FCST-04: the carry-forward stops decimating a quarter-hourly feed, and a broken cache stops stopping the add-on

Rework of the FCST-04 branch against review. Every finding below was proven by a runnable script before it was touched and re-run after; every check added or repaired was driven red by a mutation, `__pycache__` purged first (MUTATION.md, TEST-15). 29/29 test files green. ⚠️ H1 - A SUB-HOURLY FEED WAS DECIMATED TO ONE POINT AN HOUR, AND THE SPREAD WENT WITH IT. `fill_from_yesterday()` keyed the published set as `{_utc_hour(p.start): p for p in r.points}` - LAST POINT OF THE HOUR WINS, the other three dropped. `eneco_eddy` is the Belgian imbalance product and settles per QUARTER, and `_validate_raw()` accepts any strictly-increasing spacing, so this is a shape the product supports today. Measured on a 96-point quarter-hourly feed - ALL FIRM, FULLY PUBLISHED, nothing to carry: the plan reached the optimiser as 15 points, the arbitrage spread fell from 0.143 to 0.014 EUR/kWh, and the cheapest "hour" became a `:45` point whose `at()` window runs into the next one. Not the windowing effect this card is about; a feed losing 84 % of its own prints, with no flag anywhere reporting it. The fill now decides only WHICH WHOLE HOURS NEED CARRYING: every published point inside the window passes through BY IDENTITY, and a carried point is added only for a whole hour with no published coverage at all. Same feed, same fixture: 60 of 60 firm points kept, spread 0.134 over the hours still ahead. ⚠️ AND `_audit()` GOT STRICTER, NOT LOOSER - that was the condition on doing it at all. The old post-condition refused any repeated `_utc_hour`, a count that cannot express four points in one hour. Two rules replace it and SUBSUME it: a published point must BE one of that hour's published points by identity and may be emitted only once; a carried point may stand only in an hour absent from the published set, must be flagged True, and may not repeat. So the duplicated carried hour it existed to catch still raises, the same published point emitted twice now raises, and a carried point sharing an hour with a real print now raises too - which the count never looked at. Three mutations, three reds. ⚠️ H2 - A DAY RECORD THAT WAS NOT A RECORD STOPPED THE ADD-ON FROM STARTING. `_load()` validated `version` and that `days` is a dict, never that each VALUE is one. `{"version": 1, "days": {"2026-01-14": "oops"}}` loaded as ok=True, last_error=None, and then `rec.get(...)` raised AttributeError out of `words`, `yesterday()` and `resolve_tariff()` alike. `PriceHistory(...)` and `.words` are both unguarded in `Controller.__init__`, so the controller did not come up - and it never self-healed, because `_remember()` raised on the same record before it could reach `_save()`. That falsifies the CHANGELOG's "A BROKEN CACHE IS NOT A BROKEN CONTROLLER" in the one direction that matters. The asymmetry that produced it is visible in the file: `_parse_points()` is carefully defensive about every field of a POINT and `_load()` was not defensive about the record holding them. Unusable records are now dropped and REPORTED (`words` says how many and why), and `words` is total by construction as a second line. M3 - `yesterday()` WAS OFF BY ONE DAY FOR 00:00-00:59 AFTER SPRING-FORWARD. `brussels()` renders at the FIXED offset for that instant, so `brussels(stamp) - timedelta(days=1)` subtracts 24 ABSOLUTE hours - and the Brussels day before the March transition is 23 hours long. At 2026-03-30T00:30+02:00 it returned 2026-03-28's prices as "yesterday", or `no_data` naming a date that is not yesterday. The arithmetic is now on the local DATE, which has no offset to lose, and both transitions have a check - the module's headline claim had none. M4 - WITH NO /data THE STORE STILL FILLED, IN MEMORY. `_save()` short-circuited on `not ok` while `_remember()` populated `self._days` anyway: `words` said "unavailable" and `yesterday()` answered A SERIES that is gone on the next restart. The module docstring rejects exactly that ("a yesterday that silently forgets") as its reason for not calling makedirs, so the code now matches the docstring rather than contradicting it. M5 - THE PRUNE MUTATION SURVIVED BECAUSE THE STIMULUS WAS DEGENERATE. §8's clock-jump check seeded ONE day, so `if len(self._days) <= KEEP_DAYS: return` fired and the body it is named after never ran: pruning by arithmetic against the clock - which empties the store on exactly the morning the fill is for - passed the whole 29-file suite. The fixture is now written to a file and reopened (`_load()` does not prune, which is how a store legitimately holds more than the cap) with KEEP_DAYS + 2 days, and the mutant is red. M6 - THE AST ORDERING CHECK COULD NOT CATCH A REVERSAL, AND TWO DOCUMENTS CITED IT AS A GUARANTEE. `_calls.index("_remember") < _calls.index("yesterday")` grades `ast.walk` BFS DEPTH: `self._remember(...)` is a statement and `self.yesterday(...)` is an ARGUMENT of the fill, one level deeper, so `_remember` came out first whichever statement was written first - shipped and reversed both scored True. It is now line numbers AND, the half that matters, the behaviour: a feed carrying yesterday alongside today into an EMPTY store must fill the whole horizon; reversed, the lookup finds nothing and the plan is only the published hours. Reversing the two statements is red on both. L14/L9 - KEEP_DAYS = 3 IS THE STEADY STATE EXACTLY, so a newest-N cap has no slack: a feed publishing to D+2, or ONE stray point in the wrong year - which sorts LAST and can never age out - pushes yesterday out for good. `_prune()` now PINS the two days a lookup can actually ask for on top of the cap. Pinning is additive, so it can only RETAIN and the WITNESS-03 property (nothing is deleted by clock arithmetic) is preserved by construction; `stamp` stops being a dead parameter; the store is bounded at KEEP_DAYS + 2. L8 - the horizon end was not floored while `since` was, so a 48 h request made at any instant off the hour returned 49 points. §1's count was green only because its NOW was exactly :00:00, a shape production never has; the fixture now sits at :17:43. L10 - a failed `_save()` leaked its temp file; at 48 writes a day a persistent write failure leaked 48 inodes a day onto the volume that was already the problem. L15 - `{"version": 1.0}` passed, since `1.0 == 1`. L16 - the `_remember` tree walk scanned `app/*.py` only and matched a bare string; it now parses `app/`, `tools/`, `scenarios/` and `estop/` and looks for the NAME, so a comment does not fail it and a call in `tools/` cannot pass unseen. CLAIM-ACCURACY DEFECTS, WHICH THIS REPO TREATS AS REAL: * L7. DOCS said "*Tariff* on the status page reads `provisional` through the morning". FALSE - `tariff_words` is assigned only from the HA price-entity branch and from `static_status()`, both on the 30 s tick, and the plan's own sentence reaches `_LOG.info`. The doc now says what an operator can see (the Price history row, the Arbitrage row's own refusal sentence, the refresh log line) and a check pins the PAIRING - wiring the plan into the row is fine, correcting the doc is fine, doing one without the other is red. * L11. "fills EVERY unpublished hour" is conditional: a partial yesterday leaves the local hours it never held out of the series (§11.4 for those hours, and `at()` raises rather than answering). Correct behaviour, overclaimed wording, and untested - all three fixed. * L12. DECISIONS §1 leaned on `TariffPlan.cheap_now()`, which has NO PRODUCTION CALLER on the provider plan (`self.cheap` comes from `static_status()`, which builds its own). The conclusion holds on `arbitrage()`'s firm-pair refusal alone, which does reach control; the evidence is now stated at the reach it has. * L13. §1's "passed through by identity" check compared `eur_per_kwh` between two SEPARATE `resolve_tariff()` calls, where `is` could never have held. The value comparison is kept and now says what it grades; identity is graded on a read the file constructs and still holds. * L17. `PriceHistory.words` and the new status row had no check at all - and `words` is one of the two things H2 crashed. Both now have one. * L18. main.py's inline note carried the "widens what the optimiser can SEE, never what it may act on" guarantee without the windowing caveat that the CHANGELOG, DOCS and `resolve_tariff()`'s docstring all carry. It is the one place a maintainer reads it unqualified. EVIDENCE. All eight reviewer proofs re-run: p1/p1b return 29 March, p4 keeps 60 of 60 firm points at spread 0.134, p5 starts and reports, p6 D leaves no temp, p6 E refuses 1.0, p6 F records nothing, p7 returns 48 for 48. p2's script still scores True/True because it re-implements the OLD algorithm - the reversal is now driven as a mutation instead, and is red. p3's script always showed the shipped code correct; what was wrong was the fixture, and its mutant is now red. Twenty mutations driven, no survivors, tree fingerprinted and `__pycache__` purged before each: decimation (2 suites red), audit x3, unfloored `until` (2), unvalidated day records (7 checks), instant-subtraction yesterday, in-memory fill on a volume-less store, clock-arithmetic prune, cap without the pin, leaked temp, loose schema type, empty `words` (5), reversed `carry()` (2), a `_remember` leak in tools/ spelled two ways, a neighbour-filling fallback, a copy instead of identity (2), the DOCS claim in both directions, and the caveat removed from main.py. Baseline and the unmutated tree both green. No option added, config.yaml untouched, and the CHANGELOG entry still picks no version: PKG-07's rule, and only the deployer knows whether 0.9.0's image has been pulled. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BwRLw7eKYj2byVdwB6kbjU
FCST-04: the carry-forward fill reaches the plan, and the day it carries survives a restart
2026-09-09 23:54:23 +02:00 · 2026-09-09 19:37:45 +02:00 · 2026-09-09 01:05:36 +02:00 · 2026-09-09 00:58:03 +02:00 · 2026-09-09 00:29:00 +02:00 · 2026-09-08 23:51:35 +02:00
26 changed files with 6055 additions and 123 deletions
@@ -1,5 +1,105 @@
 # Changelog

+## Unreleased
+
+⚠️ **THIS HAS NO VERSION AND CANNOT REACH AN INSTALL UNTIL IT GETS ONE.**
+PKG-06's and PKG-07's rule: the Supervisor keys the built image by `version:`,
+so `ha apps update` is a silent no-op on an unbumped change. This branch
+deliberately does **not** pick the number, because the right one depends on a
+fact the branch cannot observe - whether 0.9.0's image has already been built
+and pulled. If it has not, this belongs inside 0.9.0's section before it ships;
+if it has, it needs 0.10.0. Whoever deploys decides.
+
+| | |
+|---|---|
+| _this branch_ | FCST-04 - the carry-forward fill reaches the plan, and the day it carries survives a restart |
+
+**FCST-03 shipped the fill and no caller.** `price_forecast.fill_from_yesterday()`
+was complete and checked, and so were the consumers written for it -
+`TariffPlan.firm_points()`, `cheap_now()`'s refusal of a provisional hour, and
+`arbitrage()`'s refusal 3, which drops provisional hours from the trade set
+rather than refusing the whole series. Nothing in `app/` called the fill, so no
+series ever contained a carried point and every one of those consumers was
+exercising its all-firm branch.
+
+⚠️ **THE DEFECT THAT PRODUCED WAS NOT AN ERROR ANYWHERE.** `load_prices()`
+returns what the feed published and nothing widened it to the requested horizon,
+so on a dynamic site between midnight and the 14-16 h CET day-ahead publication
+the plan covered the rest of today and **nothing more** - typically 10-24 of the
+48 hours `price_horizon_h` asked for. `arbitrage()` then took `max - min` over
+that truncated window and reported a spread. The number was real, the window was
+half the one configured, and **nothing said so**: tomorrow evening's peak, the
+hour a morning charge would be sold into, was not in the series at all, so the
+spread it would have created was never compared against anything.
+
+**What changes:**
+
+* `app/price_history.py` (new) keeps the **published** prices of the last few
+  Brussels days in `/data/price_history.json`, and `resolve_tariff()` now fills
+  an unpublished hour of the horizon from yesterday's same-wall-clock hour,
+  flagged provisional. ⚠️ **"An" unpublished hour, not "every" one:** the fill
+  can carry only an hour yesterday actually published, so a partial yesterday
+  leaves the local hours it never held **out of the series** - §11.4 for those
+  hours, and asking for one raises rather than answering. Correct behaviour,
+  and the wording is now the behaviour.
+* ⚠️ **A sub-hourly feed keeps every print.** `eneco_eddy` is the imbalance
+  product and settles per **quarter**; the fill walks whole hours to decide
+  what needs carrying, and that is not a rule about what a published hour may
+  contain. Every published point inside the horizon passes through by identity,
+  and a carried point is added only for a whole hour with no published coverage
+  at all. The first draft keyed the published set by hour - last point wins -
+  and a 96-point quarter-hourly feed reached the optimiser as 15 points with
+  its spread down from 0.143 to 0.014 EUR/kWh, on an all-firm, fully-published
+  series.
+* The horizon is **whole hours at both ends**: `since` is floored to the hour
+  that contains it, and so is `until`. Asked for 48 hours at 08:17, the plan is
+  48 hours long and not 49.
+* `resolve_tariff(history=...)` is **required and keyword-only**. A caller with
+  no durable volume passes the explicit `NoHistory()`. A default of `None`
+  meaning "skip the fill" would hand a caller who forgot the argument exactly
+  the truncated horizon this card removes, silently.
+* A new status row, **Price history**, names the days held and their hour
+  counts.
+
+⚠️ **ONE CHANGE HERE MOVES ENERGY, AND IT IS NOT THE ONE THE CARD IS NAMED FOR.**
+Carried hours are provisional, so §8.3 keeps them out of the trade set: the fill
+widens what the optimiser can **see**, never what it may **do**. What does
+change a trade is that the filled series is windowed to `[since, until)`, so
+**hours that have already elapsed are no longer in the spread**. Before, a cheap
+night that had already gone sat in the series as the day's cheapest hour - and a
+cycle cannot buy in an hour that is past, so the trade was quietly never taken.
+`test_price_history.py` section 2 grades this by finding the cheapest firm hour
+with and without the fill and showing it **moves**.
+
+⚠️ **A DEAD FEED IS STILL §11.4, NOT YESTERDAY.** The fill runs only on a
+`series` read; a feed that published nothing still degrades to the configured
+`static_profile`. The two are indistinguishable to the optimiser - both are
+100 % provisional, so arbitrage refuses on the firm-pair rung and `cheap_now()`
+is False - so preferring yesterday would change a documented degradation for a
+difference only the status page can see.
+`DECISIONS/FCST-04-the-day-that-is-carried.md` carries that and two more.
+
+⚠️ **A BROKEN CACHE IS NOT A BROKEN CONTROLLER.** No `/data`, an unreadable
+file, a schema this build does not know (including a `version` of `1.0`, which
+`== 1`), a **day record that is not a record**, a malformed point - each its own
+state with its own reason, none of them a price and none of them an empty
+series. In every one the plan is still a real provider series and the controller
+still regulates; the horizon is just as short as it was before this card.
+`.words` never raises, because `Controller.__init__` reads it unguarded and a
+status line that throws is an add-on that does not start.
+
+⚠️ **AND WHERE THERE IS NO `/data`, NOTHING IS RECORDED AT ALL** - not even in
+memory. An in-memory yesterday answers for one run and is gone on the next
+restart, which is a cache that forgets without saying so; the honest pair is
+`unavailable` on the status row and no prices carried.
+
+⚠️ **"YESTERDAY" IS A LOCAL DATE, AND THE ARITHMETIC IS ON THE DATE.**
+`brussels()` renders at a fixed offset, so subtracting `timedelta(days=1)` from
+it subtracts 24 **absolute** hours - and the Brussels day before a
+spring-forward is 23 hours long. Between 00:00 and 00:59 on the morning after
+the March transition that lands on the day *before* yesterday. **No option was
+added and `config.yaml` is untouched.**
+
 ## 0.9.0

