filled in — an invented
price is one an optimiser would trade on — and an overlap is refused because
otherwise one of two prices wins silently on list order. The trap this catches
in practice: `22:00–07:00 mon-fri` does **not** cover Saturday morning, because
a window applies on the day the clock reads.

**validate.py rule 7's stated reason expired, exactly as TARIFF-02 predicted.**
That ticket left a tripwire in `test_price.py` section 12 asserting that nothing
in `app/` consumed `load_prices()`, with instructions to fix the refusal's
wording in the same commit when it went red. It went red naming
`static_tariff.py`. The message now says what is true instead — the series *is*
consumed, but main.py acts on the static side of it and not on a fetched one —
and the tripwire is re-aimed rather than removed: the consumer set is pinned, so
a second consumer (OPT-04) goes red again, and the expired phrase is asserted
gone so a bad merge cannot restore it. Rule 7 also grew a second refusal:
`static_profile` alongside the Home Assistant price pair is the same
silent-loser conflict, and a live one.

**No new place for an option default (CFG-03 is not widened).**
`static_profile` was added to `app/price.py`'s `PRICE_OPTION_DEFAULTS` — the one
in-code mirror TARIFF-02 created and pinned to `config.yaml` by a check — rather
than to a defaults dict in the new module, which has none. main.py grew no
`o.get("static_profile", …)` fallback. That is TARIFF-02's pattern, stated
here so a reviewer does not have to infer it.

**Measured, on the data that exists.** `test_static_tariff.py` replays the two
committed days of this house's real P1 meter (`scenarios/p1-net-2026-08-{20,23}`)
priced under a Belgian day/night profile. 2026-08-20 is a Thursday: 15.220 kWh
imported, 9.606 in the day window and 5.614 in the night one, EUR 4.7185 under
day/night against EUR 4.6725 flat, and shifting the day-window energy into the
night window saves EUR 0.8069 — which is 9.606 kWh x the 0.0840 spread, derived
twice and required to agree. 2026-08-23 is a Sunday: 0.728 kWh, **all** of it
night-priced, which is the weekend row proving itself on real data. §17 names
six Belgian scenarios and a static-profile case is not among them; this is the
seventh, and it is two captured days rather than a season — the file says so
rather than implying an annual figure.

Also worth knowing before OPT-04: on realistic Belgian day/night numbers the
whole-tariff spread is about 0.08 EUR/kWh, **below** the shipped
`cycle_cost_threshold_eur` of 0.10. A correct optimiser will decline to cycle
for the tariff alone on such a site. That is a product finding, not a defect.

**Price provider plugins, and one canonical all-in price format (TARIFF-02).**
`app/price.py` ships five providers - `generic_dayahead` (wholesale EPEX BE),
`engie_dynamic`, `totalenergies_dynamic`, `eneco_eddy` and the offline
`manual_csv` - all emitting exactly `{"prices": [{"start", "eur_per_kwh",
"provisional"}]}`, tz-aware at the Brussels offset, EUR/kWh, all-in. **Nothing
consumes it yet**: this is the seam, landed first, so that the arbitrage work
that follows never learns a supplier's name.

**Double-counting the markup is refused rather than discouraged, given a
correct declaration.** A plugin does not return a price; it returns the figure
its source published and declares which of {energy, markup, taxes} that figure
contains. The framework adds the complement, once. Engie and TotalEnergies
publish a retail figure with their margin already inside; adding the configured
3 c/kWh on top would make every future arbitrage decision wrong by that amount
with nothing reporting it.

Two mechanisms, because one was not enough. `PluginOptions` raises on the
framework-owned keys - and review of PR #16 then wrote a double-counting plugin
three more ways, one of them touching nothing but `PRICE_OPTION_DEFAULTS`, a
public dict in the module every plugin already imports. So `register()` now also
**refuses at import any plugin whose own code mentions a framework key**
(`code_leaks()`, docstrings stripped via `ast` so a plugin's mandated
`SOURCES:` line does not false-positive). Honest wholesale baseline 0.188
EUR/kWh; every attack landed 0.219.

**The guarantee's precise shape**, stated because an overclaim is worse than no
claim: the framework adds the complement exactly once *given a correct `sources`
declaration*. A plugin that declares `{energy}` while its feed is really all-in
still comes out one markup high while touching no framework name, and no code
can catch that - the declaration is the plugin's unverifiable claim about
somebody else's tariff. What is mechanised is reachability; what stays human is
truthfulness.

**The product type is declared, not inferred from the plugin's name.**
`eneco_eddy` is a real-time/imbalance product and gets widened arbitrage
thresholds; that reaches a consumer through `capabilities().product` and
`volatility_regime()`, never through a name comparison. `test_price.py` asserts
mechanically that no plugin id appears anywhere in the framework's own code,
and drives a sixth provider defined inside the test file to the correct regime
with `price.py` untouched.

**"No price data" is a type, not an empty list.** Following the same rework
`sim/riglock.py` took, a fetch returns `PriceRead.series(...)` or
`PriceRead.no_data(reason)`, and only the first carries fields. An empty series
is not constructible; `no_data` has no `.points` to iterate, no `.at()`, and
`.get("points", [])` raises rather than returning the default. **A day of free
electricity and a day with no prices are different instructions** - arbitrage is
pointless under the first and forbidden under the second - and an implementation
that returns `[]` on a failed fetch cannot tell them apart. Every site-level
failure returns `no_data`; nothing raises into the control tick.

**What is not verified, and cannot be by an offline suite:** every fixture is
synthetic, so the five feed URLs and JSON shapes are checked against a belief
about what those suppliers publish, not against the feeds. The parse layer is
exercised; the contract with each supplier is not. Expect the first live fetch
of each to need a shape fix, and note that a wrong shape degrades to
no-price-data rather than to a wrong price.

**`manual_csv` is the documented offline fallback** and needs no network, no
supplier account and no HTTP stack. See DOCS.md for the format. Its `start`
column must carry a UTC offset - a bare local timestamp is refused, because one
hour every October names two moments.

**Injection is priced separately** (`export_price_eur_kwh`), a scalar rather
than a second series, and negative values are accepted. Consumption plugins
cannot reach it.

**`price_provider` may not be combined with `price_now_entity` /
`price_avg_entity`** - the add-on refuses to start on that pair. The Home
Assistant sensors are the existing delegation of pricing to HA and still work
alone; with both configured they would silently keep deciding while the status
page named a supplier that decided nothing.

