Merge release/1.0 (TARIFF-03) into SAFETY-07

The anticipated conflict again, and only it: both cards prepend to CHANGELOG.md's
Unreleased section. Both entries kept, SAFETY-07 first. app/main.py and DOCS.md
auto-merged, so the overlap claim on #18 still holds - config.yaml and
validate.py remain untouched by this card even after TARIFF-02 and TARIFF-03
widened both.

Full suite green on the merge result, test_price.py and test_static_tariff.py
included.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015VXjA3LWccN1rJsk3okyPC

goodwe_controller/CHANGELOG.md

@@ -2,6 +2,73 @@
 
 ## Unreleased
 
+**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

goodwe_controller/DOCS.md

@@ -439,6 +439,11 @@ maintenance schedule.
 
 #### Capacity-tariff exposure of the grid-following loop
 
+⚠️ **This section described the exposure BEFORE SAFETY-07. The loop is now
+capped — see "The capacity-peak guard" below — and what follows is kept because it
+is still the clearest statement of WHAT is being guarded and why the number is
+worth what it is.**
+
 ⚠️ **`peak_cap_w` caps MAINTENANCE charging. It does not cap the loop.** The
 grid-following loop is bounded by `max_w` and by the device rating, and by
 nothing else. That is deliberate — the loop exists to *reduce* what crosses the
@@ -463,16 +468,72 @@ What exists today, and what does not:
 |---|---|
 | `max_w` bounds the magnitude | ✅ — but the exposure IS the magnitude |
 | Maintenance charging capped under `peak_cap_w` | ✅ |
-| **Loop capped under `peak_cap_w`** | ❌ — not built. This is SAFETY-07 |
+| **Loop capped under `peak_cap_w`** | ✅ — SAFETY-07, on every `meter_source` but `off` |
 | Frozen meter detected | ✅ on `homewizard_local` only, **alarm only** |
-| Frozen meter *stopped* | ❌ — the alarm does not command anything |
+| Frozen meter *stopped* | ✅ — SAFETY-07 forbids importing while the meter is frozen |
 | SoC guard on the loop | ❌ — the BMS is the only energy backstop |
 
-So the frozen-meter alarm below **narrows the window; it does not close it.** On
-a capacity-tariff site where an unattended runaway is unacceptable, the mitigation
-today is operational: set `max_w` no higher than the loop actually needs (the
-tuning was measured at 2000 W and a 1.78 kW load step is tracked fine there), fit
-the e-stop, and watch for the alarm.
+On the legacy `meter_source: off` path there are no P1 telegrams, so there is no
+quarter-hour average to guard, and the status page says **NOT COVERED**. There
+the mitigation is still operational: set `max_w` no higher than the loop actually
+needs (the tuning was measured at 2000 W and a 1.78 kW load step is tracked fine
+there), fit the e-stop, and watch for the alarm.
+
+### The capacity-peak guard
+
+**What it does.** Every cycle the add-on projects the open quarter to its
+end-of-interval average — from P1 telegrams only, never from the command — and
+compares it to a ceiling. Over the ceiling it files one claim: a **lower bound on
+the setpoint**, at `P_PEAK`. That single shape is both behaviours §10.1 asks for:
+a hard clamp on charging while the bound sits above the standing charge, and
+pre-emptive discharge while it sits above zero. It can never cap discharge or
+export, because a floor cannot.
+
+**The number.** `excess = 900 × (projected − ceiling) ÷ seconds remaining`, and
+the floor is the **measured** battery power plus that. The 900/remaining factor
+is the design rather than an accident: it is 1.0 at the top of the quarter and
+grows, so the guard is gentlest when its projection rests on the least evidence
+and becomes aggressive only once the quarter is nearly measured.
+
+**The ceiling** is `peak_cap_w`, raised to the month's banked peak when that
+already stands higher — once a month's bill is set at 5013 W, holding later
+quarters to 3500 W buys nothing and costs real charging. **That relaxation is
+granted only by an OBSERVED reading:** a peak store that cannot be read leaves the
+ceiling at the cap, so a broken instrument makes the guard stricter, never looser.
+
+**When it cannot see, it forbids importing.** Three states are not evidence and
+are checked before the projection is graded: no open quarter (a fresh averager
+projects a confident 0 W, which is under every ceiling there is), a meter gap
+longer than `meter_max_age_s`, and SAFETY-18's frozen-meter conjunction. In all
+three the claim becomes a floor of **0 W**. That is the backstop for the frozen
+meter: the alarm arrives at `meter_freeze_s` (300 s) of a 900 s billing quarter
+and commands nothing, and this stops the import.
+
+**It trips, where the freeze detector only notifies,** and the captures are why.
+On the healthy reference day (2026-08-23) it engages on **zero** of 95 quarters.
+The action is level-triggered and unlatched — re-derived from a fresh projection
+every cycle, released the instant the projection drops back — so a false
+engagement costs one cycle of discharge rather than a parked battery. And it is a
+LIMIT, so every stop in the watchdog table still outranks it.
+
+**Replayed on 2026-08-20**, the day the vendor controller failed: the billed
+quarter measured **5029.8 W**, above the inverter's own 5000 W rating, and the
+guard brings it to **3485 W** — under the soft limit, without ever asking for more
+than the inverter can give. Across both committed days the quarters under the
+limit go from 189/190 to 190/190. Two days is not thirty and does not close AC 3;
+it is the comparison two days can support.
+
+**What it alarms on is saturation, not engagement.** An engagement is an ordinary
+control action and goes to the action log and the status page. A *saturated*
+guard — asking for everything the inverter has while the quarter is still
+projected over — raises one notification (`goodwe_peak_saturated`), because that
+is the one state where the money will be spent and shedding load is the only
+thing left that changes the bill.
+
+**No new option.** The ceiling is `peak_cap_w`, which already shipped. 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 — every one of them a key that already exists.
 