 ⚠️ **MINOR FOR NEW FUNCTIONALITY, NOT FOR A BREAK - AND THAT DISTINCTION IS THE
@@ -0,0 +1,181 @@
+# FCST-04 — what "yesterday" is, and what the fill is allowed to reach
+
+Four decisions, recorded because each had a defensible alternative that a later
+reader would otherwise re-open from scratch.
+
+---
+
+## 1. The fill runs on a `series` read only. A dead feed still goes to §11.4.
+
+`fill_from_yesterday()` accepts a `no_data` read and will happily build a
+100 %-carried series from yesterday alone. `resolve_tariff()` does not let it:
+a feed that published nothing still falls to `_degrade()` and §11.4's own row —
+*"No price data → fall back to the static profile, mark the plan provisional,
+no arbitrage."*
+
+**Why not prefer yesterday's real prices to a coarse day/night profile?**
+Because it buys nothing the optimiser can use, and it silently replaces a
+fallback the operator configured.
+
+Measured, not assumed — and stated at the reach it actually has:
+
+- **This one reaches control.** A fully-carried series is 100 % provisional, so
+  `TariffPlan.arbitrage()` finds fewer than two firm hours and refuses on
+  §8.3's firm-pair rung. That is the verdict the optimiser reads: `arbitrage.py`
+  calls `plan.arbitrage(...)` on `fresh_tariff_plan()`, and it is the only path
+  by which a provider plan moves energy. The §11.4 static fallback refuses on
+  the same rung, for the same reason.
+- **This one does not, and saying so is the point.** `TariffPlan.cheap_now()`
+  also returns False on a fully-carried series, because `firm_points()` is
+  empty — but it has **no production caller on the provider plan**.
+  `self.cheap` comes from `static_status()`, which builds its *own* static plan
+  and calls `cheap_now()` on that. So it is a property of the type, checked in
+  `test_static_tariff.py`, and it is not evidence about the shipped control
+  path. The conclusion below holds on the first bullet alone.
+
+The two paths therefore differ only in which numbers the status page shows.
+
+So the choice is between two identical control behaviours, and the tie-breaker
+is that `static_profile` is a documented option an operator fills in **for this
+case**. Overriding it for a cosmetic difference is a behaviour change nobody
+asked for.
+
+⚠️ **The consequence, stated so it is not discovered later as a bug:**
+`fill_from_yesterday()`'s `no_data` branch is not reached from production.
+It stays checked in `test_price_forecast.py`. Turning it on is a product
+decision — it would want a rule about how old a carried profile may be before it
+stops being offered at all — and not a wiring one.
+
+---
+
+## 2. "Yesterday" is a date, not the newest record.
+
+`newest()` is the obvious accessor and it is the one that decays. An add-on down
+from Friday to Monday would answer Friday's prices to a fill whose only
+vocabulary is `provisional: true` — a flag that says "this is a guess", never
+"this guess is three days old". Nothing downstream can tell the difference and
+nothing reports it.
+
+So `yesterday(stamp)` computes the Brussels date before `stamp`'s Brussels date
+and looks up **that** record. Anything else is `no_data` with a reason naming
+both the date wanted and the dates held.
+
+The cost is real and accepted: after an outage of more than a day, the first
+morning back has no fill. That is the honest state — there is no yesterday —
+and the horizon is short for one morning rather than wrong for several.
+
+---
+
+## 3. What is recorded is the provider's read, and the mechanism is ordering.
+
+Feeding the filled series back into the store is how this rots: day 2 carries
+day 1's guesses, day 3 carries day 2's carry of them, and after a week of a dead
+feed the "yesterday" being carried is a fossil of the last real publication.
+Every point still carries `provisional=True`, so the **flag never lies** — what
+decays is the number, behind a flag that is telling the truth.
+
+⚠️ **No value-level check can catch that**, and claiming one would be worse than
+claiming none. A carried point and a published-provisional point (`eneco_eddy`'s
+current quarter, §9.2) are the same three fields carrying the same flag. The
+difference is provenance, and provenance is not in the value.
+
+The defence is therefore an **ordering**, not a validation:
+
+- `carry()` is the only public entry point. It records its argument, then
+  returns the fill.
+- The filled series exists only as that return value. The object never sees it
+  again.
+- `_remember()` is private and called from exactly one place.
+- `test_price_history.py` parses every shipped module under `app/`, `tools/`,
+  `scenarios/` and `estop/` and fails if any of them **uses the name**
+  `_remember` — an attribute, a bare name, a `def`, or the string constant a
+  `getattr` would need. By name rather than by substring, so a comment that
+  merely mentions the rule does not fail it and a call in `tools/` cannot pass
+  unseen.
+- and it pins the order two ways: the **line numbers** of the two calls in
+  `carry()`, and — the one that matters — the **behaviour**. A feed carrying
+  yesterday alongside today, into an empty store, must fill the whole horizon;
+  reversed, the lookup finds nothing and the plan is only the hours the feed
+  published. Driven, so a reversal is red.
+
+⚠️ **The first version of that ordering check graded nothing, and this file
+said it did.** It compared positions in `ast.walk`, which is breadth-first:
+`self._remember(...)` is a statement and `self.yesterday(...)` is an *argument*
+of `fill_from_yesterday(...)`, one level deeper, so `_remember` came out first
+whichever statement was written first. Measured — shipped and reversed both
+scored the same. A decorative check that a document cites as a guarantee is
+worse than no check, which is why the correction is recorded here rather than
+quietly made.
+
+That ordering has a second benefit worth keeping: a feed that returns yesterday
+alongside today refreshes yesterday's record from live data *before* it is
+consumed as the profile.
+
+---
+
+## What this ticket does not claim
+
+It does not make a trade possible that yesterday's prices justify. Carried hours
+are dropped from the trade set by §8.3, and that is not a limitation to be fixed
+later — it is the rule.
+
+The one change that **does** move energy came from the windowing, not the
+carrying: the filled series covers `[since, until)`, so hours that have already
+elapsed are no longer in the spread. A cheap night that has gone used to be the
+series' cheapest hour, and a cycle cannot buy in an hour that is past — so the
+trade was silently never taken. `test_price_history.py` section 2 grades that by
+finding the cheapest firm hour with and without the fill and showing it moves.
+
+⚠️ **And section 1's identity check now grades identity.** It used to compare
+`eur_per_kwh` between `_before` and `_after`, which come from two *separate*
+`resolve_tariff()` calls — each parsing the feed body into its own objects, so
+`is` could never have held between them however the fill were written. FCST-03's
+whole anti-copy argument is `filled.points[i] is fetched.points[j]`, and floats
+cannot grade it. The value comparison is kept and now says what it is (the fill
+moved the window, it did not touch a price); identity is graded on a read the
+file constructs and still holds a reference to.
+
+---
+
+## 4. A sub-hourly feed passes through whole; the fill only decides which hours are missing.
+
+`eneco_eddy` is the Belgian imbalance product and settles **per quarter**, and
+`price.py`'s `_validate_raw()` accepts any strictly-increasing spacing — so a
+published hour may legitimately hold four points.
+
+FCST-04's first caller keyed the published set as `{_utc_hour(p.start): p for p
+in r.points}`. It reads perfectly ordinary and it keeps the **last** point of
+each hour. Measured on a 96-point quarter-hourly feed — all firm, fully
+published, nothing to carry: the plan reached the optimiser as **15 points**,
+the arbitrage spread fell from **0.143 to 0.014 EUR/kWh**, and the cheapest
+"hour" became a `:45` point whose `at()` window runs into the next hour. That is
+not the windowing effect above; it is a feed losing 84 % of its own prints, and
+no `provisional` flag would ever have reported it.
+
+**The alternative was rejected:** refusing sub-hourly feeds at `_validate_raw()`
+would have made the count correct by breaking the one supported product that
+legitimately publishes per quarter.
+
+**What was done instead** keeps FCST-03's existing contract rather than weakening
+it: *every* published point inside the window passes through **by identity**,
+and a carried point is added only for a whole hour with **no** published
+coverage at all. A carried hour is still a single hourly point whatever the
+feed's own spacing is, because yesterday's profile is only known per wall-clock
+hour — and carried hours are provisional, so they are out of the trade set
+anyway.
+
+⚠️ **`_audit()` got stricter, not looser, and that was the condition on doing
+this at all.** The old post-condition refused any repeated `_utc_hour` — a count
+that cannot express four points in one hour. It is replaced by two rules that
+**subsume** it:
+
+- a **published** point must *be* one of that hour's published points, by
+  identity, and may be emitted only once;
+- a **carried** point may stand only in an hour that does not appear in the
+  published set at all, must be flagged `True`, and may not repeat.
+
+So a duplicated carried hour — the DST mispricing the count existed to catch —
+still raises; so does the same published point emitted twice; and a carried
+point sharing an hour with a real print now raises too, which the count never
+looked at. Each of those three is driven by its own mutation in
+`test_price_forecast.py`.
@@ -149,8 +149,70 @@ each swept across the **whole range `config.yaml` accepts** for the two settings
 that decide how fast the site can turn around: `gain` `float(0.05,1.0)` and
 `slew_w` `int(50,5000)`.

+### ⚠️ THE HARNESS BEHIND EVERY FIGURE ABOVE IS NOT COMMITTED, AND UNTIL 2026-09-07 THAT MADE THEM UNCHECKABLE
+
+**Read this before quoting any number in this section.** The paragraph below
+said the campaign harness "is not committed here" and treated that as a
+packaging detail. It is not: a figure produced by a script that exists in
+nobody's tree cannot be re-run, cannot be reviewed, and cannot be defended when
+it is contradicted. It has now been contradicted twice.
+
+| sweep | worst peak | worst continuous | trips |
+|---|---:|---:|---:|
+| the 924-run campaign below — **harness uncommitted** | 1700 W | 505.0 s at 100 W; 5.0 s above 700 W | 0 / 924 |
+| independent reviewer sweep, 2026-09-07, 48 honest reversals | **2750 W** | **15.0 s above 700 W** | 0 / 48 |
+| `admin/goodwe` `two_master_circulation`, 840 runs — **committed** | 1048 W | 10.0 s at 100 W; 5.0 s above 700 W | **0 / 840** |
+
+**This is not proof that anything was fabricated, and it should not be reported
+as one.** The three swept different grids. What the third one adds is a
+measurement of *why* the spread is that wide — on one grid, changing nothing but
+the model:
+
+| plant / stagger | worst legitimate peak |
+|---|---:|
+| modelled plant (lag 10 s, ramp 800 W/s), staggers up to a whole cycle | 1048 W |
+| `INSTANT` (lag 0, unbounded ramp), staggers up to a whole cycle | 2250 W |
+| `INSTANT`, staggers strictly under a whole cycle | **0 W** |
+
+A factor of 2.1×, and a factor of infinity, between three defensible
+descriptions of the same rig. **The plant model and the write-stagger set are
+terms of the number**, and a worst case quoted without them is not a
+measurement. That is enough to explain the whole spread above without anybody
+having invented anything — and it is exactly what cannot be established when the
+instrument is missing.
+
+**What survives, and what does not.**
+
+* **The window argument survives, and is now committed.** The claim that
+  decides `circulation_hold_s: 75` is *"the worst continuous legitimate episode
+  above 700 W is about one control cycle"*. All three sweeps agree on the shape
+  of that (5.0 s, 15.0 s, 5.0 s — one to three cycles against a fifteen-cycle
+  window), and **all three trip zero times**. `two_master_circulation` re-runs
+  it on every pull request in `admin/goodwe`.
+* **The threshold curve does NOT survive as stated.** `505.0 s at 100 W` is the
+  number the "knee at 400 W" argument for `circulation_w: 700` rests on, and the
+  committed sweep measures **10.0 s** at 100 W on its grid — 50× smaller. The
+  curve still *falls* with the threshold (126 of 840 runs cross 100 W against 1
+  that crosses 700 W), so the knee is real in direction; its depth is not
+  established by anything anybody can run. **Do not quote 505.0 s.**
+* **Nothing about the guard is compromised.** 0 trips in 924 + 48 + 840
+  non-faulty runs, from three independent grids.
+
+**What would settle it:** the original harness, committed. Until then the
+committed sweep is the only figure in this section anybody can check, and it is
+labelled as its own grid's rather than as a reproduction.
+
 ### Reproducing it

+⚠️ **THE COMMITTED VERSION IS `admin/goodwe`'s `sim/inject.py
+two_master_circulation`** (`sim/sitesim.py:circulation_findings`, runnable on its
+own as `python3 sim/sitesim.py --circulation`, `norig`, ~36 s, graded by
+`checks.yml` on every pull request there). It produces the fault with
+`plantsim.Inverter.wedge()` — a unit answering its bus and no longer steering its
+power stage, which is HW-01 — and sweeps the legitimate family with the
+serialised write path staggered up to a whole control cycle. The paragraph below
+describes the ORIGINAL campaign, which is the one that is not committed.
+
 The rig lives in `admin/goodwe` (`sim/`) and was used **read-only**; the harness
 is not committed here because it imports that repo's `plantsim`/`frame_check`,
 which are not present in this one or in CI. What it does is small enough to
@@ -206,7 +268,11 @@ margin above the knee is what absorbs that without re-opening this card.
 **Revisit when SPIKE-04 lands, not before.**

 The **schema floor is the knee itself**, 400 W, so a legally configured install
-cannot land in the 445–505 s regime. What the threshold costs: a circulating
+cannot land in the 445–505 s regime — ⚠️ and those two figures are from the
+uncommitted campaign; the committed sweep measures 10.0 s at 100 W on its own
+grid. See the disclosure at the head of this section. The floor is still the
+right shape (fewer runs cross a higher threshold, measured 126/840 at 100 W
+against 1/840 at 700 W); its DEPTH is not established by anything re-runnable. What the threshold costs: a circulating
 loop smaller than 350 W per unit is not reported, which is deliberate — at that
 scale the round-trip loss is a few watts.

@@ -289,6 +355,32 @@ SAFETY-25 — so no stale `.pyc` could exist to be read. Byte deltas recorded;
 | M17 | `circulation.py` | default `700` → `500` | **+0 B** | **RED** — the CFG-03 pin and the not-a-round-number check |
 | M18 | `main.py` | `await self.check_circulating()` removed | −39 B | **RED** — all four notification checks |
 | M19 | `main.py` | publish `round(x or 0.0)` instead of `None` | −68 B | **survivor on the first battery — GAP CLOSED**, then RED |
+| **M20** | `main.py` | **`pos, neg = opposing(...)` → `([], [])`** — the attribution clause deleted from the alarm | −102 B | **SURVIVED ALL 26 FILES on the shipped battery — GAP CLOSED**, then RED on §6a |
+| M21 | `main.py` | `pos` and `neg` exchanged — the alarm names both packs, both the wrong way round | **+0 B** | **RED** — §6a ×3 (the mirrored run, and "only its own sentence") |
+| M22 | `main.py` | `opposing(..., min_w=cfg.min_w / 2.0)` → `min_w=0.0` | −25 B | **RED** — §6a's resting-pack check |
+
+**And four in `admin/goodwe`**, graded through `python3 sim/sitesim.py
+--circulation` rather than through the injector — deliberately, because
+`inject.py` is the file that takes the rig lock, and "run the injector to grade a
+mutation" is not a sentence anybody should have to type. Byte deltas over
+`plantsim.py` + `sitesim.py` together.
+
+| # | file | mutation | Δ | result |
+|---|---|---|---|---|
+| S1 | `plantsim.py` | `wedge()` no longer stops the latch reaching the power stage | −20 B | **RED** — the opposing-signs precondition **and** "the watchdog did not trip" |
+| S2 | `sitesim.py` | the sensor computed over the **allocation record** instead of register 0x0518 | **+0 B** | **RED** ×3 — this is M10 restated on the rig |
+| S3 | `sitesim.py` | the campaign run on `plant=INSTANT` | −51 B | **RED** — the plant is a *term* of the figures; it was a SURVIVOR until that check was added, and finding out why produced §3's 1048/2250/0 W table |
+| S4 | `sitesim.py` | the write stagger removed from the legitimate campaign | −19 B | **RED** — "the legitimate campaign produced no opposition at all" |
+
+**S3 is recorded as a survivor because it was one, and it is the most useful
+result in this table.** The scenario first shipped with a vacuity check whose
+text claimed `plant=INSTANT` would make the sweep measure 0 W. S3 was run and
+**passed**: `INSTANT` with a stagger of exactly one control cycle samples the
+plant before the second unit has stepped, so it produces opposition anyway — and
+a *larger* worst peak (2250 W) than the modelled plant does (1048 W). The claim
+in the check's own text was wrong. It was corrected against the measurement, the
+plant was pinned as an explicit term, and S3 is red now for a stated reason
+rather than by accident. §3's spread table is that measurement.

 **M19 is recorded as a survivor because it was one.** The first battery found no
 check grading what `publish()` does when a per-unit reading is missing, so
@@ -297,6 +389,30 @@ watches went unnoticed — the sensor's own `None` rule, defeated one layer out.
 A check was added and the mutation re-run; it is red now. The run that found it
 was purged, so it is a result and not an artefact.

+**M20 IS RECORDED AS A SURVIVOR BECAUSE IT WAS ONE, AND IT IS THE MORE
+UNCOMFORTABLE OF THE TWO.** The commit message for this card claimed *"18
+mutations, all RED"*, and M1–M19 above have no entry for the notification's
+attribution clause. Verification on 2026-09-07 ran it and it **survived the
+whole 26-file suite**: the alarm still fires, still stops the site, still says
+the meter cannot see it — and no longer says *which two packs* or *which way
+round*. `circulation.py`'s own docstring calls that sentence the first thing
+somebody woken by this alarm needs and *"impossible to reconstruct from the
+number alone"*.
+
+It survived for one reason, and it is not reachability: **M18 (the whole call
+removed) is red**, so `check_circulating()` is exercised. The three checks on
+that notification asserted `METER CANNOT SEE`, `CONTROL HAS BEEN STOPPED` and
+`isolator` — and **never an inverter name**. Nothing graded the attribution.
+
+`test_watchdog.py` §6a closes it, and closes it on a fixture that can tell the
+three failures apart. It runs a roster of **`["garage", "shed", "cellar"]`** —
+operator names, none of them `UNIT_A`, for test_maintenance.py §11's reason:
+on a fixture whose names are the module's own constants a check cannot tell the
+roster from the constant. It runs the site **twice, mirrored**, so the sentence
+is not symmetric and a swap (M21, `+0 B`) is a different red from a deletion.
+And it carries a **third pack resting at 5 W**, so the floor that stops a pack
+in standby being named as a participant in a 6 kW loop (M22) is graded too.
+
 ## 7. What this card does NOT close

 * **Inverter B's ramp and lag are placeholders.** Every transient number in §3
@@ -305,7 +421,9 @@ was purged, so it is a result and not an artefact.
  margin. **Do not present any figure here as a measurement of inverter B.**
 * **A wedged inverter still cannot be stopped by this add-on.** §1 says so; the
  notification says so; there is no software fix and this card does not pretend
-  to one.
+  to one. It *can* now be produced on demand: `plantsim.Inverter.wedge()`.
+* **The depth of the threshold curve is not established by anything committed.**
+  §3's disclosure. The window is; the knee's magnitude is not.
 * **A loop below ~350 W per unit is not reported.** Deliberate, §3.
 * **The saturation freeze fires under a wedged unit and leaves a standing grid
  error.** Observed on the rig while producing §2's numbers: with B wedged, the
@@ -1424,6 +1424,70 @@ distinct state from "prices, all of them zero": the plan is marked provisional
 and arbitrage is disabled, per the degradation matrix. It never falls back to a
 price of zero, which would read as free electricity.

+#### The hours the feed has not published yet (FCST-04)
+
+A day-ahead auction publishes tomorrow's prices once a day, at **14–16 h CET**.
+Before that, the feed is serving today and nothing else — so a `price_horizon_h`
+of 48 asks for 48 hours and gets back the ten or fifteen that are left in the
+day. Until FCST-04 those other hours were simply **absent** from the plan: not
+priced badly, not flagged, just missing, and nothing said so.
+
+Now the add-on keeps the **published prices of the last few days** in
+`/data/price_history.json` and fills an unpublished hour of the horizon with
+yesterday's price for the same Brussels wall-clock hour, flagged
+`provisional: true`. Two things follow, and the second one matters more:
+
+- **The plan says it is guessing** — `provisional: true` on every carried hour,
+  which is what the optimiser reads. Before, a half-length plan looked
+  completely settled.
+- **Nothing may trade on a carried hour.** §8.3 forbids acting on a provisional
+  price, so arbitrage drops those hours from the trade set entirely — it needs a
+  buy hour and a sell hour that were *both* published. Carrying a price forward
+  widens what the optimiser can **see**; it never widens what it may **do**.
+
+**"An unpublished hour", not "every unpublished hour"** — the fill can only
+carry an hour yesterday actually published. If yesterday's record is itself
+partial (a first day, an outage, a feed that only ever serves the rest of the
+day), the local hours it never held are **left out of the series**: §11.4
+applies to those hours, and asking the plan for one raises rather than
+answering. An hour with no price is not a free hour, and the fill does not
+invent a neighbour, a mean or a zero for it.
+
+**A quarter-hourly feed keeps all four of its prints.** `eneco_eddy` settles
+per quarter, and the fill decides *which whole hours need carrying* — never what
+a published hour may contain. Every point the provider published inside the
+horizon is passed through untouched; a carried point is added only for a whole
+hour the feed did not print at all.
+
+**What you can actually see on the status page.** *Price history* (below) says
+what is held. *Arbitrage* carries the strategy's own sentence, which names the
+firm/provisional split whenever that split is what refused a trade. *Tariff*
+does **not** report the fill: it is the static-profile or Home Assistant
+price-entity sentence, regenerated on the 30 s tick. The provider plan's own
+line — `dynamic_dayahead from generic_dayahead, 48 h, provisional` — goes to the
+**add-on log** at every refresh, alongside `§8.3: 33 of 48 hours carried forward
+from yesterday's profile and flagged provisional; 15 published`.
+
+The plan is also now cut to the horizon you asked for, which means **hours that
+have already elapsed are no longer in it**. That is a real change to arbitrage:
+a cheap night that has already gone used to sit in the series as the cheapest
+hour of the day, and a cycle can never buy in an hour that is past — so the
+trade was quietly never taken. The spread is now taken over hours still ahead.
+
+A new status row, **Price history**, says which days are held and how many hours
+each has. If it reads `empty` or `unavailable`, nothing is carried, the horizon
+is as short as the feed makes it, and every other guarantee is unchanged — a
+cache that cannot answer is not allowed to be a controller that cannot regulate.
+The store is rebuilt by the next successful fetch; on a brand-new install the
+first fill is possible the day after the first refresh.
+
+The same applies to a file that is *partly* broken: a day record this build
+cannot read is dropped, the row says how many were dropped and why, and the rest
+of the file still loads. And where there is no `/data` at all the store records
+**nothing** — not even in memory. An in-memory yesterday would answer for one
+run and vanish on the next restart, which is a cache that forgets without saying
+so; `unavailable` and *no prices carried* are the honest pair.
+
 **Injection is priced separately** (`export_price_eur_kwh`). In Belgium it is
 often near zero, which by itself strongly favours self-consumption over export.
 It is a single number rather than a curve because no supported supplier
@@ -1574,6 +1638,185 @@ battery power, circulating power, state of charge, maintenance phase and
 controller status by MQTT discovery. This is observability only — the controller works fine without a
 broker, and MQTT problems can never affect control.