**No timezone database is used.** The shipped Alpine image has no `tzdata`, so
Brussels offsets are computed from the EU rule and every instant is carried as
UTC and rendered at the offset that applies to it. The checks drive both 2026
transitions - the 23-hour and 25-hour days - rather than a normal day, where a
fixed 24-slot implementation and a correct one agree.

**The capacity-tariff peak is now measured, banked and survives a restart
(TARIFF-01).** New `app/tariff.py`. The 15-minute block average itself was
already built and unwired - `p1.QuarterAverager` computes clock-aligned 900 s
blocks, time-weighted over irregular telegram spacing, offtake-only and
per-phase - and `P1Ingest.blocks` was a 96-entry list a restart erased. This
adds the SQLite behind it (`/data/tariff.db`): one row per completed quarter in
the indefinite 15-min tier, one row per calendar month in the permanent peak
tier, and no `DELETE` against either anywhere in the module.

**Blocks, not a sliding window.** The card says "rolling"; 9.1 bills "the
highest 15-minute average offtake", stated per QUARTER. "Rolling" here means
continuously updated, not a trailing-15-minutes window - a trailing window
reports a peak the meter never bills, and reports it HIGH, because a burst
straddling :15 fills it completely while contributing to two block averages at
partial weight.

**A restart mid-block resumes it, and a restart that outlives the block banks
it.** `QuarterAverager` grew `snapshot()`/`restore()`; the tracker checkpoints
the open block every 15 s (worst case 1.7 % of one quarter) and clears the
checkpoint in the same breath as a block close. The held sample keeps its
ORIGINAL timestamp across the restart, so the outage is integrated under the
same `max_hold_s` rule a running process would apply and uncovered seconds drag
the closed block DOWN rather than being held forward. `quarter_start` is the
primary key and a re-bank may only ever RAISE a figure, so a replay cannot
double-count and a partial re-run cannot lower a quarter already measured.

**An unreadable store is not a zero peak.** `peak()` answers `OBSERVED` /
`NO_RECORD` / `NOT_OBSERVED` and carries `peak_w = None` unless observed - the
`riglock.holder()` fix applied to something that spends money. A peak read as
0 W hands SAFETY-07 maximum headroom and silently disarms the guard at the
moment its own storage breaks, so there is no accessor that defaults it: a
caller who forgets gets a `TypeError`, and the status page prints a different
row for "no quarter yet" than for "the database did not answer".

**Calendar months are LOCAL, blocks are UTC.** A quarter starting 22:15 UTC on
31 August bills in September. Block alignment still needs no timezone.

**No new option, and no second peak entity.** The soft limit is the existing
`peak_cap_w`; `config.yaml` and `validate.py` are untouched, so CFG-03 is not
widened. `peak_forecast_entity` stays the input to `maintenance_charge_floor()`
- repointing that consumer changes what moves energy and belongs to SAFETY-07.

**It measures and nothing else.** No claim, no clamp, no command; graded on the
parsed syntax in `test_tariff.py`, which also proves the same grader lights up
on `main.py`. Per SAFETY-20, the quantity is DELIVERED power - `tariff.py` only
ever sees `P1Sample`, the meter's own reading, never a setpoint.

**AC 3 is not met and the check says so.** It asks for 30 days of replay; two
days of this house's real meter exist (2026-08-20, 2026-08-23), now committed as
`scenarios/p1-net-*.jsonl.gz`. `test_tariff.py` part 9 replays them, reports
the percentage under the shipped limit next to the sample size in days, and
asserts nothing about the 95 % figure.

**A computed setpoint over the rating now alarms instead of being clipped in
silence (SAFETY-03).** The clamp was never the missing half - specs.md 10.3
records an unclamped integrator climbing ~130 W per 4 s past 10 kW and reporting
**14 768 W** while pinned at the 5 kW rail, and a silent clamp turns exactly that
into a flat, legal-looking 5000 W hold. `write_setpoint()` - the one path every
write takes, cycle, heartbeat, operator button and shutdown alike - now clamps to
the **validated** `max_w` and reports whether the raw value differed, and
`Controller.check_output_clamp()` raises `OUTPUT_CLAMP`: a persistent
notification (`goodwe_output_clamp`), an `ERROR`, an event row and a red status
banner. **Not a trip**: the clamped value is written and the loop keeps running,
per 11.2's "clamp *and alarm*".

**The bound is `max_w` and nothing is `min()`ed at runtime.** Section 14 requires
the loop clamp and the driver's `setpoint_limit_w` to agree, validated at startup
with a refusal on mismatch - which CFG-01 already ships. The device range is
still applied second, as defence in depth; after a successful validation it
cannot engage, so if it ever does, that is louder than the first stage and it
alarms too, naming the bound. No new config key: the guard's limit is the one
the operator already set, and its cadence is a module constant.

**`Decision` grew `i_clamped` / `clamped` booleans, and that is the load-bearing
part.** `reason` is one string that later stages overwrite - a cycle that clamps
the integrator and is then slew limited reports `slew-limited`, and one that
clamps while saturated reports `saturated-freeze`. Both *are* the winding case,
so an alarm keyed on the string is silent on exactly the cycles it exists for:
mutating main.py back onto `"clamp" in decision.reason` drops the 10.3 replay
from 2 engagements to **0**. The flags are set where the clamp happens and are
never overwritten. On a default install (`integrator_max_w` following `max_w`)
the integrator bound is the one that bites and the output clamp is unreachable,
so the integrator flag is the whole signal.

**Distinct from the saturation freeze, and checked in the negative.** A clamp is
a *software* bug signal; saturation is the *hardware* failing to follow a legal
command, and its action is to freeze the integrator, not to alarm. They share no
reason code, notification id, status row or counter. `test_clamp.py` drives a
loop that saturates for 10 consecutive cycles at 1.4 % of the rating and asserts
zero clamp notifications, zero engagements, no log row and no status row - after
first asserting that the saturation freeze *really did* engage, so the silence is
an observation rather than a scenario that did nothing.

**The action log is coalesced per episode, and that was a real bug.**
`log_event()` keeps 40 rows; one row per engagement flushed the entire ring
within 40 cycles of a winding loop, and the first thing it evicted was
SAFETY-18's own `METER FROZEN?` row - this guard silencing another guard's log.
Caught by `test_freeze.py` going red, not by inspection. One row per episode now,
carrying the exact engagement count read off the counter.

**A non-finite setpoint is refused, and it was worse than passing one through.**
`nan` is loss of input (11.1 rule 1: command zero, keep commanding), not an
out-of-range value. Measured: because every comparison against `nan` is False, a
naive `max(-limit, min(limit, nan))` keeps its *first* argument and turns a `nan`
into a command **at the full rating**, then reports it as a clamp engagement.