 ### The winding-loop alarm (the output clamp)
 

goodwe_controller/app/main.py

@@ -33,17 +33,20 @@ from datetime import datetime, timezone
 import aiohttp
 
 from .arbiter import (
-    P_LOOP, P_MAINT_SHAPE, P_MAINTENANCE, P_SAFETY, Claim, resolve,
+    P_LOOP, P_MAINT_SHAPE, P_MAINTENANCE, P_PEAK, P_SAFETY, Claim, resolve,
 )
 from .control import Tuning, compute, maintenance_charge_floor, peak_at_risk
 from .clamp import (
     REASON_CODE as CLAMP_REASON, ClampResult, ClampWatchdog, clamp_to_rating,
     is_commandable,
 )
-from .freeze import FreezeWatchdog, FreezeWatchdogConfig, rail_w
+from .freeze import (
+    FreezeWatchdog, FreezeWatchdogConfig, looks_frozen, rail_w,
+)
 from .hass import HomeAssistant
 from .maintenance import IDLE, MaintConfig, Maintenance
 from .mqtt import MqttPublisher
+from . import peakguard
 from .p1 import P1Ingest, build_source, is_enabled
 from .static_tariff import static_status
 from .tariff import PeakStore, PeakTracker
@@ -189,6 +192,22 @@ class Controller:
                                 cap_w=float(opts.get("peak_cap_w", 3500)))
         self.peak.restore(self.p1.averager)
         _LOG.info("capacity-tariff peak store: %s", self.peak.restore_state)
+
+        # SAFETY-07. The guard TARIFF-01 was built for. It is stateless by
+        # design (peakguard.py holds no history between calls), so the only
+        # state here is what an EDGE needs: what the last cycle concluded, so
+        # the action log and the notification fire on transitions rather than
+        # once per cycle for a quarter of an hour.
+        #
+        # ⚠️ NO CONFIG OF ITS OWN, for the reason the clamp watchdog gives ~90
+        # lines up and TARIFF-01 repeats: the ceiling is `peak_cap_w`, already
+        # shipped and already validated, and every other term is read off
+        # something the arbiter or the averager already owns. CFG-03 still
+        # tracks three places, not four.
+        self.peak_guard = None
+        self._peak_engaged = False
+        self._peak_saturated = False
+        self._maint_shape_suspended = False
         # Set by amain() once the transport is built, so the status page can show
         # what only the transport knows (homewizard_local's unchanged_s). Stays
         # None when P1 is off, or under a transport that has no such counter.
@@ -424,6 +443,77 @@ class Controller:
                 self.log_event("inputs missing - commanding 0 W")
             claims.append(Claim.set("safety", P_SAFETY, 0.0, "inputs missing"))
 
+        # --- SAFETY-07: the quarter-peak projection guard --------------------
+        # ONE `lo` limit at P_PEAK, and peakguard.py carries the whole argument.
+        # What is decided HERE and nowhere else is what the guard is allowed to
+        # see and what it is allowed to ask for:
+        #
+        # ⚠️ `frozen` IS EVALUATED ON THIS CYCLE'S STANDING COMMAND, NOT ON
+        # `self.freeze_wd.active`. The watchdog object is advanced at the END of
+        # cycle() (check_meter_frozen), so reading its level here would grade
+        # this quarter against the PREVIOUS cycle's conjunction - one cycle of
+        # lag, in a guard whose entire job is to act inside a 900 s quarter.
+        # `standing` is the value SAFETY-02 already computed above for exactly
+        # this meaning: the command on the wire that this meter reading is a
+        # response to. The predicate itself is freeze.py's shipped
+        # `looks_frozen()`, not a copy of it - see loop_is_following() on what a
+        # transcribed safety predicate costs.
+        #
+        # ⚠️ `achievable_w` BOUNDS THE FLOOR SO THIS CLAIM CANNOT CONTRADICT THE
+        # SAFETY LIMITS ABOVE. `lo > hi` in the arbiter is a CONTRADICTION and
+        # resolves to 0 W - the battery parked, in the one minute of the month it
+        # should be discharging hardest. maintenance_charge_w had to learn the
+        # same lesson (config.yaml, validate.py rule 6); this claim is built so
+        # the config cannot express it.
+        #
+        # ⚠️ AND IT IS FILED EVERY CYCLE, unconditionally, like SAFETY-02's. A
+        # claim is not sticky. Filed once and skipped next cycle, the floor
+        # lapses and the actuator goes silently back to whatever was pushing the
+        # peak. Which of the higher-priority claims outranks it is arbiter.py's
+        # business, not this block's - and the answer is already right: a
+        # maintenance DRAIN or HOLD owns the setpoint at P_MAINTENANCE and
+        # commands discharge or zero, neither of which can create offtake, while
+        # a maintenance CHARGE is only ever a P_MAINT_SHAPE limit (40) and is
+        # therefore bound by this floor. That is §9.1's "money outranks the
+        # maintenance schedule", now as a consequence of the priorities.
+        #
+        # ⚠️ AND IT IS NOT RUN AT ALL WITHOUT P1. `meter_source: off` is the
+        # legacy single-entity path: there is no telegram stream, so
+        # `p1.averager` has no open block, so the guard would be BLIND on every
+        # cycle forever - and blind means "no charging". That is not a
+        # conservative default, it is a controller that silently stops charging
+        # on every install that has not opted in to P1 yet; test_clamp.py's
+        # export case caught it doing exactly that. The billed quantity simply
+        # is not measured on that path, and the honest report is NOT COVERED -
+        # the same answer, for the same reason, that the status page already
+        # gives for SAFETY-18 on a transport that cannot see `unchanged_s`.
+        dev_top = self.dev_max if self.dev_max is not None else max_w
+        self.peak_guard = None if not self.p1_enabled else peakguard.decide(
+            self.peak.project(self.p1.averager, now),
+            batt_w=self.batt,
+            cap_w=float(self.o.get("peak_cap_w", 3500)),
+            reading=self.peak.peak(now),
+            achievable_w=min(max_w, abs(float(dev_top))),
+            max_hold_s=self.p1.averager.max_hold_s,
+            frozen=looks_frozen(getattr(self.p1_source, "unchanged_s", None),
+                                standing, max_w, self.freeze_wd.cfg),
+        )
+        peak_claim = (peakguard.claim(self.peak_guard)
+                      if self.peak_guard is not None else None)
+        if peak_claim is not None:
+            claims.append(peak_claim)
+        # ⚠️ EDGE-LOGGED, not level-logged. The action log holds 40 lines; a
+        # guard that appended one every 15 s for a whole quarter would push the
+        # cause of the event off the end of the log that is meant to explain it.
+        # SAFETY-03 logs every engagement because a clamp engagement IS the
+        # count that matters there; here the count is the seconds, and the
+        # engagement's own reason line already carries them.
+        engaged_now = self.peak_guard is not None and self.peak_guard.engaged
+        if engaged_now != self._peak_engaged:
+            self._peak_engaged = engaged_now
+            if self.peak_guard is not None:
+                self.log_event(self.peak_guard.explain())
+
         # --- maintenance -----------------------------------------------------
         if result.owns_setpoint:
             claims.append(Claim.set("maintenance", P_MAINTENANCE,
@@ -435,9 +525,44 @@ class Controller:
             # claimed, so the loop can discharge and shave the peak. When a
             # dedicated peak-shaving strategy arrives it will claim SET at
             # P_PEAK and outrank this limit without any code here changing.
-            if peak_at_risk(self.peak_fc, float(self.o.get("peak_cap_w", 3500))):
-                self.log_event("peak at risk - maintenance charge shaping suspended")
+            # ⚠️ AND SAFETY-07 IS THE OTHER HALF OF THIS CONDITION, NOT A
+            # SECOND MECHANISM. The shaping is `hi <= 0`; the peak guard's claim
+            # is `lo >= 0`. Both are limits, arbiter.py applies EVERY limit at or
+            # above the winner's priority, and most restrictive wins - so a peak
+            # event during a maintenance charge window files `lo > hi`, which is
+            # a CONTRADICTION and resolves to 0 W. The battery would park in the
+            # exact quarter it should be shaving. Priority does not save it:
+            # clause 2 has no notion of a higher-priority limit overriding a
+            # lower one, and it should not - "most restrictive" is what makes the
+            # safety bounds unconditional.
+            #
+            # So the higher-priority claimant SUPPRESSES the lower one rather
+            # than out-voting it, which is what this branch already did for
+            # `peak_forecast_entity`. That entity is optional and empty on a
+            # default install, so until now the suspension could not fire on the
+            # site that needed it. `self.peak_guard` is measured from the meter
+            # and is always there.
+            #
+            # ⚠️ THE CONJUNCTION LIVES IN peakguard.py, NOT HERE, and that is not
+            # tidiness. It was inline, and a reviewer mutated it back to its
+            # historical broken form - re-gated on `peak_forecast_entity` alone -
+            # and every suite in this repository stayed green while a default
+            # install parked at 0 W in the quarter that sets the month's bill.
+            # Read shaping_suspended() before touching this: both terms are
+            # there, and why neither is sufficient alone.
+            if peakguard.shaping_suspended(
+                    self.peak_guard, self.peak_fc,
+                    float(self.o.get("peak_cap_w", 3500))):
+                # Edge-logged: level-logging here would write one line per cycle
+                # for as long as the event lasts and flush the 40-line action log
+                # of the very context that explains it.
+                if not self._maint_shape_suspended:
+                    self._maint_shape_suspended = True
+                    self.log_event("peak at risk - maintenance charge shaping "
+                                   "suspended so the peak claim is not "
+                                   "contradicted")
             else:
+                self._maint_shape_suspended = False
                 hi = 0.0
                 reason = "charge-only"
                 if self.cheap:
@@ -552,6 +677,11 @@ class Controller:
         # is not, and test_freeze.py part 5 fails on exactly that.
         await self.check_meter_frozen()
 