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

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

+* ⚠️ **ARCH-02's OWN REHEARSAL HARNESS, CAUGHT BY ITS OWN MUTATION BATTERY —
+  the THIRTEENTH instance, and the third in an INSTRUMENT rather than in a
+  check.** `tools/rehearse_roster.py` reports orphaned Home Assistant entities
+  by composing the ids the add-on announces **before** the roster switch and
+  **after** it, and diffing. Both sides are composed by the same build. So the
+  two mutations that change the id composition itself — renaming the MQTT device
+  (`"GoodWe"` -> `"GoodWe Controller"`), renaming one site sensor (`"Grid power"`
+  -> `"Grid power W"`) — moved **every id on both sides equally**, and the
+  harness reported *"no orphans"* about a change that orphans the entire live
+  entity list. It read exactly like a clean transition. The first draft's
+  M5/M6 came back `exit 0` and `exit 2`, and neither was the failure they name.
+
+  The fix is habit 2 with a filename: **the reference has to be data that does
+  not move with the code.** `--live-entities` takes a list of the ids the
+  install actually has, `tools/rehearsal/live-entities-recorded.txt` is the
+  recorded stand-in that keeps it live in CI, and a build that announces an id
+  the list does not carry is now a `COMPOSITION DRIFT` finding rather than a
+  silence. ⚠️ **The anchor moved for the same reason**: it had been
+  `compose(mqtt.DEVICE["name"], p1["name"]) == "sensor.goodwe_p1_sample_age"`,
+  which any renaming of the device satisfies by moving both operands, and it is
+  now composed from two recorded **literals**.
+
+  Ask of an instrument what habit 4 asks of a check: *where did each side of
+  this comparison come from* — and if a mutation of the subject moves both
+  sides, the instrument is decoration.
+
+  ⚠️ **AND REVIEW FOUND TWO MORE IN THE SAME CARD, ONE OF THEM INSIDE THE FIX
+  FOR THE OTHER — the FOURTEENTH and FIFTEENTH.** Both are the pin, not the
+  tool, and both are the same sentence in a new costume: *a check whose data
+  cannot reach the failure it names.*
+
+  * **Fourteenth — a COUNT cannot tell one red leg from two.** The rehearsal
+    reports two independent routes to the disarmed latch (`SAME-STEP` and
+    `STAGED`) and prints every finding **twice per red leg**: once in that
+    leg's own block and once in the closing summary. The check named *"NOT THE
+    SAME-STEP PATH ALONE"* was written `out.count("LATCH CAME UP DISARMED")
+    >= 2`, so **one** red leg always satisfied it. Measured on three builds:
+    shipped code (both legs red) → 4; `migrate_unit_keys()` handed the roster
+    name (SAME-STEP fixed) → 2, **passes**; a guard-side fix (both fixed) → 0.
+    The check reported `ok` on the build whose only remaining failure *is* the
+    path it says is not the only one. The fix is to parse the legs and require
+    the finding in **each**, which is habit 2 again: grade the thing, not a
+    total over the thing.
+
+  * **Fifteenth — and it is INSIDE THAT FIX, which is the third time this file
+    has had to write that sentence.** The first leg-splitter opened on a
+    `LEG <id>` marker and never closed, so every line after the *last* marker
+    was attributed to that leg — including the closing summary, which repeats
+    every finding from every leg. The last leg therefore read red whenever
+    **any** leg was red, and the new per-leg check was the old count with extra
+    steps. It was caught by building the one fixture that separates them: a
+    tree in which `SAME-STEP` is genuinely clean and `STAGED` is not. The tool
+    now prints `END LEG <id>` and the splitter closes on it.
+
+    The rule that falls out, and it is the cheap one: **a per-thing check that
+    has never seen the things DISAGREE is a total.** Part 5b exists to make
+    them disagree on every run.
+
+    ⚠️ **AND ONE THAT HAS ONLY SEEN THEM DISAGREE IN ONE ARRANGEMENT IS A TOTAL
+    OVER THAT ARRANGEMENT.** Review found this by attacking the fix itself. Part
+    5b does make the legs disagree, on every run - and was still blind to a build
+    that merely REORDERS them, because it had only ever seen the clean leg
+    printed LAST. Three edits, two of them ordinary refactors (reorder the report
+    legs, reword a summary line), produced `all checks passed` from a splitter
+    that closed ZERO legs.
+
+    A fixture cannot enumerate the arrangements, so the STRUCTURE is asserted
+    instead: a leg may not open while one is open, and every leg that opens must
+    be explicitly closed. That fires on the FIRST edit of all three attacks,
+    whatever the order or the wording. **When a fixture cannot enumerate the
+    arrangements, assert the invariant they share.**
+
 * **CFG-04's `tools/supervisor_shape_probe.py` §B** — and this one is the rule
  reaching outside the suite. The probe is the committed evidence that
  `config.yaml`'s `inverters: []` default is load-bearing, and four documents
@@ -68,6 +68,7 @@ from . import peakguard
 from . import packshare
 from .p1 import P1Ingest, build_source, is_enabled
 from .price import Uncertainty, price_option
+from .price_history import PriceHistory
 from .static_tariff import resolve_tariff, static_status
 from .tariff import PeakStore, PeakTracker
 from .store import migrate_maint_keys, migrate_unit_keys
@@ -165,7 +166,8 @@ def pack_energy_of(inv):

 class Controller:
    def __init__(self, opts: dict, hass: HomeAssistant, store, mqtt_pub, *, stop,
-                 peak_db_path: str = "/data/tariff.db"):
+                 peak_db_path: str = "/data/tariff.db",
+                 price_history_path: str = "/data/price_history.json"):
        self.o = opts
        self.hass = hass
        self.store = store
@@ -247,7 +249,40 @@ class Controller:
            hold_s=float(opts.get("delivery_hold_s",
                                  DELIVERY_OPTION_DEFAULTS["delivery_hold_s"])),
        )
-        self.store_migration = migrate_unit_keys(store, UNIT_A)
+        # SAFETY-44. THE NAMES THE MACHINES READ, HANDED TO THE MIGRATION THAT
+        # WRITES FOR THEM. This line used to read `migrate_unit_keys(store,
+        # UNIT_A)`, and the constant was wrong in exactly the way the
+        # `maint_migration` call below says the SAME constant would be wrong -
+        # ⚠️ this comment used to cite "line 315" and "two hundred lines
+        # below", and both were wrong. The correction I wrote for them was
+        # wrong too: it said "~54 lines down", which was review's distance to
+        # the WRONG line, carried to a new referent without re-measuring -
+        # inside the sentence apologising for a wrong number. There is no
+        # distance here now, because a distance is a coordinate and the
+        # sentence above already names the thing. On a declared roster the flat
+        # latch landed under
+        # `unit/inverter-a/`, which is on no roster, was read back under the
+        # name it was handed, and stamped the marker. `roster_names()` is the
+        # naming rule itself and returns `[UNIT_A]` on the flat shape, so the
+        # upgrade path every install runs today is byte-identical.
+        #
+        # ⚠️ AND `misdirected` IS UNREACHABLE FROM **THIS** CALL SITE, WHICH THE
+        # COMMIT MESSAGE ORIGINALLY CLAIMED AS A BENEFIT. `unit` is
+        # `roster_names(opts)[0]` and `build_roster()` names its entries from
+        # the same `roster_names()` list, so `unit` is a member of `roster` by
+        # construction and the refusal can never fire here. Review measured it:
+        # dropping the `roster` argument from the `_verify_moves()` call inside
+        # `migrate_unit_keys()` survives 28/28.
+        #
+        # THE ARGUMENT STAYS ANYWAY, and the reason is the one `_verify_moves()`
+        # already gives for keeping its own `roster=None` branch: it is what a
+        # NEW caller gets, and the grading it enables is measured at the
+        # function's own level in test_maintenance.py rather than through here.
+        # What is NOT true is that this call site buys the refusal, so that
+        # claim is gone rather than restated.
+        _roster_names = roster_names(opts)
+        self.store_migration = migrate_unit_keys(
+            store, _roster_names[0], _roster_names)
        self.inverters = build_roster(opts, store, sign_cfg, delivery_cfg)
        # ⚠️ CFG-04. SAY WHICH SHAPE IS IN FORCE, BECAUSE THE OTHER ONE'S KEYS
        # ARE STILL SITTING IN THE UI. A declared `inverters:` roster reads none
@@ -549,6 +584,42 @@ class Controller:
        self.peak.restore(self.p1.averager)
        _LOG.info("capacity-tariff peak store: %s", self.peak.restore_state)

+        # FCST-04. The §8.3 carry-forward store: the published prices of the
+        # last few Brussels days, so the hours a day-ahead feed has not printed
+        # yet can be filled from yesterday's profile and FLAGGED, instead of
+        # being absent from the series entirely.
+        #
+        # ⚠️ IT IS AN INSTRUMENT'S BACKING FILE, NOT A STRATEGY. It commands
+        # nothing and claims nothing; what it changes is how many hours the
+        # plan covers. Before this, a dynamic site between midnight and the
+        # 14-16 h CET publication had a series covering only the rest of today
+        # - typically 10-24 of the 48 hours `price_horizon_h` asked for - and
+        # `TariffPlan.arbitrage()` took its spread over that truncated window
+        # without anything reporting that it had. Every carried hour is
+        # provisional, so refusal 3 drops it from the trade set: CARRYING a
+        # price widens what the optimiser can SEE, never what it may act on.
+        #
+        # ⚠️ AND THE CAVEAT THAT BELONGS WITH THAT SENTENCE, because this is
+        # the one place a maintainer reads it on its own. THE OTHER HALF OF
+        # THIS CHANGE DOES MOVE ENERGY, and it is the windowing rather than the
+        # carrying: the filled series covers `[since, until)`, so hours that
+        # have ALREADY ELAPSED are no longer in the spread. A cheap night that
+        # has gone used to sit in the series as the day's cheapest hour, and no
+        # cycle can buy in an hour that is past - so the trade was quietly
+        # never taken. `resolve_tariff()`'s docstring, DOCS.md and the CHANGELOG
+        # all carry this beside the guarantee; it may not appear here without it.
+        #
+        # ⚠️ A BROKEN OR ABSENT STORE IS NOT AN ERROR HERE. It answers `no_data`,
+        # the fill carries nothing, and the plan is the pre-FCST-04 one minus
+        # the elapsed hours - short, but honest and still refusing on the same
+        # rules. It must not be an error in the other sense either: a day
+        # record this build cannot read is DROPPED by `PriceHistory._load()`
+        # and reported in `.words`, because both the constructor above and the
+        # `.words` read below are unguarded, and a cache that raises here is an
+        # add-on that does not start.
+        self.price_history = PriceHistory(price_history_path)
+        _LOG.info("§8.3 price history: %s", self.price_history.words)
+
        # 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
@@ -751,23 +822,132 @@ class Controller:
        legacy_tripped = bool(data.get("sign_tripped", False))
        if not legacy_tripped or not self.inverters:
            return
-        # ⚠️ THE QUESTION IS "IS ANY NAMESPACED LATCH SET", AND IT IS ASKED OF
-        # THE STORE RATHER THAN OF THE ROSTER. At a roster of two a trip on B
-        # leaves A untripped, so re-arming A off a flat key that B's own trip
-        # mirrored would latch a healthy inverter on every restart - the
-        # flat-key defect this card removed, coming back through its own guard.
-        # And the STORE is the right place to ask: it holds every
-        # `unit/*/sign_tripped` this install has ever written, including one
-        # belonging to a unit that is not in TODAY's roster (a roster is
-        # configuration and it changes), whereas the roster only knows what is
-        # configured right now. The flat key carries new information only when
-        # nothing namespaced is latched at all.
-        namespaced_latched = any(
-            bool(v) for k, v in data.items()
-            if k.startswith("unit/") and k.endswith("/sign_tripped"))
-        if namespaced_latched or any(inv.sign_wd.tripped
-                                     for inv in self.inverters):
+        # ⚠️ THE QUESTION IS "IS ANY NAMESPACED LATCH SET **THAT A MACHINE
+        # READS**", AND THE LAST CLAUSE IS SAFETY-44's SECOND HALF. At a roster
+        # of two a trip on B leaves A untripped, so re-arming A off a flat key
+        # that B's own trip mirrored would latch a healthy inverter on every
+        # restart - the flat-key defect DEV-02 removed, coming back through its
+        # own guard. That much stands, and it is why the question is asked of
+        # the STORE's keys rather than of the roster's objects.
+        #
+        # ⚠️ BUT IT USED TO ACCEPT **ANY** `unit/*/sign_tripped`, and the
+        # paragraph here defended that: the store "holds every one this install
+        # has ever written, including one belonging to a unit that is not in
+        # TODAY's roster ... whereas the roster only knows what is configured
+        # right now". THE PREMISE IS TRUE AND IT ARGUES THE WRONG WAY. A latch
+        # under a name no unit answers to is not protection an operator has - it
+        # is a key nothing reads, which is the definition of the state this
+        # method exists to repair. Counting it as protection made the guard
+        # return early on precisely the install that needed it, and that is not
+        # hypothetical: it is where the migration above USED to put the latch
+        # on every declared roster, so the guard was blinded by its own upgrade.
+        #
+        # THE MIGRATION FIX DOES NOT REACH THOSE INSTALLS. A store stamped by
+        # the shipped version already has the marker, so `migrate_unit_keys()`
+        # returns at the version check and never runs again. This guard is the
+        # ONLY repair path for a site that upgraded before this card, and it
+        # re-arms onto `inverters[0]` - which lands the latch back under a name
+        # the roster reads.
+        # ⚠️ AND "OFF THE ROSTER" IS NOT ENOUGH ON ITS OWN - IT HAS TO BE THE
+        # MIGRATION'S OWN FINGERPRINT, WHICH IS `UNIT_A`. The first draft of
+        # this treated EVERY off-roster latch as litter, and test_inverter.py's
+        # DEV-02 restart case caught it: there the store carries
+        # `unit/inverter-b/sign_tripped = True` from B's own trip while the
+        # roster being graded holds only A, and the blunt rule re-armed a
+        # HEALTHY A off the flat key that B's trip had mirrored - DEV-02's
+        # defect rebuilt inside its own guard, which is exactly what the
+        # paragraph above was right to be afraid of.
+        #
+        # `UNIT_A` is the discriminator because it is the ONLY name the broken
+        # migration ever wrote. `roster_names()` falls back to a POSITIONAL
+        # `inverter-N`, so `inverter-a` is not a name this code hands to anyone
+        # else. A latch stranded under any OTHER name is a real unit that is
+        # out of today's config - a roster is configuration and it changes -
+        # and it still counts as protection.
+        #
+        # ⚠️ THE RESIDUAL, STATED RATHER THAN LEFT TO BE FOUND: a site whose
+        # own first unit is NAMED `inverter-a`, latched, and then taken out of
+        # the roster gets re-armed on `inverters[0]`. That is a healthy unit
+        # held at 0 W until an operator clears it, which is the conservative
+        # direction; SAFETY-44's is a site DRIVING on a state file that says it
+        # is latched.
+        #
+        # ⚠️ AND THE SECOND HALF OF THAT RESIDUAL IS THE **SENTENCE**, NOT THE
+        # BOOLEAN. The reason re-armed from is the FLAT mirror, which carries
+        # whatever tripped LAST - so on a multi-unit site this path can print
+        # one inverter's evidence under another inverter's name. The latch is
+        # right and the explanation may not be. It is disclosed here because the
+        # banner is what an operator acts on, and `test_maintenance.py`'s check
+        # that the reason is not the removed unit's is guaranteed by that
+        # fixture's choice of flat string, NOT by this code.
+        # ⚠️ AND THE THREE CASES ARE THREE VARIABLES, NOT ONE SHARED FLAG.
+        # THIS IS WHERE THE FIRST VERSION OF THIS FIX WAS DEFEATED, BY ONE KEY.
+        # It wrote `namespaced_latched = True` for BOTH "on the roster" and "off
+        # the roster but not UNIT_A", so a single latched key belonging to a
+        # unit the operator had removed from `inverters:` vetoed the repair of
+        # the misdirected `inverter-a` latch sitting beside it. Measured on the
+        # shipped build, roster ['garage', 'shed'], store stamped so the
+        # migration cannot reach it:
+        #
+        #   sign_tripped + unit/inverter-a/sign_tripped
+        #       -> [('garage', True), ('shed', False)]   repaired
+        #   ...AND unit/old-shed/sign_tripped
+        #       -> [('garage', False), ('shed', False)]  ⚠️ DRIVING
+        #
+        # A site DRIVING on a state file that reads `sign_tripped: true`, which
+        # is the sentence this whole card exists to make false - and
+        # `tools/rehearse_roster.py` said so, on the unmodified tree, exit 1.
+        #
+        # ⚠️ IT IS ALSO INSTANCE #17 OF HABIT 4, IN THE EXTENSION MUTATION.md
+        # ADDED TWO DAYS AGO. The two fixtures written for this card each hold
+        # ONE arrangement - `_staged_store` has only `inverter-a`, `_other_store`
+        # has only `inverter-b` - and the flag is written twice with opposite
+        # intents only when BOTH are present. A per-key check whose keys have
+        # never disagreed WITHIN ONE STORE is a check over whichever key came
+        # last. The fixture below now holds both.
+        #
+        # THE THREE CASES, AND WHY THE ANSWER DIFFERS FOR EACH:
+        #   on the roster            a machine reads it: real protection, stop.
+        #   off-roster, `UNIT_A`     the broken migration's fingerprint - this
+        #                            IS the flat latch, misdirected. Repair.
+        #   off-roster, other name   a real unit out of today's config. Its
+        #                            latch is unread too, but it is NOT the flat
+        #                            key's misdirection, and re-arming a healthy
+        #                            sibling off it is DEV-02's defect. Stop -
+        #                            UNLESS a stranded `UNIT_A` latch is also
+        #                            present, because then the flat key IS
+        #                            accounted for and the repair is the point.
+        on_roster = {inv.name for inv in self.inverters}
+        stranded, on_roster_latched, other_off_roster = [], False, False
+        for k, v in data.items():
+            if not (k.startswith("unit/") and k.endswith("/sign_tripped")):
+                continue
+            if not bool(v):
+                continue
+            # `unit_key()` builds `unit/<unit>/<key>`, so the unit is field 1 -
+            # split through the writer's own spelling rather than re-deriving.
+            unit = k.split("/")[1]
+            if unit in on_roster:
+                on_roster_latched = True
+            elif unit == UNIT_A:
+                stranded.append(k)
+            else:
+                other_off_roster = True
+        if on_roster_latched or any(inv.sign_wd.tripped
+                                    for inv in self.inverters):
            return
+        if other_off_roster and not stranded:
+            # A removed unit's own latch, and no misdirected one. The flat key
+            # is that unit's mirror, so it carries no new information and
+            # re-arming `inverters[0]` from it would latch a healthy pack.
+            return
+        if stranded:
+            _LOG.error(
+                "SAFETY-44: this store carries a LATCHED sign-disagreement "
+                "failsafe under %s, and that unit is NOT on today's roster "
+                "%s - so no machine reads it and the site came up DRIVING on "
+                "a state file that says it is latched. Re-arming below.",
+                ", ".join(sorted(stranded)), sorted(on_roster))
        first = self.inverters[0]
        # The one state this method exists for.
        reason = str(data.get("sign_reason", "") or "")
@@ -3313,7 +3493,8 @@ class Controller:
            now = datetime.now(timezone.utc)
            try:
                plan = await resolve_tariff(
-                    self.o, now, now + timedelta(hours=horizon), now=now)
+                    self.o, now, now + timedelta(hours=horizon), now=now,
+                    history=self.price_history)
                self.tariff_plan, self.tariff_at = plan, now
                _LOG.info("tariff refreshed: %s", plan.words)
            except Exception as err:                          # noqa: BLE001
@@ -4162,6 +4343,13 @@ class Controller:
            ("Maintenance due", "now" if next_due is None else next_due.strftime("%Y-%m-%d")),
            ("Cheap window", "yes" if self.cheap else "no"),
            ("Tariff", self.tariff_words),
+            # FCST-04. WHAT §8.3 HAS TO CARRY FORWARD, which is the difference
+            # between a 48-hour plan and a 12-hour one on a weekday morning.
+            # It is its own row rather than a clause of "Tariff" because it
+            # answers a different question: "Tariff" says what the plan IS, and
+            # this says what the plan could still be filled from if the feed is
+            # late again tomorrow.
+            ("Price history", self.price_history.words),
            # OPT-04. The strategy's own sentence, not a re-derivation of it -
            # every refusal it can make already carries a reason a human can
            # act on, and "Arbitrage: no" would throw all of them away.
@@ -48,7 +48,8 @@ So the flag is not passed through anywhere in this file:
    the object the provider produced is the object in the output series - so
    there is no copy that could lose or gain a flag. `filled.points[i] is
    fetched.points[j]` is a checkable property, and test_price_forecast.py
-    checks it.
+    checks it. EVERY published point in the window is passed through, not one
+    per hour: see the sub-hourly note below.
  * `_audit()` RE-DERIVES the expected flag for every output point from the set
    of hours the feed actually published, and RAISES if any point disagrees.
    It is a post-condition, not a filter: it never fixes a flag, because a
@@ -90,6 +91,24 @@ which is the row the spec actually writes for "no price data". If NOTHING can
 be produced, the result is `PriceRead.no_data(reason)` and not an empty series,
 because an empty series is not constructible on purpose (price.py trap 3).