**Review fixes (PR #12).** Three, and the first was a behavioural regression
this PR introduced: `clamp_to_rating()` **returned early** from the `max_w`
stage, so the device bound - the "applied second, defence in depth" half - was
reached only when the rating bound had NOT fired. The bigger the overshoot, the
less the guard did. At `max_w` 5000 against a +-4000 device, +9999 emitted
**+5000 W, outside the device range**, while a mere +4500 was correctly held to
+4000. That undid the pre-PR guarantee "never send a value the device will
reject outright" and undercut this module's own case for clamping to `max_w`
rather than a runtime `min()`, which rests entirely on that stage catching a
validation that did not run. Both stages now always run, the result names both
bounds when both bite (`max_w+device`), and the property is asserted over a
492-point sweep rather than on witness values - the defect was invisible to any
single-value check that happened to sit inside one stage. No reachable
production path was found for it (two validation gates and the arbiter bound sit
in front), but latent is not harmless.

Two harness defects, both of the kind where a check cannot fail for the value it
forbids: the hysteresis check issued `range(CLEAR_CYCLES - 1)` clean cycles,
which at `CLEAR_CYCLES = 1` issues **zero** and passes while printing
"0 clean cycles do NOT clear it" - green for the chatter setting the constant
exists to prevent. And the re-assert check stimulated at `REASSERT_S + 9`, green
for **any** value including 3600, i.e. the hour of silence the module promises
not to produce. Both are now written in wall-clock terms that do not move when
the constant does: **one** clean cycle must not clear the alarm, and a loop
pinned across the 900 s capacity-tariff quarter must speak at least 5 times and
on no more than 20 % of cycles. `CLEAR_CYCLES = 1`, `REASSERT_S = 3600` and
`REASSERT_S = 0` each go red.

**`test_clamp.py`** (new, assert-based, no framework) covers the clamp
arithmetic and its boundaries - exactly +-`max_w` is *not* an engagement - the
alarm's edge/re-assert/hysteresis cadence, the 10.3 runaway replayed through the
real `Controller.cycle()` against an unclamped transcription, the saturation
negative, and the direct-write path. Six author mutations reproduced red,
including "clamp kept, alarm deleted" (the historical build) and "alarm keyed on
`reason`". Two of the check's own hypotheses about how the freeze and the clamp
interact were **refused by the sweep** and rewritten to what it measured: a
frozen cycle *can* clamp, and does so exactly when the integrator reverses.

`test_freeze.py`'s notification assertions now filter by `notification_id`
rather than by service - the add-on raises more than one kind of notification,
and its railing scenarios legitimately raise this one too.

**Cheap-window maintenance charging was unreachable at the shipped defaults, and
now the config that makes it unreachable is refused (MAINT-04).** `config.yaml`
shipped `max_w: 2000` against `maintenance_charge_w: 2500`. The cheap-window
shaping files `Claim.limit(hi = -maintenance_charge_w)` against the supervised
SAFETY limit's `lo = -max_w`, so `lo > hi` — the arbiter's clause 3, correctly
resolved to **0 W** and logged as a CONTRADICTION. **Maintenance silently never
charged in a cheap window on a default install**, and the only evidence read
like a bug in the arbiter rather than like the two numbers an operator set.
Measured against the real arbiter, loop asking +500 W:

```
max_w=2000  hi=-2500  ->     +0 W  contradiction=True
max_w=2500  hi=-2500  ->  -2500 W  limited by supervised(max_w)
max_w=3000  hi=-2500  ->  -2500 W  limited by maintenance(cheap-window charge)
```

`validate_startup_config()` now refuses `maintenance_charge_w > max_w` (rule 6),
the same shape as CFG-01's `sign_disagree_w >= max_w`: a number the operator set,
overridden by a layer they cannot see. Equality is reachable (`-max_w` is a
single valid point) and deliberately not refused. Verified by starting the
add-on on the pre-fix pair: exit 78, the refusal banner, no traceback.

**`maintenance_discharge_w` was checked, not assumed, and it is NOT the same
case.** The shaping is not symmetric. The drain phase does not shape the loop,
it OWNS the actuator — `Claim.set(P_MAINTENANCE, +discharge_w)` — and a SET is
*clamped* by the limits above it rather than contradicted by them. 2500 against
a 2000 W ceiling emits **+2000 W**, and the arbiter's own `explain()` prints
`maintenance -> 2000 W, limited by supervised(max_w)` every cycle. Nothing is
silent and energy still moves; the cycle completes slower (13.8 kWh usable at
2000 W is ~6.9 h against the 720 min drain timeout). **Warned about, not
refused.**

**The shipped defaults changed so the validator does not reject them.**
`maintenance_charge_w` and `maintenance_discharge_w` are now **2000**, at the
shipped `max_w`. `max_w` was deliberately NOT raised instead: it is the
supervised safety ceiling and raising a default upward ships more power to every
new install. Raising `max_w` (the reference install runs 5000) is what unlocks a
faster maintenance cycle — raise it first, then these. `test_validate.py` now
reads `config.yaml` itself and asserts the shipped numbers pass the rule, so a
default that the validator refuses cannot ship again.

### ⚠️ Upgrading an existing default install: the add-on may refuse to start

An install still carrying `maintenance_charge_w: 2500` against `max_w: 2000`
now meets rule 6 and **exits 78 with the refusal banner** instead of starting.
That is the intended behaviour — the configuration was already broken, it was
just silent about it — but an add-on that will not come back after an update
needs the fix stated here rather than only in the log.

**The fix, either way: set `maintenance_charge_w` to `2000` or lower, OR raise
`max_w` to `2500` or higher** (Settings → Add-ons → GoodWe RS485 Controller →
Configuration), then start it. The banner in the log says the same thing and
names both numbers.

Whether you hit this at all depends on whether the Supervisor kept the old
value, and **both cases are covered above because we did not verify which one
applies here.** The Supervisor merges an add-on's *stored* options over the
defaults in `config.yaml`, so:

- **A key the operator never edited** re-derives from the new `config.yaml` and
  becomes 2000 on upgrade. Non-event, nothing to do.
- **A key the operator did set** persists at 2500 and the add-on refuses to
  start until it is changed. Note this includes anyone who used the
  Configuration page's YAML editor, which writes the whole options block back —
  so "I never touched that field" is not a guarantee the old value is gone.

⚠️ **This is reasoned from the Supervisor's documented merge behaviour, not
measured on the live install.** The reference install's stored options could not
be read (`/api/hassio/` refuses a long-lived token) and probing it was out of
scope — it regulates a real battery. It runs `max_w: 5000` and is unaffected
either way.