+        # --- SAFETY-07's alarm, and it is NOT the engagement ------------------
+        # Its boundary is unconstrained: it grades `self.peak_guard`, which this
+        # method assigned above and nothing since has touched.
+        await self.check_peak_saturated()
+
         # --- TARIFF-01, and it is LAST on purpose -----------------------------
         # Banking a quarter is bookkeeping about a moment that has already
         # happened; a slow SD card must never sit between the meter and the
@@ -643,6 +773,59 @@ class Controller:
                 f"for was itself out of range."),
         })
 
+    async def check_peak_saturated(self) -> None:
+        """SAFETY-07. Alarm ONLY when the guard has run out of inverter.
+
+        ⚠️ AN ENGAGEMENT IS NOT AN ALARM, AND THE DIFFERENCE IS §10.3's DOCTRINE
+        READ CAREFULLY. "Treat an out-of-range computed setpoint as a bug signal
+        - clamp it, then alarm" is SAFETY-03's rule and it is right there: the
+        loop computing 14 768 W is a WINDING INTEGRATOR, a defect, and clipping
+        it silently is what let it keep climbing behind a dashboard full of legal
+        numbers. A peak floor above the inverter's rating is not a defect. It is
+        a house load bigger than the inverter - 2026-08-20 15:45 ran ~5900 W
+        against a 5000 W unit - and alarming on the ordinary case is how an
+        operator learns to dismiss the notification that matters. So an
+        engagement is an ordinary control action: it goes to the §15 action log
+        and the status page, and no further.
+
+        What DOES earn the operator's attention is saturation: the guard asked
+        for more discharge than the inverter can give, so the quarter is going
+        over the ceiling anyway and the controller has already done everything
+        available to it. That is the one state where the money is going to be
+        spent and a human is the only remaining actuator.
+
+        Edge-triggered, and de-asserting arms it again - the same shape as
+        FreezeWatchdog's `evaluate()`, and for the same reason: one notification
+        per episode, not one per cycle.
+        """
+        g = self.peak_guard
+        saturated = bool(g is not None and g.saturated)
+        was, self._peak_saturated = self._peak_saturated, saturated
+        if not saturated or was:
+            return
+        self.log_event("PEAK GUARD SATURATED - " + g.reason)
+        _LOG.error(
+            "SAFETY-07 - %s. The guard is asking for everything the inverter "
+            "has and the quarter is still projected over the ceiling: this "
+            "quarter will set or raise the month's billed capacity peak. The "
+            "controller is STILL RUNNING and still shaving; there is nothing "
+            "further it can do. Shed load.", g.reason)
+        # A fixed notification_id, and a DIFFERENT one from every other alarm
+        # here - check_meter_frozen() gives the reasoning. A saturated peak guard
+        # is not a frozen meter and is not an out-of-range setpoint.
+        await self.hass.call("persistent_notification", "create", {
+            "notification_id": "goodwe_peak_saturated",
+            "title": "GoodWe: the capacity-peak guard has run out of inverter",
+            "message": (
+                f"{g.reason}.\n\nThe battery is being commanded to everything "
+                f"it can give and the quarter is STILL projected over the "
+                f"{g.ceiling_w:.0f} W ceiling, so this quarter will set or raise "
+                f"the month's billed capaciteitstarief peak. This is not a "
+                f"controller fault - the load is simply larger than the "
+                f"inverter. Shedding load in the next few minutes is the only "
+                f"thing that changes the bill."),
+        })
+
     async def check_meter_frozen(self) -> None:
         """SAFETY-18. Alarm on a meter that answers forever without changing.
 