+⚠️ A SUB-HOURLY FEED IS NOT DECIMATED TO ONE POINT AN HOUR, and the hourly walk
+below is about WHICH HOURS NEED FILLING - never about what a published hour is
+allowed to contain. `eneco_eddy` is the Belgian imbalance product and settles
+per QUARTER, `_validate_raw()` accepts any strictly-increasing spacing, and a
+published hour may therefore hold four points. So EVERY published point inside
+the window is passed through, in order, and a carried point is added only for a
+whole hour with NO published coverage at all.
+
+The alternative was measured rather than reasoned about, because FCST-04's
+first caller shipped it: keying the published set as `{_utc_hour(p.start): p
+for p in r.points}` reads perfectly ordinary and keeps the LAST point of each
+hour, dropping the other three. On a 96-point quarter-hourly feed - all firm,
+fully published, nothing to carry - the series came out 15 points long, the
+arbitrage spread collapsed from 0.143 to 0.014 EUR/kWh, and the cheapest
+"hour" became a `:45` point whose `at()` window runs into the next hour. A feed
+losing 84 % of its own prints is not a windowing effect and is not something a
+provisional flag would ever have reported.
+

 AC2: THE IRREVERSIBLE-ACTION GATE
 =================================
@@ -156,6 +175,25 @@ def _utc_hour(instant: datetime) -> datetime:
    return u.replace(minute=0, second=0, microsecond=0)


+def _published_by_hour(read: PriceRead) -> dict:
+    """UTC hour -> the provider's OWN points that start inside it, in order.
+
+    ⚠️ A TUPLE PER HOUR, NOT A POINT, AND THAT IS THE WHOLE OF THE SUB-HOURLY
+    RULE. `{_utc_hour(p.start): p for p in r.points}` reads identically, is one
+    character shorter, and silently keeps one point per hour - the last one
+    written. The mapping answers two DIFFERENT questions here: "does this hour
+    need filling at all?" is `u in published`, and "what does the output carry
+    for it?" is all of `published[u]`. Collapsing them is how a quarter-hourly
+    feed loses three quarters of its prints.
+    """
+    out: dict = {}
+    read.match(
+        series=lambda r: [out.setdefault(_utc_hour(p.start), []).append(p)
+                          for p in r.points],
+        no_data=lambda _r: [])
+    return {u: tuple(ps) for u, ps in out.items()}
+
+
 def _carried(start_utc: datetime, eur_per_kwh: float) -> PricePoint:
    """A point filled from yesterday's profile. §8.3: ALWAYS provisional.

@@ -172,37 +210,69 @@ def _carried(start_utc: datetime, eur_per_kwh: float) -> PricePoint:
 def _audit(points, published) -> None:
    """Post-condition: every flag matches the point's TRUE provenance (AC1).

-    `published` maps a UTC hour to the point the provider produced for it.
+    `published` maps a UTC hour to the TUPLE of points the provider produced
+    for it (`_published_by_hour()`). An hour ABSENT from that mapping is an hour
+    the feed did not print, and a carried point is the only thing that may
+    stand in it.
+
+    The rule, in the two forms an output point can take:
+
+      * PUBLISHED - the point must BE one of `published[u]`, by identity, and
+        may be emitted only once. Its flag is whatever the feed said, never
+        forced to False: an imbalance print is published AND provisional (§9.2,
+        TARIFF-02), and a check that "corrected" it would be a bug wearing a
+        guarantee.
+      * CARRIED - the hour must not appear in `published` AT ALL, the flag must
+        be exactly True, and that hour may hold no second point.
+
+    ⚠️ WHAT REPLACED "AT MOST ONE POINT PER HOUR", AND WHY IT IS STRICTER. This
+    audit began as a `seen` set: any repeated `_utc_hour` raised, because a
+    duplicated hour is how a DST day gets mispriced. A bare hour count cannot
+    express a quarter-hourly feed, where four distinct published points share
+    one hour legitimately - so the count was replaced by the two rules above,
+    which subsume it. A repeated CARRIED hour still raises; so does the same
+    published point emitted twice; and a carried point may no longer share an
+    hour with a published one, which the count never checked at all.

    ⚠️ RAISES, NEVER REPAIRS. A module that silently corrects its own
    provenance is one whose provenance means nothing; the only safe response to
    "this point's flag disagrees with where it came from" is to refuse to emit
    the series at all.
    """
-    seen = set()
+    carried_hours = set()
+    emitted = set()
    for p in points:
        u = _utc_hour(p.start)
-        if u in seen:
-            raise ValueError(f"the filled series repeats {u.isoformat()}; a "
-                             f"duplicated hour is how a DST day gets mispriced")
-        seen.add(u)
        src = published.get(u)
-        prov = Provenance.CARRIED if src is None else Provenance.PUBLISHED
        if src is None:
+            if u in carried_hours:
+                raise ValueError(
+                    f"the filled series repeats the carried hour "
+                    f"{u.isoformat()}; a duplicated hour is how a DST day gets "
+                    f"mispriced")
+            carried_hours.add(u)
            if p.provisional is not True:
                raise ValueError(
-                    f"{u.isoformat()} is {prov.value} and is flagged "
-                    f"provisional={p.provisional!r}. §8.3: an hour filled from "
-                    f"yesterday's profile is provisional whether or not the "
-                    f"number turns out to be right - the flag is a claim about "
-                    f"PROVENANCE, not about accuracy.")
+                    f"{u.isoformat()} is {Provenance.CARRIED.value} and is "
+                    f"flagged provisional={p.provisional!r}. §8.3: an hour "
+                    f"filled from yesterday's profile is provisional whether "
+                    f"or not the number turns out to be right - the flag is a "
+                    f"claim about PROVENANCE, not about accuracy.")
        else:
-            if p is not src:
+            if not any(p is q for q in src):
                raise ValueError(
-                    f"{u.isoformat()} is {prov.value} but the series carries a "
-                    f"COPY of that point rather than the point itself. A copy "
-                    f"is where the provisional flag gets lost; published points "
-                    f"pass through by identity.")
+                    f"{u.isoformat()} is {Provenance.PUBLISHED.value} but the "
+                    f"series carries a COPY of that point rather than the "
+                    f"point itself. A copy is where the provisional flag gets "
+                    f"lost; published points pass through by identity.")
+            if id(p) in emitted:
+                raise ValueError(
+                    f"the filled series emits the published point at "
+                    f"{p.start.isoformat()} TWICE. The provider's own points "
+                    f"are strictly increasing (`_validate_raw`), so a repeat "
+                    f"is never a second print - it is a duplicated hour, which "
+                    f"is how a DST day gets mispriced")
+            emitted.add(id(p))


 def fill_from_yesterday(read: PriceRead, *, yesterday: PriceRead,
@@ -217,11 +287,21 @@ def fill_from_yesterday(read: PriceRead, *, yesterday: PriceRead,

    Returns a `PriceRead`:

-      * a series, whose points are the provider's own objects for the hours it
-        published and `_carried()` points for the rest, ordered by instant;
+      * a series, carrying the provider's own objects - ALL of them, which on
+        a quarter-hourly feed is four inside one hour - for the hours it
+        published, and one `_carried()` point for each remaining whole hour,
+        ordered by instant;
      * `no_data` when neither source can produce a single hour - never an
        empty series, which is not constructible (price.py trap 3).

+    ⚠️ THE WINDOW IS WHOLE HOURS AT BOTH ENDS. `since` is floored to the hour
+    that CONTAINS it, because the current hour is still a tradeable hour; and
+    so is `until`, which is the half FCST-04's first draft left out. With
+    `since` floored and `until` compared raw, a 48 h horizon asked for at 08:17
+    emitted 49 hours - the hour of `until` was still `< until` - and the count
+    was one too many for every `now` that is not exactly on the hour, which is
+    every `now` a running controller has.
+
    ⚠️ IDEMPOTENT AND RE-RUNNABLE, WHICH IS THE §5.4 LATE-PUBLICATION CONTRACT.
    Day-ahead publication at 14-16 h CET can be late or fail, so the fetcher
    retries and this runs again against the fresher read. Published hours always
@@ -234,9 +314,7 @@ def fill_from_yesterday(read: PriceRead, *, yesterday: PriceRead,
            raise ValueError(f"{name} must be tz-aware; a naive instant is "
                             f"ambiguous for one hour every October")

-    published = read.match(
-        series=lambda r: {_utc_hour(p.start): p for p in r.points},
-        no_data=lambda _r: {})
+    published = _published_by_hour(read)

    # Yesterday's profile, keyed by BRUSSELS WALL-CLOCK HOUR. First point wins
    # when yesterday itself was a 25-hour day and printed local hour 2 twice.
@@ -248,11 +326,12 @@ def fill_from_yesterday(read: PriceRead, *, yesterday: PriceRead,