⚠️ **Known, deliberately left, and worth a follow-up:** `main.py` and
`maintenance.py` still fall back to **2500** when the key is absent from the
options dict, while `config.yaml` and `validate.py` now say 2000. Unreachable on
a Supervisor install (every key is always supplied) and reachable only from
tests/`sim`. The real fix is one source of truth for option defaults, which is
bigger than this ticket and touches files another ticket is in.

## 0.5.2

**A meter frozen at a wrong-but-fresh value now raises an alarm. It still is not
stopped.** (SAFETY-18)

A meter that answers `200 OK` with a stale document forever is *arriving* on
every poll, so every freshness mechanism in the stack reads it as healthy:
`meter_max_age_s` is reset by each response, the firmware watchdog is fed by the
heartbeat and by a sample age that never climbs, and the arbiter contributes only
±`max_w`. The loop integrates a constant error and rails — replaying the real
`compute()` at 15 s cycles, in 150/60/30 s at `max_w: 2000` for a meter frozen at
350/1000/≥3000 W, and 360/135/75 s at `max_w: 5000`. **The live install runs
`max_w: 5000`**, which is both the schema maximum and the device rating.

The new detector (`app/freeze.py`) is a **conjunction**, not the bare
`unchanged_s` threshold TEL-05 measured and rejected:

    unchanged_s ≥ meter_freeze_s   AND   |command| within 5 % of max_w

Both terms already existed and were already on the status page. They fail for
independent reasons: at the converged point the command is nowhere near the rail,
and at the rail against a real load the kWh registers move within seconds. At the
shipped 300 s there is also a structural relation, not just a margin — the
slowest a healthy 1 Wh register can tick and still look unchanged is 12 W, which
is inside the 15 W deadband, so the command there is parked at 0 W and the two
halves cannot both be true.

**Notification only, deliberately.** An HA persistent notification (fixed
`notification_id`, so a repeat replaces rather than stacks), an `ERROR` log line,
a UI event and a red status banner. No claim, no latch, no 0 W: a false trip
parks the battery and the operator's next move is to switch the guard off; a
false notification costs an alert. The trip is a separate decision that needs
data from a real install.

- New option `meter_freeze_s` (default 300, range 60–3600, no "off").
- **Only `meter_source: homewizard_local` is covered** — it is the only transport
  that sees the meter's own document. Everywhere else the status page says
  **NOT COVERED** rather than "armed".
- `DOCS.md` now states the **capacity-tariff exposure** this does *not* close:
  `peak_cap_w` caps maintenance charging and nothing caps the loop, so a lying
  meter can drive a 5 kW unattended import, and ≥15 minutes of that sets the
  month's billed peak (measured: one 4998 W quarter moved a monthly peak from
  4832 → 5013 W). That cap is SAFETY-07.
- `test_freeze.py` grades it in both directions: silent over 40 minutes of the
  live house's real converged capture (committed as
  `scenarios/p1-age-2026-08-26.jsonl.gz`), silent over an hour of TEL-05's
  low-power case where a bare threshold is shown to fire, and alarming on a
  genuinely frozen meter driven over real HTTP through the real
  `HomeWizardLocalSource` — against the rig's own `hwsim` `freeze` fault when the
  infra repo is reachable, an equivalent stub otherwise. Its negative twin is the
  load-bearing one: the same railed command against a meter that is still moving
  must stay silent at the same threshold.

## 0.5.1

**SAFETY-02's latch now survives an add-on restart, which is the difference
between a latch and a pause.** `auto` was persisted and `sign_wd.tripped` was
not, so reconstructing the `Controller` against the same store — exactly what a
container restart does — brought it back **armed, into the same fault, with no
operator involved**: `tripped=False`, banner back to `NOT READY`, and it resumed
commanding (measured: −1000 W, −2000 W) before re-tripping 180 s later. Under a
Supervisor that restarts a crashed add-on that is a loop of three-minute
wrong-direction bursts — and "restart the add-on" is the first thing an operator
looking at a `FAILSAFE` banner and a 0 W inverter will try.

`sign_tripped` and `sign_reason` are now `Store` keys beside the maintenance
state, written the moment the trip happens (reason first, flag second, so a crash
between the two can never persist a latch with no explanation) and restored in
`Controller.__init__`. Cleared only by the operator's start. `DOCS.md` said "it
does not clear itself"; that was false across a restart and now is not.

Also in this release, all of it review follow-up rather than new behaviour:

- **The arming gate is tested against the real arbiter.** It was the decision the
  whole ticket turns on and it was asserted in three comment blocks and verified
  by nothing — every case passed `following` in as a literal. `test_watchdog.py`
  now derives it from `resolve()` across twelve claim configurations, including
  the redirect case `winner == "loop"` alone would have got wrong and the `lo`
  limit that raises magnitude without flipping sign.
- **The MQTT `status: "failsafe"` value is tested.** The check that claimed to
  cover it restated the assertion three lines above it and never touched the
  payload.
- **A forward hazard written down where it will be read, and made to bite.**
  `winner == "loop"` disarms the watchdog under any higher-priority SET claim.
  Today only maintenance and the safety stops emit one and neither follows the
  meter; a peak-shaving strategy at `P_PEAK` **does**, and would blind the
  watchdog for every peak event.

  The gate is now `watchdog.loop_is_following()`, called by both `main.py` and
  the test, for the same reason `disagrees()` was extracted: the test held a
  *transcription* of the shipped expression, so editing the shipped gate alone to
  `winner in ("loop", "peak")` left every suite green while the watchdog went
  blind. Two copies, and the one that drifts is the one doing the grading.

  ⚠️ Extracting it was not sufficient on its own. The realistic P_PEAK row —
  peak shaving asking for *more* than the loop — still could not detect the
  widened gate, because the magnitude term rejects it regardless and masks the
  winner term. A second row with peak asking for *less* isolates `winner` as the
  only deciding term, and that is the one that turns red.
- **The worked example described a configuration that does not ship.** At the
  defaults `maintenance_charge_w` (2500) exceeds `max_w` (2000), so that claim is
  an arbiter *contradiction* resolving to 0 W, not a −2500 W charge. The example
  is now written at 1500 W. (The contradiction is a real pre-existing bug —
  a default cheap-window maintenance charge silently never happens — filed
  separately, deliberately not fixed here.)
- **`disagrees()` is public**, so `sim/replay_sign.py` stops carrying its own copy
  of the predicate it grades.
- `held_s` takes the clock instead of reading one, restoring the pure-function
  property the module advertises.
- **The `sign_reason`-before-`sign_tripped` write ordering is enforced, not just
  documented.** Swapping the two lines used to pass every check. A store that
  fails the second write now proves the property the ordering exists for: what
  survives a crash mid-trip is a reason with no latch, never a latch with no
  explanation.
- Measured figures corrected throughout: 93.4 % / 4212 of 17280 (was 92.9 % /
  4127 of 16349), per-cluster 0.3 / 0.9 / 10.5 / 0.0 %.