@@ -918,6 +1101,40 @@ class Controller:
                                       f"{self.peak.cap_w:.0f} W soft limit"
                                     + f"; {self.peak.recorded} quarter(s) banked"
                                     + f"; on start: {self.peak.restore_state}"})
+
+            # SAFETY-07. Three rows, not one, because "idle" and "cannot see"
+            # are opposite states and a page that spelled them the same way
+            # would show a green line over a guard that is switched off. `None`
+            # is a fourth state on purpose: before the first cycle the guard has
+            # concluded nothing, and saying "idle" then would be a claim about a
+            # decision that has not been taken.
+            _pg = self.peak_guard
+            if _pg is None:
+                out.append({"ok": True, "warn": True,
+                            "text": "peak guard: no decision yet - the first "
+                                    "control cycle has not run"})
+            elif _pg.saturated:
+                out.append({"ok": False, "warn": False,
+                            "text": "PEAK GUARD SATURATED - " + _pg.reason})
+            elif _pg.blind:
+                out.append({"ok": False, "warn": True, "text": _pg.explain()})
+            else:
+                out.append({"ok": not _pg.engaged, "warn": _pg.engaged,
+                            "text": _pg.explain()})
+        else:
+            # ⚠️ OUTSIDE the `p1_enabled` block, and that placement is the point.
+            # SAFETY-07 is NOT COVERED on the legacy single-entity path - the
+            # quarter-hour average is measured from P1 telegrams and this path
+            # has none - and a row that only rendered where the guard runs would
+            # leave the one install that is unprotected showing nothing at all.
+            # Same answer, same reason, as the freeze detector's NOT COVERED row
+            # a few lines up.
+            out.append({"ok": False, "warn": True,
+                        "text": f"peak guard: NOT COVERED on meter_source "
+                                f"{o.get('meter_source')} - the quarter-hour "
+                                f"average is measured from P1 telegrams and this "
+                                f"path has none, so a runaway cannot be capped"})
+
         rows = [("battery SoC", self.soc, o.get("soc_entity")),
                 ("battery power", self.batt, o.get("batt_entity"))]
         if not self.p1_enabled:

goodwe_controller/app/p1.py

@@ -286,6 +286,26 @@ class QuarterAverager:
         """Offtake watt-seconds accumulated in the open block so far."""
         return self._acc
 
+    @property
+    def held_net_w(self) -> float:
+        """The signed net power currently being held forward, in W.
+
+        PUBLIC BECAUSE SAFETY-07 DEPENDS ON IT, AND THAT IS THE WHOLE REASON.
+        `PeakTracker.project()` extrapolates the rest of the quarter from this
+        value and used to reach it as `getattr(averager, "_last_net", 0.0)`. A
+        default on a private name fails QUIET and fails in the DISARMING
+        direction: rename `_last_net` and every projection silently credits 0 W
+        held, which reads LOW, which puts the peak guard under the ceiling
+        exactly when the meter says it is over. Filed as TEST-13 item 4 and
+        fixed here rather than inherited - a money guard may not rest on a read
+        that can go wrong without a symptom.
+
+        Positive is offtake, matching `P1Sample.net_w`. It is the raw held
+        value, NOT floored at zero: the flooring is the accumulator's rule (only
+        offtake is billed) and belongs where it is applied, not here.
+        """
+        return self._last_net
+
     @property
     def offtake_avg_w(self) -> float:
         """Average offtake over the part of the open block seen so far."""