    out = []
    u = _utc_hour(since)
-    end = until.astimezone(timezone.utc)
+    end = _utc_hour(until)
    while u < end:
-        if u in published:
-            out.append(published[u])                    # BY IDENTITY. No copy.
-        else:
+        got = published.get(u)
+        if got:
+            out.extend(got)              # BY IDENTITY, and ALL of them. No copy,
+        else:                            # and no decimation to one an hour.
            src = by_local_hour.get(brussels(u).hour)
            if src is not None:
                out.append(_carried(u, src.eur_per_kwh))
@@ -0,0 +1,598 @@
+"""FCST-04 - the day the carry-forward carries, kept somewhere a restart cannot reach.
+
+WHAT WAS MISSING, AND WHY IT WAS INVISIBLE. FCST-03 shipped
+`price_forecast.fill_from_yesterday()` complete and checked, and shipped the
+consumers that were written to expect it - `TariffPlan.firm_points()`,
+`cheap_now()`'s refusal of a provisional hour, and `arbitrage()`'s refusal 3,
+which drops provisional hours from the trade set instead of refusing the whole
+series. Every one of those was live. The FILL ITSELF HAD NO CALLER: nothing in
+`app/` referenced `fill_from_yesterday`, so no series ever contained a carried
+point and every one of those consumers was exercising its all-firm branch.
+
+The defect that produced is not an error anywhere. `load_prices()` returns
+whatever the feed published and nothing widens it to the requested horizon, so
+on a dynamic site between midnight and the 14-16 h CET day-ahead publication
+the series covers the rest of TODAY and nothing more - typically 10-24 of the
+48 hours `price_horizon_h` asked for. `TariffPlan.arbitrage()` then computes
+`max - min` over that truncated window and reports a spread. The number is
+real, the window is half the one the operator configured, and NOTHING SAYS SO:
+tomorrow evening's peak - the hour a morning charge would be sold into - is
+simply not in the series, so the spread it would have created cannot clear a
+threshold it was never compared against. A missing hour is not a wrong price;
+it is an hour the optimiser never learns exists.
+
+So this module is the half FCST-03 left out, and it is deliberately the SMALL
+half: the fill's arithmetic, its DST rule and its provenance audit all stay in
+`price_forecast.py`, untouched. What lives here is the one thing that
+arithmetic needs and could not have: A YESTERDAY THAT SURVIVES A RESTART.
+
+
+WHY PERSISTENCE, AND NOT A VARIABLE ON THE REFRESH TASK
+======================================================
+An in-memory "the last full day I saw" costs nothing and is wrong exactly when
+it matters. The window the fill exists for is midnight -> publication; an
+add-on restarted at 08:00 - a config change, an HA update, a power cut - would
+come up with no yesterday at all and spend that whole window on the truncated
+series, which is the pre-FCST-03 behaviour arriving silently on the mornings
+after an operator touched something. The fill's own value is a function of
+surviving a restart, so the day it carries has to.
+
+⚠️ AND NOT IN `/data/state.json`. That file holds the maintenance history and
+SAFETY-02's sign latch, and `Store.set()` rewrites the whole of it on every
+write. The tariff refresh runs every `TARIFF_REFRESH_S` (1800 s), so folding a
+price cache into it would put 48 extra whole-file rewrites a day through the
+file whose corruption reads as "no maintenance ever ran" and "never tripped".
+The cache is worth having; it is not worth putting anywhere near that. Separate
+file, separate blast radius.
+
+⚠️ AND THE DIRECTORY IS NOT CREATED, which is `tariff.PeakStore`'s rule and its
+reason applies unchanged: `/data` is mounted by the Supervisor and always
+exists on a real install, so `makedirs` here could only ever succeed where the
+durable volume is MISSING - manufacturing a cache on the container's ephemeral
+layer that writes fine, reads fine, and is thrown away on the next restart.
+That is a yesterday that silently forgets, which is worse than one that says it
+cannot answer.
+
+⚠️ AND THAT RULE IS CARRIED OUT IN `_remember()`, NOT ONLY IN `_save()`. FCST-04
+first shipped it half-done: `_save()` short-circuited on `not self.ok` while
+`_remember()` filled `self._days` anyway, so a store with no durable volume
+answered `words` = "unavailable" and `yesterday()` = A SERIES - an in-memory
+yesterday that works all day and is gone on the next restart, which is exactly
+the silently-forgetting cache this rule exists to refuse. A store that cannot
+persist now records nothing, and both of its answers say the same thing.
+
+
+THE FOUR RULES THIS FILE IS SHAPED BY
+=====================================
+
+1. IT RECORDS WHAT THE PROVIDER PUBLISHED, NEVER WHAT THE FILL PRODUCED.
+   Feeding a filled series back in is the defect that makes the whole mechanism
+   rot: day 2 carries day 1's guesses, day 3 carries day 2's carry of them, and
+   after a week of a dead feed the "yesterday" being carried forward is a
+   fossil of the last real publication with nothing anywhere reporting its age.
+   Every point still carries `provisional=True`, so the flag never lies - what
+   decays is the NUMBER, silently, behind a flag that is telling the truth.
+
+   ⚠️ NO VALUE-LEVEL CHECK CAN CATCH THAT, and saying so is more useful than a
+   guarantee that would not hold. A carried point and a published-provisional
+   point (`eneco_eddy`'s current quarter, §9.2) are the same three fields with
+   the same flag; the difference is provenance, and provenance is not in the
+   value. So the mechanism is ORDERING, not validation: `carry()` is the only
+   public entry point, it remembers the argument and returns the fill, and the
+   filled series exists only as a return value that `carry()` itself never
+   sees again. `_remember()` is private, is called from exactly one place, and
+   `test_price_history.py` PARSES every shipped module under `app/`, `tools/`,
+   `scenarios/` and `estop/` and fails if any of them uses the NAME - an
+   attribute, a bare name, a def, or the string constant a `getattr` needs. By
+   name rather than by substring, so a comment about the rule does not fail it
+   and a call from `tools/` cannot pass unseen.
+
+   ⚠️ AND THE ORDER IS GRADED BY DRIVING IT, NOT BY READING THE SOURCE. The
+   first version of that check compared positions in `ast.walk`, which is
+   breadth-first: `self._remember(...)` is a statement and `self.yesterday(...)`
+   is an ARGUMENT of the fill, one level deeper, so `_remember` came out first
+   whichever statement was written first - shipped and reversed scored the
+   same. What grades it now is the behaviour: a feed carrying yesterday
+   alongside today, into an EMPTY store, must fill the whole horizon, and
+   reversed it fills nothing.
+
+2. "YESTERDAY" MEANS YESTERDAY, NOT "THE NEWEST DAY I HAVE".
+   `newest()` is the natural accessor and it is the one that decays: an add-on
+   down from Friday to Monday would answer Friday's prices to a fill that flags
+   them as one day old. The lookup is BY DATE - the Brussels day before the
+   Brussels day of the instant asked about - and a record for any other date is
+   NO_RECORD with a reason naming both dates. A three-day-old profile is not a
+   worse yesterday; it is not a yesterday.
+
+3. A STORE THAT CANNOT ANSWER SAYS SO, IN THE TYPE.
+   `yesterday()` returns a `PriceRead`: a `series` when a full record exists,
+   `no_data(reason)` for every other outcome - no file, no record for that
+   date, an unreadable file, a record that parses to nothing. There is no
+   empty-series answer and no `.get(..., [])`-shaped accessor, for
+   `price.py`'s reason: an empty series is how "the cache is broken" gets to
+   wear the shape of "yesterday was free". `fill_from_yesterday()` takes a
+   `PriceRead` and handles `no_data` on its own, so the degraded case needs no
+   branch at the call site - it fills nothing and the horizon stays short,
+   which is exactly the behaviour that shipped before this ticket.
+
+4. HOURS ACCUMULATE WITHIN A DAY; THEY DO NOT REPLACE IT.
+   A feed that serves only the REMAINING hours of the day (a legitimate shape -
+   nothing in the plugin contract promises a whole day) would, under
+   replace-the-record, overwrite a complete 24-hour profile with the two points
+   left at 22:00 and hand the next morning a yesterday with two hours in it. So
+   a record is MERGED by instant: a later read updates the hours it carries and
+   leaves the rest, and a revised price for an hour already held wins because
+   it is the newer statement about the same hour.
+
+⚠️ KEYED BY INSTANT, NOT BY LOCAL HOUR, AND THAT IS THE DST DECISION. Storing
+`hour -> price` would collapse the two 02:00s of a 25-hour October day into
+one, inventing a second and different collapsing rule beside the one
+`fill_from_yesterday()` already documents ("the first of its two 02:00 points
+is the one used, deterministically"). Storing the instants keeps both, in
+order, and lets the fill's own rule be the only one in the system.
+
+WHAT THIS DOES NOT DO, ON PURPOSE
+=================================
+It does not extend the fill to a feed that published NOTHING. `resolve_tariff()`
+applies `carry()` only on a `series` read; a `no_data` read still falls to
+`_degrade()` and §11.4's row - "No price data -> fall back to the static
+profile, mark the plan provisional, no arbitrage" - which is the fallback the
+operator configured `static_profile` FOR. Preferring yesterday's feed over it
+would change a documented degradation without being asked, and it would buy
+nothing measurable: a fully-carried series is 100 % provisional, so
+`arbitrage()` finds fewer than two firm hours and refuses, and `cheap_now()` is
+False, exactly as on the §11.4 fallback. The two differ only in which numbers
+the status page shows. `fill_from_yesterday()` handles a `no_data` read anyway,
+and that branch is therefore not reached from production - it stays checked,
+and turning it on is a product decision, not a wiring one. See
+DECISIONS/FCST-04-the-day-that-is-carried.md.
+"""
+
+import json
+import logging
+import os
+import tempfile
+from datetime import datetime, timedelta, timezone
+
+from .price import PricePoint, PriceProduct, PriceRead, brussels
+from .price_forecast import fill_from_yesterday
+
+_LOG = logging.getLogger("goodwe.price.history")
+
+# How many Brussels days of published prices the SORT-ORDER cap keeps. The fill
+# needs exactly one - yesterday - and rule 2 refuses anything older, so this is
+# not a horizon: it is the number of days a record may sit unused before
+# pruning removes it, and it is small because a bigger cache would only hold
+# days no lookup can ask for. Three covers today, yesterday, and the day a
+# fetch spanning midnight writes ahead of itself.
+#
+# ⚠️ IT IS A CAP, NOT THE RETENTION RULE, AND THREE HAS NO SLACK. The steady
+# state is already {yesterday, today, tomorrow}; a feed that publishes to D+2,
+# or ONE stray point in the wrong year - which sorts LAST and is therefore
+# immortal under a newest-N cap - pushes yesterday out of a three-slot store
+# and keeps it out, permanently and silently. So `_prune()` PINS the two days a
+# lookup can actually ask for on top of the cap; see its docstring. The store
+# is then bounded at KEEP_DAYS + 2 records rather than at KEEP_DAYS.
+KEEP_DAYS = 3
+
+SCHEMA = 1
+
+
+def _bdate(instant: datetime) -> str:
+    """The Brussels calendar date of an instant, as `YYYY-MM-DD`.
+
+    ⚠️ BRUSSELS, NOT UTC. 23:30 UTC in July is already tomorrow in Belgium, and
+    a cache keyed on the UTC date would file the last 90-120 minutes of every
+    day under the wrong one - so a lookup on the following morning would find a
+    "yesterday" missing its own evening hours, which are the expensive ones.
+    """
+    return brussels(instant).strftime("%Y-%m-%d")
+
+
+def _yesterday_of(instant: datetime) -> str:
+    """The Brussels date of the day BEFORE `instant`'s Brussels day.
+
+    ⚠️ THE ARITHMETIC IS ON THE LOCAL DATE, NOT ON THE INSTANT, AND THAT IS A
+    DST DEFECT THIS FILE SHIPPED. `brussels()` renders at the FIXED offset in
+    force at that instant, so `brussels(stamp) - timedelta(days=1)` subtracts 24
+    ABSOLUTE hours from a wall clock - and the Brussels day before a
+    spring-forward is 23 hours long. Measured at 2026-03-30T00:30+02:00, the
+    first hour of the morning after the March transition: it returned
+    2026-03-28, so the fill carried a TWO-DAY-OLD profile flagged as yesterday's
+    (or, with only the last three days held, answered `no_data` naming a date
+    that is not yesterday).
+
+    A calendar day minus one day is `date` arithmetic, and a `date` has no
+    offset to lose. The 25-hour October day is right for the same reason.
+    """
+    return (brussels(instant).date() - timedelta(days=1)).strftime("%Y-%m-%d")
+
+
+def _n_points(rec) -> int:
+    """How many points a day record holds. 0 for anything that is not one.
+
+    ⚠️ TOLERANT, AND IT IS THE SECOND LINE OF DEFENCE, NOT THE FIRST. `_load()`
+    drops a malformed day record and says so, so this should never see one. It
+    is written this way because `words` is read from `Controller.__init__` and
+    from every status render: a status LINE that raises takes the add-on down,
+    and "the cache is broken" must never be able to mean "the controller does
+    not start". That is exactly what FCST-04 shipped - `{"days": {"2026-01-14":
+    "oops"}}` gave `ok=True`, `last_error=None`, and an `AttributeError` out of
+    `words`, `yesterday()` and `resolve_tariff()` alike, with no self-heal
+    because `_remember()` raised before it could rewrite the file.
+    """
+    if not isinstance(rec, dict):
+        return 0
+    pts = rec.get("points")
+    return len(pts) if isinstance(pts, (list, tuple)) else 0
+
+
+class PriceHistory:
+    """The published price series of the last few Brussels days, in /data.
+
+    ⚠️ `ok` IS ABOUT THE FILE, NOT ABOUT THE DATA. A healthy store with no
+    record for yesterday is `ok` and answers `no_data` - a normal state on a
+    site's first day, and a different thing from a store that could not be
+    opened. Both are visible: `ok`/`last_error` describe the file, and
+    `yesterday()`'s reason describes the answer.
+    """
+
+    def __init__(self, path: str = "/data/price_history.json"):
+        self.path = path
+        self.ok = False
+        self.last_error: str | None = None
+        self._days: dict = {}
+        self._open()
+
+    # -- the file ------------------------------------------------------------
+    def _open(self) -> None:
+        parent = os.path.dirname(self.path)
+        # See the module note: NOT created. A cache on the ephemeral layer is
+        # one that forgets without saying so.
+        if parent and not os.path.isdir(parent):
+            self.last_error = f"no durable volume at {parent}"
+            # INFO, not ERROR: on a dev host or in a test container there is no
+            # /data and nothing is wrong.
+            _LOG.info("no price history: %s - §8.3 carry-forward has nothing "
+                      "to carry and the horizon stays as short as the feed",
+                      self.last_error)
+            return
+        self.ok = True
+        self._load()
+
+    def _load(self) -> None:
+        try:
+            with open(self.path, "r", encoding="utf-8") as fh:
+                blob = json.load(fh)
+        except FileNotFoundError:
+            _LOG.info("no price history yet at %s - the first refresh writes "
+                      "one, and the fill has nothing to carry until the day "
+                      "after that", self.path)
+            return
+        except (json.JSONDecodeError, OSError) as err:
+            # A corrupt cache must not stop the controller, and must not be
+            # silently re-created either: the record is discarded, the reason
+            # is kept, and `yesterday()` reports it rather than answering
+            # "no record", which would read as a normal first day.
+            self.last_error = f"unreadable ({err})"
+            _LOG.warning("price history at %s is %s - starting from empty. "
+                         "§8.3 carry-forward has nothing to carry until the "
+                         "next refresh has written a full day",
+                         self.path, self.last_error)
+            return
+        found = blob.get("version") if isinstance(blob, dict) else None
+        # ⚠️ THE TYPE IS PART OF THE VERSION. `1.0 == 1` in Python, so a float
+        # written by some other tool would pass a bare `!=` and be read as
+        # schema 1 - a version check that cannot tell 1 from 1.0 is one that
+        # would not notice a writer this build has never met.
+        if not isinstance(blob, dict) or isinstance(found, bool) \
+                or not isinstance(found, int) or found != SCHEMA:
+            self.last_error = f"schema {found!r}, not {SCHEMA}"
+            _LOG.warning("price history at %s has %s - discarding it rather "
+                         "than guessing what its points mean",
+                         self.path, self.last_error)
+            return
+        days = blob.get("days")
+        if not isinstance(days, dict):
+            self.last_error = "no `days` object"
+            _LOG.warning("price history at %s has %s", self.path, self.last_error)
+            return
+        self._days = self._usable_days(days)
+
+    def _usable_days(self, days: dict) -> dict:
+        """The day records that are shaped like day records. The rest are dropped.
+
+        ⚠️ THE ASYMMETRY THAT PRODUCED THE DEFECT: `_parse_points()` below is
+        carefully defensive about every field of a POINT, and `_load()` was not
+        defensive about the record holding them at all. `{"version": 1, "days":
+        {"2026-01-14": "oops"}}` therefore loaded as `ok=True`,
+        `last_error=None` - and then `rec.get(...)` raised `AttributeError` out
+        of `words`, `yesterday()` and `resolve_tariff()`. `PriceHistory(...)`
+        and `.words` are both unguarded in `Controller.__init__`, so THE ADD-ON
+        DID NOT START; and it never self-healed, because `_remember()` raised on
+        the same record before it could reach `_save()`.
+
+        A broken cache is a short horizon. It is not a controller that will not
+        regulate a battery, and §6's rule - absent machinery must not disable
+        the loop - does not stop applying because the machinery is a JSON file.
+
+        Dropped, not repaired: an unusable record has nothing in it to keep, and
+        the next `_remember()` writes a fresh one under that date.
+        """
+        kept, dropped = {}, []
+        for key, rec in days.items():
+            if not isinstance(key, str) or not isinstance(rec, dict) \
+                    or not isinstance(rec.get("points"), list):
+                dropped.append(str(key))
+                continue
+            kept[key] = rec
+        if dropped:
+            # Recorded, not silent: `yesterday()` and `words` both report
+            # `last_error`, and a cache that quietly discards half a file reads
+            # to an operator exactly like a normal first day.
+            self.last_error = (f"{len(dropped)} unusable day record(s) dropped "
+                               f"({', '.join(sorted(dropped))})")
+            _LOG.warning("price history at %s: %s - the controller starts "
+                         "anyway and the horizon is as short as the feed makes "
+                         "it", self.path, self.last_error)
+        return kept
+
+    def _save(self) -> None:
+        if not self.ok:
+            return
+        # ⚠️ THE TEMP FILE IS REMOVED ON EVERY FAILING PATH. A write that dies
+        # between `mkstemp` and `os.replace` used to leave its temp behind, and
+        # the refresh task runs 48 times a day: a persistent write failure -
+        # a full volume being the likeliest cause - leaked 48 inodes a day onto
+        # the volume that was already the problem.
+        tmp = None
+        try:
+            fd, tmp = tempfile.mkstemp(dir=os.path.dirname(self.path))
+            with os.fdopen(fd, "w", encoding="utf-8") as fh:
+                json.dump({"version": SCHEMA, "days": self._days}, fh, indent=1)
+            # Atomic: a power cut mid-write must not leave a truncated file
+            # that parses as an empty day.
+            os.replace(tmp, self.path)
+            tmp = None                      # it IS the file now; do not unlink
+        except OSError as err:
+            self.last_error = f"could not write ({err})"
+            _LOG.error("could not persist price history: %s", err)
+        finally:
+            if tmp is not None:
+                try:
+                    os.unlink(tmp)
+                except OSError:             # nothing further to try, and a
+                    pass                    # failed cleanup is not a fault
+
+    # -- the one public entry point -----------------------------------------
+    def carry(self, read: PriceRead, *, since: datetime, until: datetime,
+              stamp: datetime) -> PriceRead:
+        """Remember what this read published, then §8.3-fill the missing hours.
+
+        THE ORDER IS DELIBERATE AND IS THE POINT OF THE METHOD. The read is
+        recorded FIRST, so a feed that hands back yesterday alongside today
+        refreshes yesterday's record from live data before it is consumed as
+        the profile; and the filled series - the one thing that must never be
+        recorded - exists only as this method's return value, which this object
+        never sees again. See rule 1, including how that order is graded.
+
+        Returns whatever `fill_from_yesterday()` returns: a series over the
+        whole hours of [since, until), carrying EVERY point the provider
+        published inside it - by identity, and four of them in one hour on a
+        quarter-hourly feed - plus one carried point for each remaining whole
+        hour FOR WHICH YESTERDAY HELD THAT WALL-CLOCK HOUR. An hour neither
+        source can price is left out (§11.4 for that hour), and `no_data` comes
+        back when that is every hour.
+        """
+        self._remember(read, stamp)
+        return fill_from_yesterday(read, yesterday=self.yesterday(stamp),
+                                   since=since, until=until, stamp=stamp)
+
+    # -- the halves, private because rule 1 is an ordering ------------------
+    def _remember(self, read: PriceRead, stamp: datetime) -> None:
+        """File a provider read's points under their own Brussels days.
+
+        ⚠️ PRIVATE, AND CALLED FROM EXACTLY ONE PLACE. A caller that can reach
+        this can record a filled series, which is the decay in rule 1. There is
+        no value-level defence against that - a carried point and a published
+        provisional point are the same three fields - so the defence is that
+        nothing outside this file names it, and `test_price_history.py` checks
+        the tree for that rather than trusting the underscore.
+        """
+        if not self.ok:
+            # ⚠️ A STORE THAT CANNOT PERSIST RECORDS NOTHING. See the module
+            # docstring's third warning: filling `self._days` here would give a
+            # site with no /data an in-memory yesterday that answers a series
+            # all day and is gone on the next restart - a cache that forgets
+            # without saying so, which is the one thing that rule refuses. The
+            # honest degraded state is `words` = "unavailable" and `yesterday()`
+            # = `no_data`, and those two must agree.
+            return
+        points = read.match(series=lambda r: r.points, no_data=lambda _r: ())
+        if not points:
+            return
+        product = read.match(series=lambda r: r.product.value,
+                             no_data=lambda _r: None)
+        provider = read.match(series=lambda r: r.provider,
+                              no_data=lambda _r: None)
+
+        touched = set()
+        for p in points:
+            day = _bdate(p.start)
+            rec = self._days.setdefault(
+                day, {"product": product, "provider": provider, "points": []})
+            # Rule 4: MERGE by instant. A later read updates the hours it
+            # carries and leaves the rest; a revised price for an hour already
+            # held wins, because it is the newer statement about that hour.
+            raw = rec.get("points")
+            held = {q["start"]: q for q in (raw if isinstance(raw, list) else ())
+                    if isinstance(q, dict) and "start" in q}
+            held[p.start.isoformat()] = p.canonical()
+            rec["points"] = sorted(held.values(), key=lambda q: q["start"])
+            rec["product"], rec["provider"] = product, provider
+            touched.add(day)
+
+        self._prune(stamp)
+        self._save()
+        _LOG.debug("price history: %d point(s) filed under %s; %d day(s) held",
+                   len(points), ", ".join(sorted(touched)), len(self._days))
+
+    def _prune(self, stamp: datetime) -> None:
+        """Drop records no lookup can reach. A sort-order cap, plus two pins.
+
+        ⚠️ NOTHING IS DELETED BY ARITHMETIC AGAINST THE CLOCK. A clock that
+        jumped - the failure mode WITNESS-03 spent a session on - would
+        otherwise delete every record in one pass and leave the fill with
+        nothing on the morning after, which is the one morning it is for.
+        Deleting on a bad clock is unrecoverable; holding a few days too many
+        is not, so the cap is `sorted(...)[-KEEP_DAYS:]` and stays that way.
+
+        ⚠️ AND A CAP ALONE IS DEFEATED BY ONE POINT IN THE WRONG YEAR. Sorting
+        is lexicographic on the date string, so a stray far-future record sorts
+        LAST and is immortal - it consumes a slot for good, and with the steady
+        state already {yesterday, today, tomorrow} filling all three, yesterday
+        is what falls out. A feed that publishes to D+2 does the same thing
+        without any stray at all. So the two days `yesterday()` can ACTUALLY be
+        asked for are pinned on top of the cap.
+
+        `stamp` is READ here, which it was not before: the pin is the only use
+        of it, and it is additive. Pinning can only RETAIN a record, never
+        delete one, so a jumped `stamp` pins two dates that are not held and
+        changes nothing - the WITNESS-03 property is preserved by construction
+        rather than by remembering to check it.
+        """
+        if len(self._days) <= KEEP_DAYS:
+            return
+        keep = set(sorted(self._days)[-KEEP_DAYS:])
+        keep |= {_bdate(stamp), _yesterday_of(stamp)}
+        for day in [d for d in self._days if d not in keep]:
+            del self._days[day]
+
+    def yesterday(self, stamp: datetime) -> PriceRead:
+        """The published series of the Brussels day BEFORE `stamp`'s, or no_data.
+
+        ⚠️ THE DATE IS COMPUTED, NOT LOOKED UP. See rule 2: `newest()` would
+        answer Friday's prices to a Monday fill, and the fill would flag them as
+        one day old because that is the only thing its flag can say.
+        """
+        if stamp.tzinfo is None or stamp.utcoffset() is None:
+            raise ValueError(
+                f"yesterday({stamp!r}): a naive datetime is not an instant, "
+                f"and 'which Brussels day was yesterday' is exactly the "
+                f"question a wall-clock reading cannot answer.")
+        want = _yesterday_of(stamp)
+        if self.last_error and not self._days:
+            return PriceRead.no_data(
+                f"the price history at {self.path} is {self.last_error}, so "
+                f"§8.3 has no profile to carry forward for {want}; the plan "
+                f"covers only the hours the feed published")
+        rec = self._days.get(want)
+        if rec is None:
+            have = ", ".join(sorted(self._days)) or "no days at all"
+            return PriceRead.no_data(
+                f"no published prices recorded for {want} (held: {have}), so "
+                f"§8.3 has no profile to carry forward. A record for another "
+                f"day is not a yesterday")
+        points, bad = _parse_points(rec.get("points"))
+        if not points:
+            why = bad or "it is empty"
+            return PriceRead.no_data(
+                f"the {want} record holds no usable point ({why}), and an "
+                f"empty series is not a free day")
+        try:
+            product = PriceProduct(rec.get("product"))
+        except ValueError:
+            return PriceRead.no_data(
+                f"the {want} record names product {rec.get('product')!r}, "
+                f"which is not a PriceProduct. The optimiser switches on that "
+                f"enum, so a record that cannot name one cannot be a series")
+        if bad:
+            _LOG.warning("price history: %s in the %s record were dropped; "
+                         "%d point(s) carried forward", bad, want, len(points))
+        return PriceRead.series(points, product,
+                                str(rec.get("provider") or "price_history"),
+                                stamp)
+
+    # -- observability -------------------------------------------------------
+    @property
+    def words(self) -> str:
+        """One line for the status page. Says what is held, or why nothing is.
+
+        ⚠️ THIS MUST NOT RAISE, EVER. It is read once from `Controller.__init__`
+        and again on every status render, and neither call site guards it - so
+        an exception here is an add-on that does not start. It went out that
+        way: a day record that was a string instead of a dict raised
+        `AttributeError` from the `len(...)` below. `_load()` now drops such a
+        record and `_n_points()` is total, so this has two independent reasons
+        not to fail on a broken file.
+        """
+        if not self.ok:
+            return f"unavailable - {self.last_error}"
+        if not self._days:
+            why = f" - {self.last_error}" if self.last_error else ""
+            return f"empty{why} - nothing to carry forward yet"
+        held = "; ".join(f"{d} ({_n_points(self._days[d])} h)"
+                         for d in sorted(self._days, key=str))
+        # A file that PARTLY loaded says so, or the dropped days read as days
+        # that were simply never written.
+        return held + (f" [{self.last_error}]" if self.last_error else "")
+
+
+def _parse_points(raw) -> tuple:
+    """Canonical dicts -> PricePoints, and a phrase naming what was dropped.
+
+    ⚠️ A MALFORMED POINT IS DROPPED, NOT DEFAULTED, and the drop is reported.
+    `float(row.get("eur_per_kwh", 0))` is the shape of the whole family of
+    defects this repo has been removing: a cache that lost a field would answer
+    "that hour was free" to a fill that flags it as an ordinary guess.
+    """
+    if not isinstance(raw, list):
+        return (), "the `points` field is not a list"
+    out, dropped = [], 0
+    for row in raw:
+        if not isinstance(row, dict):
+            dropped += 1
+            continue
+        try:
+            start = datetime.fromisoformat(row["start"])
+            value = float(row["eur_per_kwh"])
+            flag = row["provisional"]
+        except (KeyError, TypeError, ValueError):
+            dropped += 1
+            continue
+        if start.tzinfo is None or start.utcoffset() is None \
+                or not isinstance(flag, bool):
+            dropped += 1
+            continue
+        out.append(PricePoint(start=start, eur_per_kwh=value, provisional=flag))
+    out.sort(key=lambda p: p.start.astimezone(timezone.utc))
+    return tuple(out), (f"{dropped} malformed point(s)" if dropped else "")
+
+
+class NoHistory:
+    """The explicit "this caller keeps no history" - a NAME, not a `None`.
+
+    ⚠️ IT EXISTS SO THE PARAMETER CAN BE REQUIRED. `history=None` meaning "skip
+    the fill" is the shape this repo keeps removing: a caller who forgets the
+    argument gets the truncated-horizon behaviour of the bug this ticket is
+    about, silently, and the code reads as though the fill is happening. With
+    no default there is nothing to forget, and a caller with no durable volume
+    - a check, a replay harness - has to say `NoHistory()` and be visible in
+    the diff.
+
+    It carries out the ordering faithfully: it records nothing and carries
+    nothing, so `carry()` returns the provider's read unchanged and the plan is
+    exactly what it was before FCST-04.
+    """
+
+    ok = False
+    last_error = "no history is kept by this caller"
+    words = "not kept"
+
+    def carry(self, read: PriceRead, *, since, until, stamp) -> PriceRead:
+        return read
+
+    def yesterday(self, stamp: datetime) -> PriceRead:
+        return PriceRead.no_data(
+            "this caller keeps no price history (NoHistory), so §8.3 has no "
+            "profile to carry forward")
@@ -1019,7 +1019,7 @@ def static_status(opts: dict, now: datetime):