## 0.5.0

**SAFETY-02 — the sign-disagreement watchdog.** The add-on now watches whether
the setpoint it is emitting and the grid figure it is regulating on point the
same way. If they oppose each other, both over `sign_disagree_w` (300 W), for
`sign_disagree_s` (180 s) **continuously**, it files a failsafe claim at
`P_SAFETY`, commands **0 W**, and latches there until an operator starts control
again. New module `app/watchdog.py`; new options `sign_disagree_w` and
`sign_disagree_s`.

This is the gap the existing failsafes leave open. The stale-input watchdog asks
whether the meter is still talking, the saturation freeze asks whether the
inverter can follow, the output clamp asks whether the number is too big — a
loop can pass all three while charging the battery out of the grid, because the
only thing wrong is the sign. On 2026-08-20 a commercial controller did exactly
that four times in one day, for up to 77 minutes at a stretch, and reached a
commanded −14 768 W against a ±5000 W inverter.

Validated against that day. `sim/replay_sign.py` in the infra repo drives this
module over the committed RS485 captures of both 2026-08-20 (the four episodes)
and 2026-08-23 (healthy): **18 trips on the fault day, every one inside a window
`NOTES.md` §4.7 accounts for, and 0 on the healthy day.** The four episodes do
share one signature rather than merely resembling each other — sign agreement
runs 0.3 / 0.9 / 10.5 / 0.0 % inside them against 93.4 % on the healthy day.

Three decisions worth knowing about:

- **It is disarmed whenever the grid-following loop is not what is driving.**
  Sign agreement is a property of a loop that follows the meter. A maintenance
  grid charge opposes the meter *by design* — 2500 W of charge into a house
  drawing 500 W is +3000 W on the meter against a −2500 W command, for hours —
  and armed unconditionally this would trip on every maintenance cycle. The gate
  tests the emitted number, not the winning claim's name: a limit claim can let
  the loop "win" and still replace its value with one pointing the other way.
- **The trip latches.** Commanding 0 W removes the command side of the
  comparison, so the disagreement always clears the moment the trip works.
  Un-latched, this would trip, clear, and hand the actuator back to the loop that
  was driving the wrong way, every three minutes forever.
- **`sign_disagree_w` at or above `max_w` is refused at startup** (validator rule
  5). The arbiter never lets a command exceed `max_w`, so such a threshold makes
  the watchdog incapable of firing while the status page still reports it armed.
  A disarmed guard that advertises itself as armed is the failure mode this
  project keeps refusing to ship.

What it does **not** catch, stated so nobody assumes otherwise: an inverter that
acts opposite to what it was told moves the meter *further* the same way, so
command and meter agree while everything runs away. That is the saturation
freeze's job, which compares the command against the battery's actual power.

`test_validate.py`'s "the smallest sane max_w still starts" case now names a
reachable sign threshold, because a 1 W ceiling genuinely does disarm the 300 W
default and is now correctly refused.

## 0.4.0

**CFG-01.** The add-on now refuses to start on a configuration it cannot honour,
instead of logging a line and carrying on. `validate_startup_config()` runs in
`amain()` before the broker, the web server or any task exists, so a refusal
cannot leave half a controller running behind a UI.

It refuses only where continuing produces one of two outcomes: **(a)** a stated
limit is untrue - the operator configured a bound and a lower layer silently
enforces a different one - or **(b)** the add-on is inert and indistinguishable
from healthy. (b) is not a special case of (a): with no `setpoint_entity`, layer
1 pinning the inverter at 0 W is the failsafe *working*. The system does not
become unsafe, it becomes a convincing impression of a working controller.

Four things qualify:

- **`max_w` above (or below) the setpoint device's own rating.** This is the one
  that changed behaviour. It used to warn `the device wins` and run anyway, with
  `write_setpoint()` silently clamping, so the configured ceiling was never the
  real one and the only evidence had scrolled off the log. The refusal names the
  configured value, the device's value, and which wins.
- **No `setpoint_entity`.** A controller with no actuator: every write is a
  no-op, the device rating is never read, and the ESP32 heartbeat that holds off
  the layer-1 failsafe never happens. It ran, looked healthy, and moved nothing.
- **A `meter_source` that can never produce a reading** - missing `p1_host`,
  missing `p1_net_entity`, a blank MQTT topic, a phase list disagreeing with
  `meter_phases`, an unknown source name, or a `meter_poll_s` at or past
  `meter_max_age_s`. These were already caught, but they only *disabled
  ingestion*, which leaves the grid reading None forever: permanent "inputs
  missing", 0 W, and a status page indistinguishable from a healthy meter nobody
  has sent anything to yet.
- **`meter_max_age_s` above the firmware's ingest veto.** The schema accepts
  `int(5,300)`; the ESP32's threshold is a compile-time `max_age_s: "30"` and it
  trips layer 1 at `p1_age_s >= 30`. Set 120 and the add-on treats a 119 s
  reading as fresh and keeps commanding, while the firmware has already vetoed
  and is holding the inverter at 0 W - healthy meter, running controller,
  nothing moving, and no line anywhere saying the number you set was overruled.
  Structurally the same fault as the `max_w` case. Only *above* the veto is
  refused: below it the add-on gives up first, which is fail-safe and visible on
  its own status page. Mirrored as `validate.FIRMWARE_INGEST_VETO_S` - it cannot
  be imported from a compile-time YAML substitution, so it carries a note about
  the one direction in which it can go stale.

`max_w` at or below zero is refused too. Schema-blocked on a real install, but
`validate_startup_config()` is now also the simulation suite's entry point, and
a non-positive `max_w` inverts the arbiter's `supervised` bounds into a
contradiction that resolves to 0 W - i.e. it would read as a safety trip rather
than as the typo it is.

The predicates for the meter group still live in `p1.validate_meter_options()`,
which `build_source()` also calls - one definition, two callers, not two
competing validators. Two transports gained a check they never had: `ha_dsmr`
with only one of its two entities, and `mqtt_p1` with a blank topic, both of
which previously built a transport that could never read.