goodwe_controller/app/peakguard.py

@@ -0,0 +1,517 @@
+"""SAFETY-07 - the quarter-peak projection guard, and the only thing in this
+add-on that is allowed to spend the loop's freedom to save money.
+
+WHAT IT IS. TARIFF-01 built the INSTRUMENT: `tariff.py` measures the number the
+Belgian capaciteitstarief bills on - the highest 15-minute average offtake of the
+calendar month - and banks it where a restart cannot reach. It files no claim and
+commands nothing, and says so in its own first paragraph. This module is the
+CONSUMER it named. It turns that measurement into one arbiter claim and nothing
+else: no writes, no I/O, no clock, no state between calls. Given a projection and
+two measured numbers it returns a decision; `main.py` files it.
+
+------------------------------------------------------------------------------
+THE HAZARD, WITH THE EURO FIGURE
+
+specs.md:583 records a measured case: ONE 4998 W quarter moved a monthly peak
+from 4832 W to 5013 W and billed the whole month at it. The capacity tariff has
+no averaging and no forgiveness - the month's single worst quarter is the bill.
+
+SAFETY-18 shipped the ALARM for the way this system produces such a quarter by
+itself: a meter that answers 200 OK forever with a document that never changes,
+which every freshness mechanism in this stack reads as healthy while the loop
+integrates a constant error to the rail and holds it there. freeze.py's own
+table: at the live install's `max_w: 5000`, a meter frozen at >= 3000 W rails the
+command in 75 s. SAFETY-18 notifies at `meter_freeze_s` (300 s default), against
+a 900 s billing quarter. That is about nine minutes of human reaction time and
+nothing else - freeze.py says so in the words "THIS CUT DOES NOT CAP THAT."
+
+This is the cap. It is deliberately NOT a second freeze detector: it consumes
+freeze.py's shipped predicate rather than transcribing it.
+
+------------------------------------------------------------------------------
+THE MECHANISM: ONE `lo` LIMIT AT `P_PEAK`, AND WHY THAT SHAPE
+
+specs.md 10.1 row 2 asks for exactly two behaviours - "pre-emptive discharge as
+the projected quarter average approaches the soft limit; hard clamp on charging
+above it" - and 9.1 constrains the lever: "Discharge and export are never capped
+by the peak guard; only charging and import are."
+
+A LOWER BOUND on the setpoint is both of those in one object, and it is the only
+shape that is:
+
+  * incapable of capping discharge - a floor never stops the loop discharging
+    HARDER, so 9.1's asymmetry is structural here rather than remembered;
+  * incapable of capping export, for the same reason;
+  * a hard clamp on charging whenever the floor is above the standing charge;
+  * pre-emptive discharge whenever the floor is above zero.
+
+It is a LIMIT, not a SET, and that matters at both ends of the priority table.
+`P_PEAK` (50) is above `P_LOOP` (10), so it binds the grid-following loop. It is
+BELOW `P_MAINTENANCE` (60) and `P_SAFETY` (100), so a maintenance phase that owns
+the actuator and every layer-1 stop still outrank it: arbiter.py's clause 2 only
+applies limits whose priority is >= the WINNING SET's, so a P_SAFETY 0 W stop is
+not dragged off zero by this floor. That is the correct order - a controller
+pushing power the wrong way is wrong however much money it is trying to save.
+
+⚠️ AND A LIMIT DOES NOT OUT-VOTE ANOTHER LIMIT. This was assumed, and it is
+false: arbiter.py clause 2 applies EVERY limit at or above the winner's priority
+and takes the MOST RESTRICTIVE, with no notion of priority between them - which
+is exactly what makes the safety bounds unconditional and must not be changed.
+So this floor (`lo >= 0`) and the maintenance charge-only shaping (`hi <= 0`) do
+not resolve in P_PEAK's favour when both are filed. They resolve to `lo > hi`, a
+CONTRADICTION, which arbiter.py answers with 0 W - the battery parked in the one
+quarter of the month it should be shaving hardest. The rule for two limits that
+can cross is therefore SUPPRESSION, not priority: main.py drops the maintenance
+shaping for as long as this guard is engaged, which is the same thing that branch
+already did for `peak_forecast_entity` and is why 9.1's "money outranks the
+maintenance schedule" reads as a suppression there rather than as a priority
+here. The floor's clamp to `achievable_w` closes the only other pair that could
+cross - the P_SAFETY rating bounds.
+
+AND IT IS WHY OPT-04 NEEDS NO CHANGE HERE. 9.1 requires peak shaving to outrank
+arbitrage. Arbitrage does not exist yet, but it will arrive as a SET below
+P_PEAK, and arbiter.py clause 2 will bind it with this floor on the day it lands,
+with no edit to this file. Expressing the priority as a claim rather than as an
+if/else in main.py is the whole reason that holds.
+
+------------------------------------------------------------------------------
+THE FLOOR IS DERIVED, AND EVERY TERM OF IT IS MEASURED
+
+Let `ceiling_w` be the quarter average that must not be exceeded and `remaining`
+the seconds left in the open block. `project()` gives
+
+    projected = (acc + held * remaining) / 900
+
+so the held level that would land the quarter exactly on the ceiling is
+`held + 900 * (ceiling - projected) / remaining`, and the watts that must come
+off the meter are the difference between the two:
+
+    excess_w = 900 * (projected - ceiling) / remaining
+
+That single expression carries the whole control law, and two properties of it
+are the argument for this design:
+
+  1. THE CLAIM IS FILED WHILE THE FLOOR IS POSITIVE, NOT WHILE THE EXCESS IS.
+     `excess` is what must come off the METER, so it falls to zero the moment the
+     battery is already taking it - and a guard that disengaged there would hand
+     the actuator back, watch the offtake spring straight to the raw load, and
+     engage again next cycle. That is not a hypothetical: it is what the first
+     cut of this file did to 2026-08-20's billed quarter on replay - seventeen
+     engage/release cycles in 900 s, and the quarter still finished 49 W over the
+     limit it had spent a quarter of an hour holding. 10.2 names the cost
+     exactly: "an EMS that oscillates is worse than no EMS, both for hardware
+     wear and - per 9.1 - for the peak it is trying to avoid."
+
+     `batt_w + excess` IS the quantity to threshold: the offtake there would be
+     with the battery at zero, less the level this quarter can still afford. Read
+     as a LEVEL rather than as a switch, the same arithmetic gives a claim that
+     is filed for as long as a positive floor is needed and decays continuously
+     to nothing as the load falls. So there is still no hysteresis constant in
+     this file - not because oscillation was ruled out by argument, but because
+     the thresholded quantity is the one that does not oscillate.
+
+  2. THE GAIN IS 900/remaining, WHICH IS 1.0 AT THE START OF THE QUARTER AND
+     GROWS. This is the answer to the degeneracy at the top of a block, where the
+     projection is an extrapolation from almost no evidence: the guard is at its
+     GENTLEST exactly when its evidence is weakest, and becomes aggressive only
+     once most of the quarter is behind it and the projection is nearly a
+     measurement. The formula supplies that; no minimum-evidence threshold was
+     invented, because inventing one would delay the guard in the case it exists
+     for - a sustained rail from second zero.
+
+The floor itself is then
+
+    floor_w = batt_w + excess_w
+
+`batt_w` IS THE MEASURED BATTERY POWER, NOT THE STANDING COMMAND. Both were
+available. The measured one is right for two independent reasons. First, 9.1's
+settled constraint is that the utility bills what FLOWED - the arithmetic
+`grid = load - batt` is true of delivered power and false of intent. Second, and
+concretely: when the inverter cannot follow (the saturation case control.py
+already freezes the integrator for), a floor built on the command asks for
+`command + excess` every cycle while the command never arrives, and winds itself
+to the rail; a floor built on the measurement asks for `delivered + excess`,
+which stops rising the moment delivery stops rising. The delivered form cannot
+wind up. The intent form can.
+
+------------------------------------------------------------------------------
+THE CEILING IS `peak_cap_w`, RAISED ONLY BY AN OBSERVATION
+
+11.2's row is "Quarter peak | Projected quarter average approaches THE MONTHLY
+PEAK | Clamp the command", and 9.1's soft limit is `peak_cap_w`. Both matter,
+because the money is spent on `max(month's peak so far, this quarter)`:
+
+    ceiling_w = max(peak_cap_w, the month's banked peak)
+
+Once a month's peak already stands at 5013 W, holding every later quarter to
+3500 W buys nothing - the bill for that month is already 5013 W - while the
+holding itself costs real charging the site wanted. Raising the ceiling to a peak
+that has already been paid for is what stops one bad quarter from parking the
+battery's charging for three weeks.
+
+AND THAT RELAXATION IS GRANTED ONLY BY AN `OBSERVED` READING. `NO_RECORD` and
+`NOT_OBSERVED` both leave the ceiling at `peak_cap_w`, the tighter of the two, so
+a peak store that cannot be read makes this guard STRICTER and never looser. That
+direction is the whole point: tariff.py's `PeakReading` exists because a database
+that did not answer must not become a number, and the number it would have become
+here - a zero peak - is exactly the one that reads as "no peak recorded, full