 async def resolve_tariff(opts: dict, since: datetime, until: datetime,
-                         transport=None, now=None) -> TariffPlan:
+                         transport=None, now=None, *, history) -> TariffPlan:
    """The full §9.2 + §11.4 resolution. MAY PERFORM I/O - see `static_plan`.

    ⚠️ NOT ON THE 30 s TICK, and that is a deliberate boundary rather than an
@@ -1033,9 +1033,42 @@ async def resolve_tariff(opts: dict, since: datetime, until: datetime,
      price_provider off  + profile  -> the CONFIGURED product (§9.2 rows 3-4)
      price_provider off  + none     -> no prices at all (§11.4, nothing to
                                        fall back to)
-      provider series                -> the feed (§9.2 rows 1-2)
+      provider series                -> the feed, §8.3-FILLED (§9.2 rows 1-2)
      provider no_data    + profile  -> §11.4 FALLBACK, provisional, no arbitrage
      provider no_data    + none     -> no prices at all
+
+    ⚠️ `history` IS REQUIRED, KEYWORD-ONLY, AND HAS NO DEFAULT (FCST-04). It is
+    the §8.3 carry-forward store - `price_history.PriceHistory`, or the explicit
+    `price_history.NoHistory()` for a caller that keeps none. A default of
+    `None` meaning "skip the fill" would give a caller who forgot the argument
+    the TRUNCATED HORIZON this ticket exists to remove, silently, while the code
+    read as though the fill were happening. There is nothing to forget now.
+
+    ⚠️ AND THE FILL RUNS ONLY ON A `series` READ. A feed that published nothing
+    still falls to `_degrade()` and §11.4's own row - the static profile the
+    operator configured FOR that case - rather than to yesterday's feed. The two
+    are indistinguishable to the optimiser (both are 100 % provisional, so
+    `arbitrage()` finds fewer than two firm hours and refuses, and `cheap_now()`
+    is False), so preferring yesterday would change a documented degradation for
+    a difference only the status page can see. See price_history.py's closing
+    note and DECISIONS/FCST-04-the-day-that-is-carried.md.
+
+    ⚠️ THE FILL IS PER WHOLE HOUR AND ITS COVERAGE IS CONDITIONAL. Every point
+    the provider published inside the window passes through BY IDENTITY - four
+    of them in one hour on a quarter-hourly product like `eneco_eddy` - and a
+    carried point is added only for a whole hour the feed did not print at all
+    AND yesterday did. An hour neither source can price is left out, which is
+    §11.4 for that hour: `plan.price_at()` raises rather than answering, and an
+    hour with no price is not a free one.
+
+    ⚠️ THE FILLED SERIES IS WINDOWED TO [since, until), AND THAT IS A CHANGE.
+    `load_prices()` returns whatever the feed sent, past hours included, and
+    `TariffPlan.arbitrage()` took its spread over all of it - so an hour that
+    had already gone could be the cheapest hour of a cycle nothing can now buy
+    in. `fill_from_yesterday()` walks the requested horizon and emits exactly
+    that, so the spread is over hours that are still ahead. Both directions of
+    the difference are real: the window gains the unpublished hours it was
+    missing and loses the elapsed ones it should never have had.
    """
    stamp = now or datetime.now(timezone.utc)
    pid = str(price_option(opts, "price_provider") or "off").strip()
@@ -1044,6 +1077,20 @@ async def resolve_tariff(opts: dict, since: datetime, until: datetime,

    read = await load_prices(opts, since, until, transport)
    return read.match(
+        series=lambda r: _provider_plan(opts, r, since, until, stamp, history),
+        no_data=lambda r: _degrade(opts, r, since, until, stamp))
+
+
+def _provider_plan(opts, read, since, until, stamp, history) -> TariffPlan:
+    """§8.3: record what the feed published, carry yesterday into the gaps.
+
+    The fill can itself return `no_data` - a feed whose points all fall outside
+    the horizon, with nothing in the history to fill it from - and that is the
+    §11.4 row arrived at one step later, so it degrades exactly as a failed
+    fetch does rather than being special-cased here.
+    """
+    filled = history.carry(read, since=since, until=until, stamp=stamp)
+    return filled.match(
        series=lambda r: TariffPlan(r, _quiet_profile(opts), "provider"),
        no_data=lambda r: _degrade(opts, r, since, until, stamp))

@@ -253,7 +253,7 @@ class MigrationReport:
        return out


-def migrate_unit_keys(store, unit: str) -> MigrationReport:
+def migrate_unit_keys(store, unit: str, roster=None) -> MigrationReport:
    """Move the pre-DEV-02 flat per-inverter keys into `unit`'s namespace.

    `unit` is the FIRST inverter in the roster, which on every install that can
@@ -261,6 +261,24 @@ def migrate_unit_keys(store, unit: str) -> MigrationReport:
    version that knew about more than one inverter is already at
    SCHEMA_VERSION and is not touched.

+    ⚠️ `roster` IS THE NAMES THE MACHINES READ, AND IT IS SAFETY-44. This
+    function shipped WITHOUT it, taking a hardcoded `inverter.UNIT_A` at its one
+    call site - which is, exactly, the MAINT-05 defect `_verify_moves()` spends
+    a paragraph describing, still live in the sibling that paragraph was written
+    beside. On a store that ran pre-DEV-02 flat and declares
+    `inverters: [garage, shed]`, the latch moved to `unit/inverter-a/
+    sign_tripped`, was read back under that same name, was SATISFIED, and
+    stamped the marker - so the next start skipped the migration too and NO
+    inverter came up latched on an install whose state file says `sign_tripped:
+    true`. A disarmed SAFETY-02 guard, every symptom of a clean migration.
+
+    `roster_names(opts)` answers this BEFORE any `Inverter` exists - it is
+    defined to, for `MqttPublisher`, which is constructed before the roster for
+    the same reason - so the "there is no roster yet" that kept this argument
+    out was never true. On the flat shape it returns `[UNIT_A]`, so the upgrade
+    every install runs today passes `unit == roster[0] == UNIT_A` and is
+    byte-identical.
+
    THE THREE RULES, each one a case that would otherwise be a wrong answer:

    1. IT NEVER OVERWRITES AN EXISTING NAMESPACED VALUE. A store that has both
@@ -302,7 +320,7 @@ def migrate_unit_keys(store, unit: str) -> MigrationReport:

    rep.ran = True
    _move_flat_keys(store, unit, LEGACY_UNIT_KEYS, rep)
-    _verify_moves(store, rep)
+    _verify_moves(store, rep, roster)

    if rep.ok:
        store.set(SCHEMA_KEY, SCHEMA_VERSION)
@@ -383,12 +401,26 @@ def _verify_moves(store, rep: MigrationReport, roster=None) -> None:
    itself. `test_maintenance.py` section 9c calls this twice with the SAME
    wrong `unit` over one store, which is the shape nothing in the suite had.

-    `roster=None` means "not graded", and the one caller that passes it is
-    `migrate_unit_keys()`, which runs BEFORE `build_roster()` for the reason its
-    own docstring gives - the `Inverter` objects restore their latch in their
-    constructors, so the keys have to be in place before there is a roster to
-    check against. That gap is real and is recorded as a running check rather
-    than left to a comment: see test_maintenance.py section 13.
+    `roster=None` still means "not graded", and SAFETY-44 is what it cost. This
+    paragraph used to end here, saying that the one caller passing it was
+    `migrate_unit_keys()`, "which runs BEFORE `build_roster()` ... so the keys
+    have to be in place before there is a roster to check against", and calling
+    that gap real. THE FIRST HALF IS TRUE AND THE CONCLUSION IS NOT. The keys do
+    have to be written before the `Inverter`s restore from them - but the NAMES
+    do not come from the `Inverter`s. `roster_names(opts)` reads them straight
+    off the options, before anything is constructed, which is the whole reason
+    it exists (`MqttPublisher` is built before the roster too). So the argument
+    that could have graded that write was available at the call site the entire
+    time, and the sibling function three screens down had already been fixed by
+    passing it.
+
+    ⚠️ SO NO CALLER PASSES `None` TODAY, and this branch is kept for one reason:
+    it is what a NEW caller gets by DEFAULT. A future migration that forgets the
+    roster gets the ungraded read-back and stamps its marker over a namespace no
+    machine reads - silently, exactly as this one did. That is the shape of both
+    defects, and the gap is recorded as a running check rather than left to a
+    comment: test_maintenance.py section 13, and test_rehearse.py part 5, which
+    drives it end-to-end through the harness an operator actually runs.
    """
    for target, value in rep.moved.items():
        got = getattr(store, "data", {}).get(target, "<<missing>>")
@@ -21,6 +21,7 @@ Read MUTATION.md before you report a survivor.

 import asyncio
 import dataclasses
+import logging
 import re
 import sys
 from datetime import datetime, timedelta, timezone
@@ -34,7 +35,8 @@ from app.inverter import UNIT_A
 from app.main import Controller
 from app.store import (
    DEFAULTS, MAINT_SCHEMA_KEY, MAINT_SCHEMA_VERSION, MAINT_UNIT_KEYS,
-    migrate_maint_keys, unit_key,
+    SCHEMA_KEY, SCHEMA_VERSION, migrate_maint_keys, migrate_unit_keys,
+    unit_key,
 )

 fails = []
@@ -1411,45 +1413,39 @@ check("...while the SAME site with the house importing more than the drain "

 # --------------------------------------------------------------------------- #
 print()
-print("13. RECORDED, NOT FIXED: SAFETY-02's migration has the same "
-      "wrong-namespace hole, and it is a LATCH")
+print("13. FIXED (SAFETY-44): SAFETY-02's migration takes the roster too, and "
+      "the guard is no longer defeated by the key the defect wrote")