Deliberately **not** refused, because it is untidy rather than unsafe:
`integrator_max_w` above `max_w` (the arbiter still bounds the wire to ±max_w,
so the cost is unwind latency, not amperage - warned), a missing `soc_entity` /
`batt_entity` (fail-safe, and already a red NOT READY banner naming it), a
`setpoint_entity` that has not appeared on the Core API yet (ESPHome is routinely
late; refusing would make a slow boot need a human - re-checked once the real
rating arrives), and a thin-but-nonzero `meter_poll_s` margin.

A refusal is legible as a refusal: a banner that says in words that this is a
configuration problem and nothing is broken, the numbered problems, no traceback,
and exit status 78 (`EX_CONFIG`).

**SAFETY-12.** The control loop now evaluates on a clock as well as on a change.
`run_control()` ran `cycle()` only when the meter value differed from the
previous poll, so a meter that repeated a value ran **no control cycles at
all** - while the heartbeat went on re-emitting the last command and the
firmware went on putting it on the wire every 2 s. Layer 1 stayed fed, every
indicator read healthy, and a command computed from an old reading kept being
executed through whatever the load did next. Measured on the rig during TEL-05:
70 s at a literally constant 350.0 W with 14 arrivals and zero cycles.

**It was worse than a latency limit: it was a safety claim that could not
fire.** `stale_input_s` and the `inputs missing -> 0 W` claim both live *inside*
`cycle()`. Pre-fix, if the meter value happened to be constant, `cycle()` was
never called again - so that claim had no way to run. Constant meter, SoC entity
dying at t=4 s, same binary:

```
pre-fix    writes: [(0.0, 220.0), (0.0, 220.0), (10.1, 220.0)]
           final target: 220.0 W   reason: 'loop -> 220 W'
post-fix   writes: [(0.0, 220.0), (0.0, 220.0), (3.0, 430.0), (9.1, 0.0), (10.1, 0.0)]
           final target: 0.0 W     reason: 'safety -> 0 W'
```

The pre-fix controller held 220 W through the entire input outage with its
status line still reading `loop -> 220 W`, and layer 1 does not cover that case
on a `meter_source: off` install: no P1 age sensor is published, so the firmware
has nothing to veto on, while the heartbeat keeps its watchdog fed. So the old
behaviour was not "holds a stale command safely" - it was a failsafe that a
repeating meter switched off. `test_loop.py` asserts both halves.

`cycle_floor_s` (default **15 s**) is a deadline on the gap between cycles, not
a second scheduler: every cycle resets it, whichever trigger fired. On a healthy
meter it never elapses, so the cycle rate, the write sequence and therefore the
per-cycle gain the tuning was measured against are unchanged - asserted in
`test_loop.py` by driving the real loop twice on one binary (a huge
`cycle_floor_s` reproduces the old behaviour exactly) and comparing the
commanded values.

The reset is not the only thing holding that up, and the other argument is the
sturdier one: the combined cycle rate is `max(arrival rate, 1/floor)`, never the
sum. At 1/15 s the floor is strictly slower than the ~1/5 s arrival rate, so
even a hypothetical floor racing the arrivals could not exceed the nominal
healthy rate. The re-tune is unreachable by construction. ⚠️ That holds only
while `cycle_floor_s` stays above the meter's update interval, and the schema
minimum is 5 s - which is exactly the reference meter's interval. Nothing
enforces the relationship; `validate.py` is the obvious home for it and does not
have it yet.

Observed over six real minutes on the same rig, one binary, meter held at a
literally constant 350.0 W: **before, 1 cycle** (at startup) and then nothing,
the command stuck at 220 W against a 350 W import for the rest of the window.
**After, 24 cycles**, worst gap 15.14 s, and the ordinary PI ramp doing its job -
220, 430, 650, 860, 1080, 1300, 1510, 1730, 1940, 2000 W, one step per cycle,
rail at t=136 s and flat after. Three minutes with the meter parked at the
target instead: **1 cycle before, 12 after, and 19 writes either way - the same
19**. Three minutes on a normally-changing meter: **36 cycles and 49 writes on
both sides, the same values in the same order.** The heartbeat write count never
moves; the only writes the fix adds are the ten the loop was supposed to be
making all along.

15 s is bracketed, not round.

**Below it, ~11 s, and it arises more routinely than a dropped poll.**
`HomeWizardLocalSource` polls at 5.0 s against a meter that updates every 4.97 s
(NOTES.md:640). Two near-equal periods beat, so roughly once every ~14 minutes
the poller steps over an update entirely - a ~10 s gap between changed values on
a perfectly healthy meter, plus up to 1 s for the 1 s tick to notice. A failed
HTTP poll (retried on the next tick, no backoff) adds another 5 s. The floor
sits above the beat-skip so it does not fire on that alone; what it takes to
reach 15 s is a beat-skip *and* a failed poll coinciding.

**Above it, 30 s.** `meter_max_age_s` is 30 s, past which there is no reading to
evaluate, and the firmware's ingest veto is also 30 s - so the loop is
guaranteed to have re-decided at least twice inside any window in which layer 1
might otherwise be the only thing deciding. ⚠️ The default is therefore not
independently justified at 15: if the firmware's veto ever moves, this moves
with it. Noted beside the constant in `main.py`.

It is also exactly `stale_input_s`, this system's existing answer to "how long
may we hold without re-deciding".

A floor cycle runs the **full** law on the same inputs, integrator included. In
P1 mode the freshness question is already answered upstream and answered harder
- past `meter_max_age_s` the reading is None, which is "inputs missing", which
is a `P_SAFETY` 0 W claim - so a floor cycle there can only ever run on a reading
this system already calls fresh. ⚠️ That is transport-conditional: with
`meter_source: off`, `read_inputs()` takes the grid figure straight from
`hass.number()` with **no age gate at all**, so a Home Assistant entity that has
gone quiet reads as current on that path, with or without this change. Running
the full law is still right there - the alternative is a degraded second law
reasoning about an age nobody measured - but the floor does not make that path
fresh. The integrator advances per *cycle*, not per second,
so bounding the gap bounds the accumulation; `integrator_max_w` bounds it after
that. A gentler second law for floor cycles would be a second set of dynamics
nobody can test on this hardware.

⚠️ **This is not a freeze detector**, and TEL-05 established one cannot be built
here: at the converged -10 W target a 1 Wh register needs ~6 minutes to move
while the power figure legitimately repeats, so any threshold on `unchanged_s`
false-trips at the exact operating point we run at. Nothing in this change
decides whether a repeated reading is *true*. It stops "the number is the same"
from meaning "do not think". The consequence, stated plainly: against a meter
genuinely frozen at a wrong value the loop will now correct toward that wrong
value rather than hold a stale command, bounded by `max_w`, the arbiter and the
device rating. That was already what happened to a lying meter whose value
merely jittered; bit-exact repetition was never a safety property.