+freedom". It is also why this module does NOT consume `headroom_w()`, whose
+disarming failure it would otherwise inherit: a peak read as 0 W returns FULL
+headroom, and full headroom is a guard that is not there.
+
+------------------------------------------------------------------------------
+BLINDNESS COMES FIRST, AND IT IS A DIFFERENT ACTION
+
+A projection is evidence only if the quantity under it was actually measured.
+Three ways it is not, all of them checked BEFORE the projection is graded,
+because grading a stimulus where the quantity it grades is degenerate is how a
+guard passes its own tests while being switched off:
+
+  * NO OPEN BLOCK. `project()` answers a fresh averager with
+    `Projection(0.0, 0.0, 0.0, 0.0, False, None, 0.0)` - a ZERO projection, which
+    compares as comfortably under any ceiling. That is not "the quarter is fine",
+    it is "there is no quarter". Graded as a number it silently disarms the guard
+    on exactly the startup path where nothing else is watching.
+
+  * COVERAGE. A meter gap longer than one `max_hold_s` window leaves seconds that
+    `QuarterAverager` deliberately does not accumulate (p1.py: holding a dead
+    meter's last value forward is how a fabricated peak gets onto a permanent
+    record). Those seconds make the projection read LOW, which is the
+    under-guarding direction. The threshold is `max_hold_s` itself - the point
+    p1.py already stops trusting a held sample - not a new number.
+
+  * FROZEN. SAFETY-18's conjunction, evaluated through `freeze.looks_frozen()`
+    rather than re-written here. A frozen meter is the worst case of all: the
+    samples keep ARRIVING, so coverage stays perfect and the projection is
+    confident - and if the stuck value is an export, the projection is a
+    confident ZERO while the real meter runs at the rail. This is the case
+    SAFETY-07 exists to backstop, and it is the case a projection cannot see.
+
+THE BLIND ACTION IS `floor_w = 0.0`: the battery may not IMPORT while the billed
+quantity is unobservable. It is the minimal action that closes the hazard, and it
+is one-directional - forbidding charging can only ever lower delivered offtake,
+never raise it - so a false blindness costs a charge the site wanted and cannot
+cost energy the pack did not want to give. Contrast the graded branch, which can
+force discharge; blindness deliberately does not earn that, because there is by
+definition no measurement to size it from.
+
+------------------------------------------------------------------------------
+WHY THIS ONE TRIPS WHERE SAFETY-18 ONLY NOTIFIES
+
+SAFETY-18 shipped notification-first, and config.yaml says why in one line: "A
+false trip parks the battery; a false alarm costs an alert." That reasoning is
+right for a DETECTOR whose negative case had to be established on a rig run. It
+does not carry to this guard, and the two committed captures say so:
+
+  * MEASURED FALSE-ENGAGEMENT RATE ON THE HEALTHY DAY: zero. 2026-08-23's worst
+    quarter is a few hundred watts against a 3500 W ceiling and nothing in its 95
+    quarters comes near engaging. test_peakguard.py re-measures that every run
+    from the committed capture rather than quoting it here.
+  * THE ACTION IS LEVEL-TRIGGERED AND UNLATCHED. It is re-derived from a fresh
+    projection every cycle and releases the instant the projection drops back
+    under the ceiling. There is no state to clear and no operator to wake. A
+    false engagement costs one cycle of discharge, not a parked battery - which
+    is precisely the cost SAFETY-18 could not bound and this one can.
+  * THE ACTION IS A LIMIT, NOT A SET. Every safety stop in 11.2's table outranks
+    it, so the worst thing a bug in this file can do is fail to relax a floor -
+    and a floor is bounded above by `achievable_w`, which main.py takes from the
+    same `max_w` and device rating the arbiter already enforces.
+
+So it trips. It is also the only guard here whose trigger IS the billed quantity
+itself, measured from parsed P1 telegrams, rather than a proxy for it.
+
+------------------------------------------------------------------------------
+SATURATION IS REPORTED, NOT ALARMED AS A BUG
+
+10.3's doctrine - "treat an out-of-range computed setpoint as a bug signal, clamp
+it AND alarm" - is SAFETY-03's, and it is right there: a loop computing 14 768 W
+is a winding integrator, i.e. a defect. A peak floor above the inverter's rating
+is NOT a defect. It is the ordinary physics of a 5900 W house load on a 5000 W
+inverter, which is literally what 2026-08-20 15:45 was. So the floor is clamped
+to `achievable_w` and the decision carries `saturated`, and main.py raises ONE
+notification on that - not because the controller is broken, but because it is
+the one case where the money WILL be spent and the controller has already done
+everything it can. An engagement that is not saturated is an ordinary control
+action and goes to the action log, like the clamp's engagements do.
+
+------------------------------------------------------------------------------
+NO NEW OPTION. CFG-03 STILL TRACKS THREE PLACES.
+
+`peak_cap_w` already ships, is already validated `int(500,15000)`, and is already
+read by `control.peak_at_risk()`. `max_hold_s` comes off the averager the caller
+already owns; `achievable_w` off `max_w` and the device rating the arbiter
+already enforces; the freeze predicate off `meter_freeze_s`. config.yaml and
+validate.py are untouched by this ticket, exactly as TARIFF-01 left them, and
+14's `capacity_peak_limit_kw` stays declined - a kW key beside a W key is the
+mismatch trap 14's own warning box describes.
+
+------------------------------------------------------------------------------
+⚠️ WHAT THIS GUARD DOES NOT HAVE: AN ENERGY BACKSTOP. SEE SAFETY-24.
+
+The graded branch can FORCE DISCHARGE, and nothing in this file or in the loop
+below it consults the state of charge - `soc_floor` belongs to the maintenance
+cycle, and freeze.py already records 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. Against a pack sitting near 9 % that is not nothing, and the guard will
+keep asking as long as the meter says the quarter is over the ceiling.
+
+That is deliberate for THIS card and it is filed as SAFETY-24. It is also the
+reason this must not go to a client install before SAFETY-24 lands: on the
+reference site the BMS is a known quantity and an operator is present, and on a
+client site neither is true.
+
+------------------------------------------------------------------------------
+AND IT READS NO PER-PHASE QUANTITY, ON PURPOSE
+
+The sum/max/net reduction is an open product question with the owner. TARIFF-01
+put it behind `billed_w()` and defaulted to NET. Everything this module consumes
+- `PeakReading.peak_w`, `Projection.projected_w` - is already reduced by that one
+function, so whichever way the owner answers, this file does not change and does
+not need re-testing. Nothing here indexes a phase.
+"""
+
+from dataclasses import dataclass
+
+from .arbiter import P_PEAK, Claim
+from .control import peak_at_risk
+from .tariff import QUARTER_S
+
+# Why the projection may not be graded. Reasons, not booleans: the status page
+# and the action log both have to say WHICH, and "blind" alone is unactionable.
+BLIND_NO_BLOCK = "no open quarter - the meter has produced no sample yet"
+BLIND_COVERAGE = "meter gap"
+BLIND_FROZEN = "SAFETY-18: the meter is answering without changing"
+
+# Where the ceiling came from. Only the second is a relaxation, and only an
+# OBSERVED reading may produce it.
+CEILING_CAP = "peak_cap_w"
+CEILING_MONTH_PEAK = "the month's banked peak"
+
+SOURCE = "peak"
+
+
+@dataclass(frozen=True)
+class PeakDecision:
+    """What the guard concluded this cycle, with everything it concluded it from.
+
+    Every field is derived from a measurement taken in the same cycle. There is
+    no `floor_w_or(default)` and there must not be one, for `PeakReading`'s
+    reason: an accessor that turns "not measured" into a number re-creates the
+    defect the type exists to remove.
+    """
+
+    engaged: bool
+    floor_w: float | None            # the `lo` the claim carries; None = no claim
+    ceiling_w: float
+    ceiling_from: str
+    blind: str | None = None
+    projected_w: float | None = None
+    excess_w: float = 0.0
+    wanted_w: float | None = None    # the floor before `achievable_w` clamped it
+    saturated: bool = False
+    reason: str = ""
+
+    def explain(self) -> str:
+        if not self.engaged:
+            if self.projected_w is None:
+                return f"peak guard idle - {self.reason}"
+            return (f"peak guard idle - quarter projects "
+                    f"{self.projected_w:.0f} W under a {self.ceiling_w:.0f} W "
+                    f"ceiling ({self.ceiling_from})")
+        return self.reason
+
+
+def quarter_ceiling_w(cap_w, reading) -> tuple[float, str]:
+    """The average this quarter may not exceed, and where the number came from.
+
+    ONLY an OBSERVED reading may raise it. `NO_RECORD` (the store answered, this
+    month has nothing yet) and `NOT_OBSERVED` (the store did not answer) both
+    return the cap, which is the tighter answer. A failed read therefore makes
+    the guard stricter, never looser - see the module note.
+    """
+    cap = float(cap_w)
+    if (reading is not None and getattr(reading, "is_observed", False)
+            and reading.peak_w is not None and float(reading.peak_w) > cap):
+        return float(reading.peak_w), CEILING_MONTH_PEAK
+    return cap, CEILING_CAP
+
+