-# ⚠️ FOUND WHILE FIXING THIS CARD'S OWN INSTANCE, AND WRITTEN DOWN AS A RUNNING
-# CHECK RATHER THAN A COMMENT SO IT CANNOT BE FORGOTTEN OR QUIETLY "FIXED" BY AN
-# UNRELATED EDIT. It is the SAME defect one module over, and it costs more.
+# ⚠️ THIS SECTION USED TO BE THE SPECIFICATION OF A DEFECT. It was written while
+# fixing MAINT-05's own instance, as a RUNNING check rather than a comment, with
+# the instruction that "the card that fixes it should expect them to go red and
+# should replace them". SAFETY-44 is that card and this is that replacement.
 #
-# WHAT IT IS. `Controller.__init__` calls `migrate_unit_keys(store, UNIT_A)` -
-# a hardcoded constant, where the maintenance migration now passes the roster's
-# own first name. It has to run BEFORE `build_roster()`, because each
-# `Inverter` restores its sign latch in its constructor, so there is no roster
-# to name yet and `_verify_moves()` is passed `roster=None`. On a DECLARED
-# roster the pre-DEV-02 flat `sign_tripped: true` therefore lands under
-# `unit/inverter-a/sign_tripped`, which no unit on this site reads.
+# WHAT IT WAS. `Controller.__init__` called `migrate_unit_keys(store, UNIT_A)` -
+# a hardcoded constant, where the maintenance migration already passed the
+# roster's own first name. On a DECLARED roster the pre-DEV-02 flat
+# `sign_tripped: true` landed under `unit/inverter-a/sign_tripped`, which no
+# unit on the site reads, and `_verify_moves()` read it back under the name it
+# had been handed and was satisfied.
 #
-# WHY THE EXISTING GUARD DOES NOT CATCH IT, WHICH IS THE PART WORTH PINNING.
-# `_grade_store_migration()` re-arms the latch when the flat key says tripped
-# and NOTHING NAMESPACED IS LATCHED. The misdirected write IS namespaced and IS
-# latched, so `namespaced_latched` is True, the guard returns, and no inverter
-# on the roster comes up tripped. The guard is defeated by the very key the
-# defect wrote.
+# AND THE GUARD DID NOT CATCH IT, WHICH WAS THE PART WORTH PINNING.
+# `_grade_store_migration()` re-arms when the flat key says tripped and NOTHING
+# NAMESPACED IS LATCHED - and the misdirected write IS namespaced and IS
+# latched. The guard was defeated by the very key the defect wrote.
 #
-# WHY IT IS NOT FIXED HERE. The remedy is to name the migration's target from
-# the config before any `Inverter` exists, which means factoring
-# `build_roster()`'s naming rule out of `build_roster()` so there is still only
-# ONE spelling of it - inverter.py and SAFETY-02's latch semantics, with its own
-# mutation battery over test_inverter.py and test_watchdog.py. MAINT-05 moves
-# the maintenance history and nothing about the sign latch.
-#
-# ⚠️ AND IT IS A DISARMED SAFETY GUARD, NOT LOST HISTORY. Read the checks below
-# as the SPECIFICATION OF A DEFECT: every one of them is asserting the WRONG
-# behaviour, on purpose, and the card that fixes it should expect them to go
-# red and should replace them.
+# ⚠️ THE REASON IT WAS LEFT - "there is no roster to name yet, because each
+# `Inverter` restores its latch in its constructor" - WAS THE HALF-TRUTH THAT
+# KEPT IT ALIVE. The keys must be written before the `Inverter`s read them; the
+# NAMES do not come from the `Inverter`s. `roster_names(opts)` reads them off
+# the options before anything is constructed, which is what it already existed
+# to do for `MqttPublisher`. The factoring this section said the fix would need
+# had been done a card earlier.
 #
 # ⚠️ NO `schema_version` IN THIS STORE, AND THAT IS THE PRE-DEV-02 SHAPE
 # ITSELF: an install that has never run a version with namespaced keys, which is
 # what makes `migrate_unit_keys()` run at all.
 _lat_store = MemStore({"sign_tripped": True,
-                       "sign_reason": "SIGN_DISAGREEMENT: measured pre-DEV-02"})
+                       "sign_reason": "SIGN_DISAGREEMENT: SYNTHETIC FIXTURE, not measured"})
 _lat, _ = controller(NAMED_OPTS, store=_lat_store)
 print(f"    pre-DEV-02 flat latch + declared roster {_C2_NAMES} -> "
      f"{_lat.store_migration.explain()}")
@@ -1457,29 +1453,179 @@ print(f"    latch keys in the store: "
      f"{sorted(k for k in _lat_store.data if k.endswith('sign_tripped'))}")
 print(f"    roster came up tripped: "
      f"{[(i.name, i.sign_wd.tripped) for i in _lat.inverters]}")
-check("the flat pre-DEV-02 latch is migrated into `unit/inverter-a/` - a name "
-      "that is NOT on this roster, so no unit reads it. This is the same "
-      "wrong-namespace write `migrate_maint_keys()` now refuses, in the "
-      "migration that runs before there is a roster to grade against",
-      _lat_store.data.get(unit_key(UNIT_A, "sign_tripped")) is True
-      and unit_key(_C2_NAMES[0], "sign_tripped") not in _lat_store.data)
-check("⚠️ AND NO INVERTER ON THE ROSTER COMES UP LATCHED, on a store whose file "
-      "still reads `sign_tripped: true`. That is DEV-02's own sentence for the "
-      "worst case - a disarmed safety guard on an install that believes it is "
-      "protected",
-      not any(i.sign_wd.tripped for i in _lat.inverters)
+check("⭐ the flat pre-DEV-02 latch is migrated into the ROSTER'S OWN first "
+      "namespace and NOT into `unit/inverter-a/` - the same wrong-namespace "
+      "write `migrate_maint_keys()` refuses, now refused in the migration that "
+      "runs before the `Inverter` objects exist",
+      _lat_store.data.get(unit_key(_C2_NAMES[0], "sign_tripped")) is True
+      and unit_key(UNIT_A, "sign_tripped") not in _lat_store.data)
+check("⭐ AND AN INVERTER ON THE ROSTER COMES UP LATCHED, which is the whole "
+      "card: a store whose file reads `sign_tripped: true` no longer produces "
+      "a site that drives",
+      _lat.inverters[0].sign_wd.tripped
      and _lat_store.data["sign_tripped"] is True)
-check("...and `_grade_store_migration()` is DEFEATED BY THE KEY THE DEFECT "
-      "WROTE, which is why this needs a card rather than a tweak: its "
-      "'is anything namespaced latched' test is satisfied by the misdirected "
-      "value itself, so it returns instead of re-arming",
-      any(k.startswith("unit/") and k.endswith("/sign_tripped") and v
-          for k, v in _lat_store.data.items())
-      and not any(i.sign_wd.tripped for i in _lat.inverters))
+check("...and the latch that came back carries the REASON, not just the "
+      "boolean - an operator meeting a FAILSAFE banner with no sentence under "
+      "it has been told the site stopped and not why",
+      "SIGN_DISAGREEMENT" in _lat.inverters[0].sign_wd.reason)
+check("⚠️ ...and the OTHER units on the roster are NOT latched by it. The flat "
+      "key is one site-level boolean and re-arming every unit off it would be "
+      "DEV-02's defect in the opposite direction - a healthy pack parked "
+      "because its sibling once disagreed",
+      not any(i.sign_wd.tripped for i in _lat.inverters[1:]))
+
+# ⚠️ AND THE OTHER PATH, WHICH THE MIGRATION FIX ALONE DOES NOT REACH: a store
+# ALREADY STAMPED by the shipped version. `migrate_unit_keys()` returns at its
+# `from_version >= SCHEMA_VERSION` check and never moves anything, so the
+# stranded `unit/inverter-a/sign_tripped` just sits there. This is the STAGED
+# shape - "upgrade now, declare the roster later" - and it is the one an
+# operator actually produces. Only the GUARD repairs it.
+_staged_store = MemStore({
+    "sign_tripped": True,
+    "sign_reason": "SIGN_DISAGREEMENT: SYNTHETIC FIXTURE, not measured",
+    unit_key(UNIT_A, "sign_tripped"): True,
+    unit_key(UNIT_A, "sign_reason"): "SIGN_DISAGREEMENT: SYNTHETIC FIXTURE, not measured",
+    SCHEMA_KEY: SCHEMA_VERSION})
+_staged, _ = controller(NAMED_OPTS, store=_staged_store)
+print(f"    already-stamped store + declared roster -> ran="
+      f"{_staged.store_migration.ran}, came up tripped: "
+      f"{[(i.name, i.sign_wd.tripped) for i in _staged.inverters]}")
+check("⚠️ PRECONDITION: the migration really does NOT run on this store - the "
+      "marker is already there - so anything below is the guard's doing and "
+      "not the migration's",
+      not _staged.store_migration.ran)
+check("⭐ and the roster's first unit STILL comes up latched: the guard no "
+      "longer counts a latch stranded under `UNIT_A`, which is not on this "
+      "roster, as protection an operator has",
+      _staged.inverters[0].sign_wd.tripped)
+check("...and the re-armed latch is written where a machine reads it, so the "
+      "NEXT start finds it without needing the guard again",
+      _staged_store.data.get(unit_key(_C2_NAMES[0], "sign_tripped")) is True)
+
+# ⚠️ THE NEGATIVE CONTROL FOR THAT NARROWING, because "off the roster" on its
+# own was too blunt and test_inverter.py's DEV-02 restart case caught it: a
+# latch stranded under a name that is NOT `UNIT_A` is a real unit out of
+# today's config, and it MUST still count as protection.
+_other_store = MemStore({
+    "sign_tripped": True,
+    "sign_reason": "SIGN_DISAGREEMENT: B disagreed, mirrored flat",
+    unit_key("inverter-b", "sign_tripped"): True,
+    unit_key("inverter-b", "sign_reason"): "SIGN_DISAGREEMENT: B disagreed",
+    SCHEMA_KEY: SCHEMA_VERSION})
+_other, _ = controller(NAMED_OPTS, store=_other_store)
+print(f"    stranded latch under `inverter-b` -> came up tripped: "
+      f"{[(i.name, i.sign_wd.tripped) for i in _other.inverters]}")
+check("⚠️ a latch stranded under a name that is not `UNIT_A` is NOT re-armed "
+      "onto the roster: it belongs to a unit the operator has taken out of "
+      "the config, the flat key is its own mirror, and re-arming a healthy "
+      "sibling off it is the defect DEV-02 removed",
+      not any(i.sign_wd.tripped for i in _other.inverters))
+
+# ⭐ AND THE ARRANGEMENT NEITHER FIXTURE ABOVE HOLDS: **BOTH** KEYS AT ONCE.
+# ⚠️ THIS IS INSTANCE #17 OF HABIT 4 AND IT WAS FOUND IN THIS CARD'S OWN FIX.
+# The first version of the guard classified each key and then wrote ONE shared
+# `namespaced_latched` flag for both "on the roster" and "off-roster but not
+# UNIT_A". `_staged_store` above holds only an `inverter-a` latch and
+# `_other_store` below holds only an `inverter-b` one, so the flag was never
+# written twice with opposite intents in a single store - and a per-key check
+# whose keys have never disagreed WITHIN ONE STORE is a check over whichever
+# key happened to come last. Measured on that build, one added key:
+#
+#   sign_tripped + unit/inverter-a/sign_tripped   -> garage latched, repaired
+#   ...AND unit/old-shed/sign_tripped             -> ⚠️ NOTHING latched
+#
+# A site DRIVING on a state file that reads `sign_tripped: true` - the sentence
+# this card exists to make false - reached by an operator taking one unit out of
+# `inverters:` after it had tripped. "A roster is configuration and it changes"
+# is this module's own phrase, and the second inverter goes in this month.
+_both_store = MemStore({
+    "sign_tripped": True,
+    "sign_reason": "SIGN_DISAGREEMENT: SYNTHETIC FIXTURE, not measured",
+    unit_key(UNIT_A, "sign_tripped"): True,
+    unit_key(UNIT_A, "sign_reason"): "SIGN_DISAGREEMENT: SYNTHETIC FIXTURE, not measured",
+    unit_key("old-shed", "sign_tripped"): True,
+    unit_key("old-shed", "sign_reason"): "SIGN_DISAGREEMENT: old-shed, since removed",
+    SCHEMA_KEY: SCHEMA_VERSION})
+_both, _ = controller(NAMED_OPTS, store=_both_store)
+print(f"    stranded `inverter-a` AND an unrelated off-roster `old-shed` -> "
+      f"{[(i.name, i.sign_wd.tripped) for i in _both.inverters]}")
+check("⭐ A STRANDED `inverter-a` LATCH IS STILL REPAIRED WHEN AN UNRELATED "
+      "OFF-ROSTER LATCH SITS BESIDE IT. One shared flag let `old-shed` veto the "
+      "repair, and the site came up driving on a file that says it is latched",
+      _both.inverters[0].sign_wd.tripped)
+check("...and the repair still lands on the roster's first unit, not on the "
+      "removed one",
+      _both_store.data.get(unit_key(_C2_NAMES[0], "sign_tripped")) is True)
+check("⚠️ ...and the reason carried across is the FLAT one - the misdirected "
+      "latch's own - and not `old-shed`'s, which belongs to a different "
+      "physical inverter",
+      "old-shed" not in _both.inverters[0].sign_wd.reason)
+
+# ⚠️ AND THE OPERATOR MESSAGE IS ASSERTED, because review deleted it whole
+# (`if stranded:` -> `if False:`) and 28/28 stayed green. `stranded` has no
+# other consumer, so the list existed only to print a line nothing checked -
+# and the line is the entire operator-facing half of this repair: it names the
+# key that was stranded and the roster it was measured against. A repair an
+# operator cannot see is a repair they cannot confirm.
+_log_store = MemStore({
+    "sign_tripped": True,
+    "sign_reason": "SIGN_DISAGREEMENT: SYNTHETIC FIXTURE, not measured",
+    unit_key(UNIT_A, "sign_tripped"): True,
+    unit_key(UNIT_A, "sign_reason"): "SIGN_DISAGREEMENT: SYNTHETIC FIXTURE, not measured",
+    SCHEMA_KEY: SCHEMA_VERSION})
+_records = []
+_h = logging.Handler()
+_h.emit = lambda r: _records.append(r.getMessage())
+logging.getLogger("goodwe").addHandler(_h)
+try:
+    _logged, _ = controller(NAMED_OPTS, store=_log_store)
+finally:
+    logging.getLogger("goodwe").removeHandler(_h)
+_s44_lines = [m for m in _records if "SAFETY-44" in m]
+print(f"    SAFETY-44 operator lines logged: {len(_s44_lines)}")
+check("⭐ the repair SAYS SO, at ERROR, naming SAFETY-44 - deleting this line "
+      "left the whole suite green, and it is how an operator learns their "
+      "latch was re-armed rather than never lost",
+      len(_s44_lines) == 1)
+check("...and it names the STRANDED KEY, so the operator can find it in the "
+      "state file rather than being told something happened",
+      any(unit_key(UNIT_A, "sign_tripped") in m for m in _s44_lines))
+check("...and the ROSTER it was measured against, which is the half that says "
+      "WHY no machine read it",
+      any("garage" in m and "shed" in m for m in _s44_lines))
+
+# ⭐ AND `misdirected` IS MEASURED AT THE FUNCTION'S OWN LEVEL, BECAUSE IT IS
+# UNREACHABLE FROM THE CALL SITE. `Controller.__init__` passes
+# `roster_names(opts)[0]`, which `build_roster()` also names its first entry
+# from, so `unit` is a member of `roster` by construction and the refusal can
+# never fire there. Review measured that: dropping the `roster` argument from
+# `_verify_moves()` inside `migrate_unit_keys()` survives 28/28. The argument is
+# kept for the reason `_verify_moves()` gives for its own `roster=None` branch -
+# it is what a NEW caller gets by default - and a guard nothing exercises is a
+# guard nobody knows works, so it is exercised HERE rather than claimed.
+_mis_store = MemStore({"sign_tripped": True,
+                       "sign_reason": "SIGN_DISAGREEMENT: SYNTHETIC FIXTURE, not measured"})
+_mis = migrate_unit_keys(_mis_store, UNIT_A, ["garage", "shed"])
+print(f"    migrate_unit_keys(store, 'inverter-a', ['garage','shed']) -> "
+      f"ok={_mis.ok} misdirected={sorted(_mis.misdirected)}")
+check("⭐ a WRONG `unit` - one that is not on the roster it was handed - comes "
+      "back `misdirected`, which is the refusal `migrate_maint_keys()` already "
+      "had and this one did not",
+      bool(_mis.misdirected) and not _mis.ok)
+check("...and the schema marker is NOT stamped, so the migration runs again on "
+      "the next start rather than being recorded as finished - the one thing "
+      "store.py's header says must never happen",
+      _mis_store.data.get(SCHEMA_KEY) != SCHEMA_VERSION)
+check("⚠️ ...while the SAME call with a unit that IS on the roster stamps the "
+      "marker - so the refusal is caused by the name and not by the fixture",
+      migrate_unit_keys(
+          MemStore({"sign_tripped": True, "sign_reason": "x"}),
+          "garage", ["garage", "shed"]).ok)
+
 # The counterfactual, and it is the accepted baseline habit 7 asks for: the
 # same store, the same migration, one thing different - the roster shape.
 _flat_lat_store = MemStore({"sign_tripped": True,
-                            "sign_reason": "SIGN_DISAGREEMENT: measured pre-DEV-02"})
+                            "sign_reason": "SIGN_DISAGREEMENT: SYNTHETIC FIXTURE, not measured"})
 _flat_lat, _ = controller(FLAT_OPTS, store=_flat_lat_store)
 print(f"    the same store on the FLAT config -> roster "
      f"{[(i.name, i.sign_wd.tripped) for i in _flat_lat.inverters]}")
@@ -272,16 +272,19 @@ check("the provenance audit REFUSES a carried hour flagged non-provisional - "
      "the 'stale but plausible' point, which is SAFETY-18's frozen-but-fresh "
      "meter one layer up",
      raises(lambda: _fc._audit([_stale_copy], {}), ValueError))
+# ⚠️ `published` MAPS AN HOUR TO A TUPLE, not to a point (FCST-04 H1). An hour
+# may legitimately hold four published points - see the quarter-hourly block
+# below - so the audit takes the whole set and checks membership by identity.
 check("...and refuses a published hour that arrived as a COPY rather than as "
      "the published object, which is where a flag gets lost",
      raises(lambda: _fc._audit(
          [PricePoint(start=TODAY_PART.points[0].start,
                      eur_per_kwh=TODAY_PART.points[0].eur_per_kwh,
                      provisional=TODAY_PART.points[0].provisional)],
-          {D0: TODAY_PART.points[0]}), ValueError))
+          {D0: (TODAY_PART.points[0],)}), ValueError))
 check("...and never repairs a flag: the audit's only outcomes are silence and "
      "an exception, so a series it passed was already correct",
-      _fc._audit(list(FILLED.points), {u: TODAY_PART.points[i]
+      _fc._audit(list(FILLED.points), {u: (TODAY_PART.points[i],)
                                       for i, u in enumerate(DAY[:PUBLISHED_N])})
      is None)

@@ -298,6 +301,81 @@ check("a PUBLISHED point that the feed itself flagged provisional keeps that "
      and all(p.provisional for p in _eddy_filled.points[:2])
      and _eddy_filled.points[0] is _eddy.points[0])

+# ⚠️ AND A SUB-HOURLY FEED IS NOT DECIMATED TO ONE POINT AN HOUR (FCST-04 H1).
+# `eneco_eddy` settles per QUARTER and `_validate_raw()` accepts any
+# strictly-increasing spacing, so a published hour can hold four points. The
+# fill walks WHOLE HOURS to decide what needs carrying; that must not become a
+# rule about what a published hour may contain. Measured on FCST-04's first
+# draft, which keyed the published set `{_utc_hour(p.start): p ...}`: a
+# 96-point feed came out 15 points long and its arbitrage spread fell from
+# 0.143 to 0.014 EUR/kWh, on an ALL-FIRM, fully-published series.
+_QN = 16                                             # four hours of quarters
+_QUARTERS = PriceRead.series(
+    [PricePoint(start=brussels(D0 + timedelta(minutes=15 * n)),
+                eur_per_kwh=round(0.100000 + 0.011300 * n, 6),
+                provisional=False)
+     for n in range(_QN)],
+    PriceProduct.REALTIME, "eneco_eddy", D0)
+_QHOURS = {_fc._utc_hour(p.start) for p in _QUARTERS.points}
+check(f"the quarter-hourly fixture really is sub-hourly - {_QN} points across "
+      f"{len(_QHOURS)} hours - or every check below passes for an hourly feed "
+      f"too and grades nothing",
+      len(_QUARTERS.points) == _QN and len(_QHOURS) == _QN // 4)
+_qfill = derived(lambda: fill_from_yesterday(_QUARTERS, yesterday=YESTERDAY,
+                                             since=D0, until=D1, stamp=D0))
+_qfirm = derived(lambda: [p for p in _qfill.points if not p.provisional])
+check(f"every one of the {_QN} published points survives the fill BY IDENTITY "
+      f"({len(_qfirm)} kept). One-per-hour keying would leave {_QN // 4}, and "
+      f"the spread over the hours the feed actually published would collapse "
+      f"with them",
+      len(_qfirm) == _QN
+      and all(any(p is q for q in _QUARTERS.points) for p in _qfirm))
+check(f"...and the {len(DAY) - len(_QHOURS)} whole hours it published NOTHING "
+      f"for are carried, ONE point each: a carried hour is hourly whatever the "
+      f"feed's spacing is, because yesterday's profile is only known per "
+      f"wall-clock hour",
+      _qfill.kind == "series"
+      and len([p for p in _qfill.points if p.provisional])
+      == len(DAY) - len(_QHOURS))
+check("...and the filled series is still ordered by instant across the two "
+      "kinds, so `at()` and `window()` are not reading a shuffled list",
+      _qfill.kind == "series"
+      and [p.start.astimezone(UTC) for p in _qfill.points]
+      == sorted(p.start.astimezone(UTC) for p in _qfill.points))
+
+_PUB_Q = _fc._published_by_hour(_QUARTERS)
+# ⚠️ `derived`, NOT A RAW CALL. Under the decimating mutation this check
+# exists for, `_audit` RAISES here - and a harness that dies on the mutation it
+# is grading reports a traceback instead of a FAIL, which MUTATION.md forbids.
+check("the audit ACCEPTS four published points inside one hour, which is the "
+      "case the repeated-hour COUNT it replaced could not express",
+      derived(lambda: _fc._audit(list(_QUARTERS.points), _PUB_Q)) is None)
+check("...and still REFUSES a repeated CARRIED hour, which is the DST "
+      "mispricing that count existed to catch",
+      raises(lambda: _fc._audit([_fc._carried(D0, 0.11),
+                                 _fc._carried(D0, 0.12)], {}), ValueError))
+check("...and refuses the SAME published point emitted twice - the provider's "
+      "points are strictly increasing, so a repeat is never a second print",
+      raises(lambda: _fc._audit([_QUARTERS.points[0], _QUARTERS.points[0]],
+                                _PUB_Q), ValueError))
+check("...and refuses a CARRIED point standing in an hour the feed DID "
+      "publish, which the old count never looked at",
+      raises(lambda: _fc._audit([_fc._carried(D0, 0.11)], _PUB_Q), ValueError))
+
+# ⚠️ THE HORIZON IS WHOLE HOURS AT BOTH ENDS (FCST-04 L8). `since` is floored
+# and `until` was not, so a 48 h request made at any instant off the hour came
+# back 49 hours long. Graded at an OFF-THE-HOUR instant, because on the hour
+# the two implementations agree and the check would grade nothing.
+_off = D0 + timedelta(hours=8, minutes=17, seconds=43)
+for _h in (1, 12, 48):
+    _n = derived(lambda h=_h: len(fill_from_yesterday(
+        TODAY_PART, yesterday=YESTERDAY, since=_off,
+        until=_off + timedelta(hours=h), stamp=_off).points))
+    check(f"a {_h} h horizon asked for at {_off.isoformat()} - 17 min 43 s "
+          f"past the hour, the shape a running controller always has - is "
+          f"{_h} points long, not {_h + 1}",
+          _n == _h)
+
 # The structural half: the constructor has nothing to pass, and the file is
 # scanned so the pass-through idiom cannot come back looking ordinary.
 check("_carried() has NO `provisional` parameter, so a caller has nothing to "
@@ -69,6 +69,7 @@ from app.price import (
    Absence, PricePoint, PriceProduct, PriceRead, arbitrage_allowed, brussels,
    plan_is_provisional, required_spread, uncertainty_of,
 )
+from app.price_history import NoHistory
 from app.static_tariff import (
    MEAN_WINDOW_DAYS, ProfileInvalid, StaticProfile, TariffPlan,
    parse_static_profile, resolve_tariff, static_plan, static_status,
@@ -746,7 +747,8 @@ _cfg = opts(TWO_PERIOD, price_provider="generic_dayahead",
            price_feed_url="https://api.energy-charts.info/price?bzn=BE",
            price_timeout_s=1, price_markup_eur_kwh=0.03,
            price_taxes_eur_kwh=0.06)
-_fb = run(resolve_tariff(_cfg, THU, THU + timedelta(days=1), transport=_dead))
+_fb = run(resolve_tariff(_cfg, THU, THU + timedelta(days=1), transport=_dead,
+                         history=NoHistory()))
 check("a dead feed falls back to the static profile rather than to nothing",
      _fb.origin == "fallback" and _fb.has_prices)
 check("...with the SAME prices the configured-product path produces, because "
@@ -769,7 +771,8 @@ check("...and the provisional plan has no cheap window either: §8.3 forbids an
      not any(_fb.cheap_now(at_local(THU, h)) for h in range(24)))
 _fb_none = run(resolve_tariff(opts(None, price_provider="generic_dayahead",
                                   price_timeout_s=1),
-                              THU, THU + timedelta(days=1), transport=_dead))
+                              THU, THU + timedelta(days=1), transport=_dead,
+                              history=NoHistory()))
 check("a dead feed with NO static_profile has nothing to fall back to, and "
      "says that rather than inventing a tariff",
      _fb_none.origin == "none" and not _fb_none.has_prices)
@@ -785,7 +788,7 @@ async def _live(url, timeout_s):