`DOCS.md`'s "Saturation is counted in cycles, not seconds" section documented
the old behaviour as an accepted deviation. It is **qualified, not deleted**:
the deviation survives (the window is still counted in cycles, and 3 x 15 s is
still not the specification's 10 s) but it is now bounded at 45 s instead of
open-ended, and the passage's second justification - "nothing accumulates while
it is stalled" - is called out as having covered integrator windup and not
staleness, which was the actual hazard. The same wrong argument in
`control.py`'s saturation comment is corrected too.

**TEL-04.** `meter_source: ha_signed` reads a single *signed* power entity, which
is the shape a HomeWizard P1 publishes. Both transports 0.3.0 shipped require two
unsigned registers and refuse negatives, so neither fits the meter installed here.

**TEL-05.** `meter_source: homewizard_local` polls the meter's own local API
directly. This exists because of something the HA paths cannot do: Home Assistant
fires no state change for a repeated reading and exposes no arrival signal at
all, so `sensor.p1_sample_age_s` on those transports measures time since the
*value changed*, not since the meter *reported*. Real captured data has gaps of
42 s and 97 s between changes on a perfectly healthy meter - long enough for a
30 s staleness threshold to command 0 W for no reason. Every HTTP response to the
local API is a genuine arrival, so the age means what it says.

Verified on the rig against a literally constant 350.0 W for 70 s - the exact
state where the HA path recorded zero samples in 45 s - with 14 arrivals and the
age never exceeding the poll cadence.

Set `p1_host` to the meter's address and `meter_poll_s` (default 5 s, matching
the meter's own ~5 s update). Still defaults to `off`; nothing changes for an
install that does not opt in.

Freeze detection is separate and deliberately not thresholded: a meter answering
200 OK with a stale number *is* arriving, so no arrival stamp can flag it. The
energy registers carry that as `unchanged_s`, surfaced on the status line for a
human to read rather than alarmed on - at a converged -10 W a 1 Wh register needs
about six minutes to move, so any threshold would false-trip at the target
operating point.

## Unreleased

**TEL-05.** A fourth `meter_source`, `homewizard_local`, which polls a
HomeWizard P1's **own local API** (`GET /api/v1/data`) instead of watching a
Home Assistant entity. Set `p1_host` to the meter's address; `meter_poll_s`
(default 5 s, the meter's own update rate) sets the cadence.

✅ **This is the first transport whose `sensor.p1_sample_age_s` measures when
the meter *reported*, and therefore the first one a firmware watchdog may
threshold.** Every HTTP response is an arrival: the meter answered, now, with
its current reading, and whether the *number* moved is not consulted. Home
Assistant cannot express that at all — a repeated reading emits no
`state_changed`, advances `last_reported` on neither serialiser, and
`state_reported` is not subscribable ("Event filter is required"). On
`ha_signed` that made a healthy meter under a flat load indistinguishable from
a dead one, and our own capture of this house's meter goes 42.2 s and 97.0 s
between changes — both past the default `meter_max_age_s` of 30, i.e. a false
trip to 0 W on a meter that is fine. If you have a HomeWizard P1, move to this
mode.

Everything TEL-01 established is reused, not re-implemented: ingest
timestamping, `meter_max_age_s`, the clock-recomputed age, the plausibility
bounds and the §20 unsigned-decode rejection, and the same `split_signed` sign
convention `ha_signed` uses. A failed or timed-out poll submits nothing, so it
is a *missing* reading — never 0 W — and it does not reset the age.

Verified on the ENV-01 rig against `sim/hwsim.py`, steady and with `--fault
freeze` injected. Still defaults to `off`.

⚠️ **An arrival stamp still cannot see a *frozen* meter**, and no arrival
detector can: a meter answering `200 OK` forever with a stale number is
arriving. The local API does expose what the HA path never had — the
`total_power_*_kwh` registers stop advancing — and the transport tracks it as
`unchanged_s`, but that is deliberately **not** folded into the age and not
thresholded: this controller regulates grid power toward ~0 W, and at a
converged −10 W the export register needs six minutes to move by its 1 Wh
resolution while the power figure legitimately repeats. Thresholding it at 30 s
would rebuild the false-trip limit cycle at the exact operating point we aim
for. Freeze detection needs the low-power case solved first, separately. It is
shown on the status page's P1 line instead — leaving it unthresholded only
holds up if a human can read it, so now they can.

**TEL-04.** A third `meter_source`, `ha_signed`, reading **one signed** Home
Assistant entity: positive = import, negative = export. That is the shape a
HomeWizard P1 publishes (`sensor.p1_meter_active_power`), and it is the meter
actually fitted here - which neither TEL-01 transport can read, because
`ha_dsmr` needs two unsigned registers and refuses a negative one, i.e. every
exporting telegram. Set `p1_net_entity`, and `p1_phase_net_entities` for the
per-phase capacity-tariff figures on a three-phase connection.

Everything TEL-01 established is inherited rather than re-implemented - the
new transport is a subclass of the `ha_dsmr` one overriding only which
entities it wants and how they become a sample. So ingest timestamping,
`meter_max_age_s`, `sensor.p1_sample_age_s` recomputed against the clock and
republished once a second, the plausibility bounds, and `unavailable` /
`unknown` treated as a *missing reading and never 0 W* all behave identically
across the three sources.

Still defaults to `off`; an existing install is unaffected until it opts in.

⚠️ **`sensor.p1_sample_age_s` is published on `ha_signed`, but must not yet be
thresholded by the ESP32 stale-input watchdog.** On the HA WebSocket paths the
age is stamped from `state_changed`, so it measures time since the value
*changed*, not since the meter *reported* - and Home Assistant exposes no
arrival signal for a repeated reading (no `state_changed`, no `last_reported`
movement on either serialiser, and `state_reported` is not subscribable over
the WebSocket). Measured on the ENV-01 rig against the real HomeWizard
integration. `ha_dsmr` mostly escapes it because a telegram moves several
entities at once; `ha_signed` has one, so a healthy meter under a flat load is
indistinguishable from a dead one. Our own capture has the house meter going
42.2 s and 97.0 s between changes. Raising `meter_max_age_s` does not fix that,
it only chooses which error you get; the fix is an arrival stamp from the meter
itself and is a separate ticket. Full detail in DOCS.md.