+def excess_w(projection, ceiling_w) -> float:
+    """Watts that must come off the meter NOW to land the quarter on the ceiling.
+
+    The whole control law, and the only place it is written down:
+
+        excess = 900 * (projected - ceiling) / remaining
+
+    Positive means the quarter is projected over; zero or negative means it is
+    not. `remaining <= 0` is not a large number, it is an undefined one - a
+    quarter with no seconds left cannot be influenced - so it answers 0.0 rather
+    than dividing.
+    """
+    remaining = QUARTER_S - float(projection.elapsed_s)
+    if remaining <= 0.0:
+        return 0.0
+    return QUARTER_S * (float(projection.projected_w) - float(ceiling_w)) / remaining
+
+
+def uncovered_s(projection) -> float:
+    """Seconds of the open block the meter did not actually cover.
+
+    Derived from what `Projection` already carries rather than re-measured:
+    `covered_frac` is `elapsed / elapsed_wall`, and `elapsed_s` IS that wall.
+    """
+    return max(0.0, float(projection.elapsed_s)
+               * (1.0 - float(projection.covered_frac)))
+
+
+def blindness(projection, max_hold_s, frozen: bool) -> str | None:
+    """Why this projection may not be graded, or None if it may be.
+
+    Evaluated BEFORE the projection is compared to anything. Each of the three is
+    a way for the projected number to be confident and wrong, and the first - a
+    fresh averager's all-zero `Projection` - is confident, wrong AND under every
+    ceiling there is.
+    """
+    if frozen:
+        return BLIND_FROZEN
+    if projection.block_start is None:
+        return BLIND_NO_BLOCK
+    gap = uncovered_s(projection)
+    if max_hold_s is not None and gap > float(max_hold_s):
+        return (f"{BLIND_COVERAGE}: {gap:.0f} s of this quarter uncovered, past "
+                f"the {float(max_hold_s):.0f} s the averager will hold a sample")
+    return None
+
+
+def decide(projection, *, batt_w, cap_w, reading, achievable_w,
+           max_hold_s=None, frozen: bool = False) -> PeakDecision:
+    """The whole guard. Pure: no clock, no I/O, no state carried between calls.
+
+    `batt_w`       MEASURED battery power, positive = discharge. None when the
+                   sensor for it is missing - and then there is nothing to size a
+                   floor from, so no floor is invented.
+    `achievable_w` the largest discharge the arbiter could actually grant
+                   (`min(max_w, device max)`). The floor is clamped to it so this
+                   claim can NEVER produce `lo > hi` in the resolver - a
+                   contradiction there resolves to 0 W, which would park the
+                   battery in the exact minute it should be discharging hardest.
+    """
+    ceiling, from_where = quarter_ceiling_w(cap_w, reading)
+    ceil_top = abs(float(achievable_w))
+
+    blind = blindness(projection, max_hold_s, frozen)
+    if blind is not None:
+        # 0.0, not a sized floor. There is no trustworthy measurement to size one
+        # from; the claim that remains is "do not add to the offtake you cannot
+        # see", which is one-directional and cannot cost pack energy.
+        return PeakDecision(
+            engaged=True, floor_w=0.0, ceiling_w=ceiling, ceiling_from=from_where,
+            blind=blind, projected_w=None,
+            reason=(f"PEAK GUARD BLIND ({blind}) - the battery may not import "
+                    f"while the billed quarter cannot be measured; commanding no "
+                    f"charging (floor 0 W)"))
+
+    projected = float(projection.projected_w)
+    if batt_w is None:
+        # main.py's P_SAFETY stop already owns a cycle with a missing input, at
+        # 0 W. Filing nothing is the honest answer: a floor sized from a
+        # measurement that is not there is a made-up number.
+        return PeakDecision(
+            engaged=False, floor_w=None, ceiling_w=ceiling,
+            ceiling_from=from_where, projected_w=projected,
+            reason="no battery measurement to size a floor from")
+    if QUARTER_S - float(projection.elapsed_s) <= 0.0:
+        # A block with no seconds left cannot be influenced. This is degenerate
+        # rather than urgent - `excess_w` answers 0.0 there and a floor equal to
+        # the battery's own current output would be a claim that says nothing.
+        return PeakDecision(
+            engaged=False, floor_w=None, ceiling_w=ceiling,
+            ceiling_from=from_where, projected_w=projected,
+            reason="the quarter has no seconds left to influence")
+
+    ex = excess_w(projection, ceiling)
+    wanted = float(batt_w) + ex
+    if wanted <= 0.0:
+        # ⚠️ THE TEST IS ON THE FLOOR, NOT ON `excess > 0`, AND THE REPLAY IS WHY.
+        # `excess` is the watts that must come off the METER, so it goes to zero
+        # as soon as the battery is already removing them - and a guard that
+        # disengaged there handed the actuator straight back, the offtake sprang
+        # back to the raw load, and the next cycle engaged again. On 2026-08-20's
+        # billed quarter that bang-bang cycled 17 times in 900 s and left the
+        # quarter 49 W over the limit it had just spent a quarter of an hour
+        # holding. §10.2: "an EMS that oscillates is worse than no EMS, both for
+        # hardware wear and - per §9.1 - for the peak it is trying to avoid."
+        #
+        # `batt_w + excess` is the offtake that would exist with the battery at
+        # zero, minus the level the quarter can afford. Testing THAT is the same
+        # arithmetic read as a level rather than as a switch: the claim is filed
+        # for as long as a positive floor is needed and decays continuously to
+        # nothing as the load falls, so there is still no hysteresis constant -
+        # the quantity being thresholded is simply the right one.
+        return PeakDecision(
+            engaged=False, floor_w=None, ceiling_w=ceiling,
+            ceiling_from=from_where, projected_w=projected, excess_w=ex,
+            wanted_w=wanted,
+            reason="the load is under the level this quarter can afford")
+
+    floor = min(wanted, ceil_top)
+    saturated = floor < wanted
+    return PeakDecision(
+        engaged=True, floor_w=floor, ceiling_w=ceiling, ceiling_from=from_where,
+        blind=None, projected_w=projected, excess_w=ex, wanted_w=wanted,
+        saturated=saturated,
+        reason=(
+            f"PEAK GUARD: quarter projects {projected:.0f} W against a "
+            f"{ceiling:.0f} W ceiling ({from_where}) with "
+            f"{QUARTER_S - projection.elapsed_s:.0f} s left, so {ex:.0f} W must "
+            f"come off the meter; battery measured {float(batt_w):+.0f} W, "
+            f"floor {floor:+.0f} W"
+            + (f" - SATURATED, {wanted:+.0f} W was needed and the inverter can "
+               f"only give {ceil_top:.0f} W" if saturated else "")))
+
+
+def shaping_suspended(decision, peak_forecast_w, cap_w) -> bool:
+    """Must the maintenance charge-only shaping stand down this cycle?
+
+    THE ONE PLACE THIS CONJUNCTION IS WRITTEN, and it is public for the reason
+    `watchdog.loop_is_following()` and `freeze.looks_frozen()` are: the check
+    that grades it has to assert on THE SHIPPED EXPRESSION. A safety predicate
+    transcribed into its own test is one copy plus a future disagreement about
+    what the check means, and the copy that drifts is always the grading one.
+    It was inline in `main.py` and nothing could reach it; a reviewer mutated it
+    back to its historical broken form and all twelve suites stayed green.
+
+    ⚠️ WHY SUPPRESSION AND NOT PRIORITY. The shaping is a LIMIT (`hi <= 0`) and
+    the peak guard's claim is a LIMIT (`lo >= 0`). arbiter.py clause 2 applies
+    EVERY limit at or above the winner's priority and takes the MOST RESTRICTIVE,
+    with no notion of one limit outranking another - which is precisely what
+    makes the safety bounds unconditional and must not be changed. Two crossing
+    limits are therefore a CONTRADICTION, which resolves to 0 W: the battery
+    PARKED, in the one quarter of the month that sets the bill. P_PEAK 50 >
+    P_MAINT_SHAPE 40 does not save it and never could.
+
+    ⚠️ AND THE FIRST TERM IS NOT ENOUGH ON ITS OWN. `peak_forecast_entity` is
+    OPTIONAL and is empty on a default install, so `peak_forecast_w` is None
+    there and `peak_at_risk()` answers False forever. Before this card the
+    suspension therefore could not fire on the very site that needed it. The
+    second term is the MEASURED guard, which is always present wherever P1 is.
+    Deleting either term - or re-gating the whole thing on the forecast - parks
+    a default install at 0 W, silently, in the billed quarter. test_peakguard.py
+    part 7 drives the real cycle with each of those mutations in place.
+    """
+    if peak_at_risk(peak_forecast_w, float(cap_w)):
+        return True
+    return decision is not None and decision.engaged
+
+
+def claim(decision: PeakDecision) -> Claim | None:
+    """The decision as one arbiter claim, or None when the guard is idle.
+
+    A LIMIT with only `lo` set. `hi` stays `+inf`, which is 9.1's "discharge and
+    export are never capped by the peak guard" expressed as a fact about the
+    object rather than as a rule somebody has to remember.
+    """
+    if not decision.engaged or decision.floor_w is None:
+        return None
+    return Claim.limit(SOURCE, P_PEAK, lo=float(decision.floor_w),
+                       reason=("blind" if decision.blind else "quarter peak"))