## 0.3.0

**SAFETY-04.** The control law's integrator is now an explicit accumulator,
bounded independently of the output clamp instead of inheriting whatever
headroom the clamp happened to leave. It also freezes while the inverter is
not tracking, rather than continuing to wind up against a command nothing is
acting on. `integrator_max_w` (default `0`) governs the bound; `0` means
"follow `max_w`", which is the existing behaviour.

Behaviour is unchanged at the defaults - a 4,928-case equivalence sweep
against the previous control law confirms it decides identically at
`integrator_max_w: 0`.

**TEL-01.** P1 meter ingestion, so a Belgian P1's two unsigned registers
(consumption, injection) no longer need a hand-written signed template
sensor: the subtraction moves into the add-on, done once and tested. Two
transports, chosen with the new `meter_source` option: `ha_dsmr` subscribes
to the DSMR integration over the HA WebSocket, `mqtt_p1` reads a topic.
Defaults to `off`, which keeps the existing `meter_entity` path untouched -
nothing changes for an install that does not opt in.

Enabling it publishes `sensor.p1_sample_age_s`: seconds since the newest
accepted telegram, recomputed against the clock and republished roughly once
a second rather than only when a telegram lands. That is deliberate - Home
Assistant only pushes a state on change, so a meter sitting at a genuinely
constant reading would otherwise look identical to a dead one. Watching the
age instead means a frozen meter shows a climbing age, not a flat line. The
firmware watchdog subscribes to this exact entity id.

Known limits, both already in DOCS.md: on `mqtt_p1`, a bridge stuck
republishing its last telegram still "arrives", so the age cannot detect
that particular failure - prefer `ha_dsmr` where both are available. And a
dead P1 meter takes 45 s to reach 0 W commanded (30 s for `meter_max_age_s`
to call the reading stale, then 15 s of `stale_input_s` on top), which is
`meter_max_age_s` and `stale_input_s` stacking, not either one alone.

## 0.2.1

`target_grid_w` (default -10 W): what the meter should rest at. The deadband
holds any resting point inside it indefinitely, and the meter bills import and
export on separate registers, so resting at +14 W import costs 0.34 kWh/day
with the loop behaving perfectly. Biasing the target slightly negative moves
that residue onto the export register. Configurable in the Configuration tab;
see DOCS.md for the trade-off table.

Behaviour is unchanged at `target_grid_w: 0`.

## 0.2.0

Precedence between strategies is now a first-class object instead of an if/else
ladder, ahead of there being more than three of them.

Every strategy returns a CLAIM each cycle - `set` ("I want X") or `limit` ("the
result must stay within these bounds") - and `arbiter.py` resolves them by one
rule:

  1. Highest-priority `set` wins; no claim at all means 0 W.
  2. Then every `limit` whose priority is >= that set's priority applies, most
     restrictive first.
  3. Contradictory limits are a BUG: command 0 W and say so.

Clause 2 is why "money outranks maintenance" is now a consequence of the
priorities rather than a special case in a Jinja template: the maintenance
charge-only limit binds the loop, but will not bind a higher-priority peak
shaving claim when one exists.

- Maintenance shaping (charge-only, cheap-window floor) moved out of the control
  law. `control.py` is once again only a controller that tracks the meter.
- The loop computes from the ARBITER's last output, not its own last wish. If
  something outranked it, that is what the hardware actually did, and tracking
  anything else makes it jump when it regains control.
- Every decision is explainable: "loop -> 0 W, limited by maintenance
  (charge-only)" now appears in the UI and the log, instead of a bare number.
- Safety limits (device rating, supervised max_w) bind every strategy including
  the highest, and are still enforced a second time at the point of writing.

## 0.1.6

Findings from installing this on a live system, replacing a working YAML
implementation. Every one of these was silent - the add-on looked healthy while
being completely unable to do its job.

- **`run.sh` must use `#!/usr/bin/with-contenv sh`.** The HA base images run
  s6-overlay, which starts services with a SANITISED environment. With a plain
  shebang, SUPERVISOR_TOKEN is simply absent and every Home Assistant call
  returns 401 - while `homeassistant_api: true` makes the permissions look
  correctly granted. Startup now logs the token length and probes the Core API,
  so the next person sees it in one line.
- **Bump `version:` for every change.** Supervisor keys the built image by
  version, so editing source and rebuilding silently reuses the old image. Two
  fixes appeared not to work because of this.
- **Dependencies come from apk, not pip.** Alpine is musl and there are no musl
  wheels for aiohttp; pip would compile it on the client's Pi.
- **paho-mqtt 1.x and 2.x are both supported.** Alpine ships 1.x, which has no
  `CallbackAPIVersion`; that raised and took the whole add-on down with it.
- **MQTT can no longer take down control.** Publisher construction is wrapped -
  observability must never stop the controller.
- **MQTT discovery is published from `on_connect`.** paho drops QoS-0 publishes
  issued before the CONNACK, so announcing straight after `connect()` published
  nothing at all while logging "MQTT connected".
- **Repeated failures log at most once a minute.** The control loop retries every
  second; unthrottled warnings rolled the log buffer and destroyed the startup
  diagnostics needed to debug the 401 above.
- **`auto_start` works.** The store's defaults supplied `auto: False`, so the
  fallback to the option could never fire.

Known issue: after deleting the MQTT entities from the registry during
development, Home Assistant would not re-adopt them from retained discovery -
not even after clearing the retained topics and reconnecting. The add-on
publishes correct discovery and live state (verified on the broker); this is an
HA-side adoption problem and affects status entities only, never control.

## 0.1.0

First packaged release. Ports the control loop and the monthly maintenance
cycle from the reference Home Assistant implementation into an add-on.

- Grid-following control: gain/slew/clamp/deadband with anti-windup, all tuned
  against measured hardware behaviour (see FIELD-GUIDE.md §14).
- Saturation freeze **with the duration term** — three consecutive diverging
  cycles, not one. The instantaneous test fires on every large correction,
  because the plant itself needs 3-6 s to settle.
- Monthly maintenance cycle as an ownership state machine: drain / charge /
  hold, with exactly one writer of the setpoint at any moment.
- Capacity-tariff awareness: the maintenance charge is capped by quarter-hour
  peak headroom, and peak shaving outranks the maintenance schedule.
- Failsafe behaviour: commands 0 W on missing inputs, on stop, and on shutdown.
  Never replays a stale setpoint - the reference implementation did, and the
  hardware watchdog cannot catch that.
- Ingress UI with a commissioning checklist that names problems in words.
- Optional MQTT discovery for status entities.

Known limits:

- Home Assistant OS / Supervised only (add-ons cannot run on Container/Core).
- The inverter protocol is reverse-engineered; no vendor contract.
- Without the optional RS485 e-stop, nothing covers the host machine dying.