goodwe_controller/app/tariff.py

@@ -522,6 +522,7 @@ class Projection:
     covered_frac: float
     complete: bool
     block_start: datetime | None
+    held_w: float
 
     def explain(self) -> str:
         return (f"quarter projects {self.projected_w:.0f} W "
@@ -657,19 +658,38 @@ class PeakTracker:
         """
         start = averager.block_start
         if start is None:
-            return Projection(0.0, 0.0, 0.0, 0.0, False, None)
+            # ⚠️ A ZERO PROJECTION, AND IT IS NOT "THE QUARTER IS FINE". It is
+            # "there is no quarter". `block_start is None` is the flag that says
+            # which, and SAFETY-07 grades that flag BEFORE it grades the number -
+            # see peakguard.blindness(). A consumer that reads only
+            # `projected_w` here reads 0 W, which is under every ceiling there
+            # is, on exactly the startup path where nothing else is watching.
+            return Projection(0.0, 0.0, 0.0, 0.0, False, None, 0.0)
         elapsed_wall = max(0.0, (now - start).total_seconds())
         elapsed = averager.elapsed_s
         acc = averager.partial_ws
         remaining = max(0.0, QUARTER_S - elapsed_wall)
-        held = max(getattr(averager, "_last_net", 0.0), 0.0)
+        # ⚠️ NO `getattr` DEFAULT, AND THAT IS THE FIX, NOT A TIDY-UP. This read
+        # was `getattr(averager, "_last_net", 0.0)`: a private name with a
+        # fallback, so renaming the attribute would have kept every projection
+        # green while crediting 0 W held - a projection that reads LOW, which is
+        # the UNDER-GUARDING direction for SAFETY-07's floor. TEST-13 item 4.
+        # `held_net_w` is now a documented public property on QuarterAverager and
+        # is read without a default, so the same rename raises AttributeError on
+        # the first cycle instead of quietly disarming the money guard.
+        held = max(averager.held_net_w, 0.0)
         return Projection(
             projected_w=(acc + held * remaining) / QUARTER_S,
             elapsed_w=(acc / elapsed) if elapsed > 0 else 0.0,
             elapsed_s=elapsed_wall,
             covered_frac=(elapsed / elapsed_wall) if elapsed_wall > 0 else 0.0,
             complete=elapsed_wall <= 0 or elapsed >= elapsed_wall - 1e-9,
-            block_start=start)
+            block_start=start,
+            # Carried so a consumer can see WHAT was extrapolated, not only the
+            # extrapolation. It is the signed held value; `projected_w` uses the
+            # offtake-only form above, and the two differ whenever the site is
+            # exporting - which is the healthy operating point here.
+            held_w=averager.held_net_w)
 
 
 @dataclass(frozen=True)