 _live_plan = run(resolve_tariff(_cfg, THU, THU + timedelta(days=1),
-                                transport=_live))
+                                transport=_live, history=NoHistory()))
 check("a live feed is used in preference to the static profile",
      _live_plan.origin == "provider"
      and _live_plan.product is PriceProduct.DAY_AHEAD)
@@ -722,18 +722,41 @@ print(f"       e.g. {_sustained.reason!r}")

 print()
 print("   ⚠️ AND NOT ON THE TRANSIENT, WHICH IS THE FALSE-ALARM CASE")
-# ⚠️ THE SHAPES BELOW ARE THE RIG'S, NOT INVENTED. Measured on TEST-19 over 924
-# closed-loop runs in which every unit executed the share it was given: the
-# worst continuous opposition above 700 W is 5.0 s - ONE control cycle - at a
-# peak of 1700 W. That is the serialised write path: on a reversing cycle B's
-# new number lands up to a whole cycle after A's, both slew through zero at
-# 800 W/s behind a 10 s lag, and between the two crossings they genuinely
-# oppose. A guard that trips on that is a false-alarm generator, and this
-# project's doctrine on crying wolf is explicit (SAFETY-15, SAFETY-19).
+# THE SHAPE BELOW IS THE RIG'S: the serialised write path means that on a
+# reversing cycle B's new number lands up to a whole control cycle after A's,
+# both slew through zero behind a real ramp and a real lag, and between the two
+# crossings they genuinely oppose. A guard that trips on that is a false-alarm
+# generator, and this project's doctrine on crying wolf is explicit (SAFETY-15,
+# SAFETY-19).
+#
+# ⚠️ THE MAGNITUDE IS A FIXTURE INPUT, NOT A MEASUREMENT, AND THIS COMMENT USED
+# TO SAY OTHERWISE. It read "the worst single transient the rig can produce -
+# 1700 W", sourced to a 924-run campaign whose harness IS NOT COMMITTED IN
+# EITHER REPOSITORY (DECISIONS/SAFETY-31-circulating-power.md §3 says so in its
+# own words). Three sweeps have now been run against that claim and no two agree:
+#
+#     924-run campaign, harness uncommitted     peak 1700 W, 505.0 s at 100 W
+#     independent reviewer sweep, 2026-09-07    peak 2750 W,  15.0 s above 700 W
+#     goodwe `two_master_circulation`, 840 runs peak 1048 W,   5.0 s above 700 W
+#
+# ⚠️ THAT IS NOT EVIDENCE THAT ANY OF THEM IS WRONG. The three swept different
+# grids, and the last one measured WHY the spread is so wide: on its own grid the
+# worst legitimate peak is 1048 W on the modelled plant, 2250 W on `INSTANT`, and
+# 0 W on `INSTANT` with no write stagger reaching a whole cycle. The plant model
+# and the stagger set are TERMS of the number, and a worst case quoted without
+# them is not a measurement. It is evidence that the claim cannot be adjudicated
+# while the instrument is missing.
+#
+# So what this check now says is what it can support: 1700 W is LARGER than
+# anything any of the three sweeps produced legitimately, and one control cycle
+# of it does not trip. The committed, re-runnable version of the claim is
+# `admin/goodwe`'s `sim/inject.py two_master_circulation`, which runs on every
+# pull request there and requires ZERO of ~840 non-faulty runs to trip.
 _one_cycle = [0.0] * 5 + [1700.0] + [0.0] * 30
 _t1, _ = circ_drive(_one_cycle)
-check("the worst single transient the rig can produce - 1700 W for ONE control "
-      "cycle - does not trip", not _t1.tripped)
+check("a single transient of 1700 W - larger than the worst peak any committed "
+      "sweep of the legitimate family has produced - lasting ONE control cycle "
+      "does not trip", not _t1.tripped)
 _repeated = ([0.0] * 3 + [1700.0]) * 20
 _t2, _ = circ_drive(_repeated)
 check("...and neither does it REPEATED every fourth cycle for 400 s, which is "
@@ -1091,6 +1114,158 @@ check("...and it does not raise the same card on every cycle - a per-PROCESS "
      "this' a different question from 'is this site tripped'",
      len(_notes2) == 1)

+
+print()
+print("6a. AC 4: THE ALARM NAMES WHICH PACKS, AND WHICH WAY ROUND")
+# ⚠️ THIS BLOCK EXISTS BECAUSE A MUTATION SURVIVED ALL 26 FILES. Replacing
+#
+#     pos, neg = opposing([inv.batt for inv in self.inverters],
+#                         min_w=self.circ_wd.cfg.min_w / 2.0)
+#
+# in `check_circulating()` with `pos, neg = ([], [])` (-102 B) silently deletes
+# the "Right now X is discharging while Y charges" clause from the notification,
+# and the whole suite stayed GREEN. The block above asserts three tokens -
+# METER CANNOT SEE, CONTROL HAS BEEN STOPPED, isolator - and NEVER an inverter
+# name, so the one sentence `circulation.py` calls "the first thing somebody
+# woken by this alarm needs" and "impossible to reconstruct from the number
+# alone" was graded by nothing at all.
+#
+# ⚠️ IT IS NOT DEAD CODE, WHICH IS THE PART THAT MAKES IT A CHECK GAP RATHER
+# THAN A REACHABILITY ONE. Ledger M18 - the `await self.check_circulating()`
+# call removed - goes red on four checks above, so this function runs. What
+# nothing asserted was the ATTRIBUTION.
+#
+# ⚠️ AND THE ROSTER IS ["garage", "shed", "cellar"], WHICH IS THE OTHER HALF OF
+# THE FIX. Every fixture above runs on `inverter-a`/`inverter-b`;
+# test_maintenance.py section 11 records what that costs - on a fixture whose
+# names are the module's own constants, a check cannot tell the roster from the
+# constant, and `UNIT_A` is not even what a nameless entry falls back to. Here
+# the names are an operator's, the two participants pull OPPOSITE ways so the
+# sentence is not symmetric, and a third pack rests at 5 W. That makes three
+# DIFFERENT mutations - the clause dropped, the two directions swapped, and the
+# `min_w` floor dropped - into three DIFFERENT red checks. A fixture naming both
+# units the same, or one in which the direction did not matter, would see the
+# first and neither of the others.
+
+_CIRC_ROSTER = [
+    dict(name="garage", setpoint_entity="number.sp_a", soc_entity="sensor.soc_a",
+         batt_entity="sensor.batt_a", max_w=5000),
+    # ⚠️ B_RATING_HYPOTHESIS, and part 4's note above says why it is not 4600:
+    # inverter B does not physically exist and a datasheet is not a measurement.
+    dict(name="shed", setpoint_entity="number.sp_b", soc_entity="sensor.soc_b",
+         batt_entity="sensor.batt_b", max_w=B_RATING_HYPOTHESIS),
+    # THE BYSTANDER. A third pack resting at standby draw, on the roster, read
+    # every cycle, and NOT a participant in anybody's loop.
+    dict(name="cellar", setpoint_entity="number.sp_c", soc_entity="sensor.soc_c",
+         batt_entity="sensor.batt_c", max_w=B_RATING_HYPOTHESIS),
+]
+_CIRC_NAMES = [e["name"] for e in _CIRC_ROSTER]
+_REST_W = 5.0
+
+
+class _NamedCircHass(_Hass):
+    """Three named packs; the meter still reads exactly zero.
+
+    Same house as `_CircHass` - grid = load - sum(B) with two units cancelling -
+    with one extra pack idling at standby draw so that "who is named" and "who
+    is merely on the roster" are different questions.
+    """
+
+    def __init__(self, a_w, b_w, c_w=_REST_W):
+        super().__init__()
+        self.vals = {"sensor.batt_a": a_w, "sensor.batt_b": b_w,
+                     "sensor.batt_c": c_w}
+
+    async def number(self, entity, invert=False):
+        return {"sensor.grid": 0.0, "sensor.soc_a": 50.0,
+                "sensor.soc_b": 50.0, "sensor.soc_c": 50.0,
+                **self.vals}.get(entity)
+
+
+def _named_trio(hass, store, *, hold_s=1, min_w=700):
+    opts = {"meter_entity": "sensor.grid", "meter_source": "off",
+            "max_w": 5000, "gain": 0.6, "auto_start": True,
+            "circulation_w": min_w, "circulation_hold_s": hold_s,
+            "inverters": _CIRC_ROSTER}
+    c = Controller(opts, hass, store, _P(), stop=asyncio.Event())
+    c.auto = True
+    for inv in c.inverters:
+        inv.dev_min, inv.dev_max = -5000.0, 5000.0
+    return c
+
+
+async def _circ_named(a_w, b_w):
+    hass = _NamedCircHass(a_w, b_w)
+    c = _named_trio(hass, _MemStore())
+    for _ in range(14):
+        await c.read_inputs()
+        await c.cycle()
+        await asyncio.sleep(0.25)
+    notes = [x for x in hass.calls
+             if x[0] == "persistent_notification"
+             and x[2].get("notification_id") == "goodwe_circulating_power"]
+    return c, notes
+
+
+_cg, _notes_g = asyncio.run(_circ_named(+3000.0, -3000.0))
+_cs, _notes_s = asyncio.run(_circ_named(-3000.0, +3000.0))
+_msg_g = _notes_g[0][2]["message"] if _notes_g else ""
+_msg_s = _notes_s[0][2]["message"] if _notes_s else ""
+print(f"       roster                     : {_cg.units}")
+print(f"       garage discharging, measured: "
+      f"{dict(zip(_cg.units, [i.batt for i in _cg.inverters]))}")
+print(f"         -> {_msg_g.split(chr(10))[2] if _msg_g.count(chr(10)) > 2 else _msg_g!r}")
+print(f"       shed discharging,   measured: "
+      f"{dict(zip(_cs.units, [i.batt for i in _cs.inverters]))}")
+print(f"         -> {_msg_s.split(chr(10))[2] if _msg_s.count(chr(10)) > 2 else _msg_s!r}")
+
+# --- THE PRECONDITIONS, AS THEIR OWN CHECKS (MUTATION.md rule 1) ------------
+check("PRECONDITION: the roster carries three DISTINCT operator-chosen names, "
+      f"none of them `UNIT_A` ({UNIT_A!r}) - so every claim below is a claim "
+      "about the roster and not about a constant that happens to equal it",
+      _cg.units == tuple(_CIRC_NAMES) and len(set(_CIRC_NAMES)) == 3
+      and UNIT_A not in _CIRC_NAMES)
+check("PRECONDITION: the two runs are MIRROR IMAGES - same house, same meter at "
+      "0 W, same magnitudes, and the two participants swapped. Without this the "
+      "direction claim below could be satisfied by a sentence that ignores the "
+      "measurement entirely",
+      [i.batt for i in _cg.inverters] == [+3000.0, -3000.0, _REST_W]
+      and [i.batt for i in _cs.inverters] == [-3000.0, +3000.0, _REST_W]
+      and _cg.grid == 0.0 and _cs.grid == 0.0)
+check("PRECONDITION: both runs actually tripped and actually raised the card, "
+      "so an absent sentence below is an absent sentence and not an absent alarm",
+      _cg.circ_wd.tripped and _cs.circ_wd.tripped
+      and len(_notes_g) == 1 and len(_notes_s) == 1)
+
+# --- AC 4: THE ATTRIBUTION ITSELF -------------------------------------------
+_WHO_G = "Right now garage is discharging while shed charges."
+_WHO_S = "Right now shed is discharging while garage charges."
+check("⚠️ THE ALARM NAMES THE TWO PACKS AND SAYS WHICH WAY EACH ONE IS PULLING. "
+      "`circulation.py` calls this the first thing somebody woken at 3 a.m. "
+      "needs and impossible to reconstruct from the watts alone, and until this "
+      "line the sentence could be deleted with all 26 files green",
+      _WHO_G in _msg_g)
+check("⚠️ AND WITH THE TWO SWAPPED, THE SENTENCE SWAPS. This is the half a "
+      "single fixture cannot have: `pos` and `neg` exchanged, or a hard-coded "
+      "sentence, satisfies the check above and fails this one. The operator is "
+      "being told which pack to walk to",
+      _WHO_S in _msg_s)
+check("...and each run carries ONLY its own sentence - the mirrored alarm does "
+      "not also contain the other direction, so the clause is derived from this "
+      "cycle's measurement rather than assembled from the roster",
+      _WHO_S not in _msg_g and _WHO_G not in _msg_s)
+check("...and the RESTING pack is not named. `cellar` is on the roster, is read "
+      "every cycle and is measured at 5 W of standby draw - `opposing()` is "
+      "called with a floor of half the trip threshold precisely so a pack idling "
+      "beside a 6 kW loop is not reported as one of its two participants",
+      "cellar" not in _msg_g and "cellar" not in _msg_s
+      and _cg.inverters[2].batt == _REST_W)
+check("...and the sentence sits with the figure, not instead of it: the same "
+      "card still carries the watts, the window and the three things only this "
+      "alarm can say",
+      all("CIRCULATING_POWER" in m and "METER CANNOT SEE" in m
+          and "CONTROL HAS BEEN STOPPED" in m and "isolator" in m
+          for m in (_msg_g, _msg_s)))
 print()
 print("   the operator's press is the only way out, and it clears BOTH halves")

@@ -0,0 +1,48 @@
+# WHAT THIS BUILD IS SUPPOSED TO ANNOUNCE, recorded so that a change to the
+# composition has something to disagree with. THREE COLUMNS:
+#
+#     <entity_id>   <unique_id>   <device identifier>
+#
+# ⚠️ ALL THREE, AND THE SECOND AND THIRD ARE THE HALF THE FIRST DRAFT MISSED.
+# A reference that carries only entity_ids cannot see the two changes that
+# orphan an install without changing a single composed id:
+#
+#   `BASE` renamed (`goodwe_ctl` -> anything)      every unique_id changes
+#   `DEVICE["identifiers"]` changed                HA sees a NEW device
+#
+# Home Assistant's MQTT integration keys its entity registry on `unique_id` and
+# its device registry on the device identifier. Change either and HA creates a
+# NEW entity beside the old one - and because discovery is published with
+# `retain=True` and nothing publishes a removal payload, the OLD config stays on
+# the broker and the dead entity stays in the registry, still named by every
+# automation that referenced it, never updating again. The composed entity_id is
+# identical on both, so an entity_id-only comparison reports a clean transition.
+# That is the same shape as the habit-4 this file's own card already caught
+# once: a reference that does not carry the thing that moved.
+#
+# ⚠️ AND `object_id` IS DELIBERATELY NOT A COLUMN. `app/mqtt.py` lines 36-42 and
+# 61-76 record that HA 2026.8 does NOT honour `object_id` in a discovery
+# payload - the id is composed from device name plus entity name - so renaming
+# one is a true no-op and a column for it would make this file refuse a change
+# that costs nothing. The check is graded on what HA actually keys on.
+#
+# PROVENANCE, stated exactly:
+#   - recorded 2026-09-08 from `release/1.0` at `meter_source: firmware`,
+#     roster of one, by reading `mqtt.discovery_payloads(units=[])`. It is what
+#     the code announces.
+#   - `sensor.goodwe_p1_sample_age` is additionally recorded in `app/mqtt.py` as
+#     VERIFIED AGAINST THE LIVE SYSTEM on 2026-08-26, and the ESP32's ingest
+#     guard subscribes to that exact string.
+#   - ⚠️ IT IS NOT A DUMP OF THE LIVE INSTALL'S ENTITY REGISTRY. That is what
+#     `--live-entities` is for. This file is about THE BUILD.
+#
+# Editing this file is how you say "the announcement changed on purpose". Do not
+# edit it to make a check pass.
+sensor.goodwe_setpoint            goodwe_ctl_setpoint     goodwe_rs485_controller
+sensor.goodwe_grid_power          goodwe_ctl_grid         goodwe_rs485_controller
+sensor.goodwe_battery_power       goodwe_ctl_battery      goodwe_rs485_controller
+sensor.goodwe_battery_soc         goodwe_ctl_soc          goodwe_rs485_controller
+sensor.goodwe_circulating_power   goodwe_ctl_circulation  goodwe_rs485_controller
+sensor.goodwe_maintenance_phase   goodwe_ctl_phase        goodwe_rs485_controller
+sensor.goodwe_controller_status   goodwe_ctl_status       goodwe_rs485_controller
+sensor.goodwe_p1_sample_age       goodwe_ctl_p1_age       goodwe_rs485_controller
@@ -0,0 +1,22 @@
+# ARCH-02 fixture: THE SHAPE THE LIVE BOX IS IN TODAY (a flat, single-inverter
+# config). ⚠️ ILLUSTRATIVE ENTITY NAMES, NOT THE LIVE ONES - the owner runs the
+# rehearsal against a copy of the real options blob, and this file exists so
+# that the harness has a synthetic transition it can be graded on.
+setpoint_entity: number.goodwe_setpoint
+soc_entity: sensor.goodwe_battery_soc_pct
+batt_entity: sensor.goodwe_battery_power_w
+batt_invert: false
+max_w: 5000
+bms_soc_floor: 10
+meter_source: 'off'
+meter_entity: sensor.p1_active_power
+inverters: []
+gain: 0.6
+slew_w: 0
+deadband_w: 15
+target_grid_w: 0
+step_w: 10
+heartbeat_s: 10
+stale_input_s: 15
+maintenance_discharge_w: 2000
+maintenance_charge_w: 2000
@@ -0,0 +1,39 @@
+# THE ENTITY IDS THE INSTALL HAS — a stand-in for the owner's real export.
+#
+# ⚠️ THIS FILE AND `announced-reference.txt` ANSWER DIFFERENT QUESTIONS AND ARE
+# NOT INTERCHANGEABLE:
+#
+#   announced-reference.txt   what the BUILD announces. Loaded on EVERY run.
+#                             Three columns, because HA keys on `unique_id` and
+#                             on the device identifier as well as on the
+#                             composed entity_id.
+#   this file (--live-entities)
+#                             what the INSTALL has. One id per line. Orphans
+#                             are graded against it: an id the install has and
+#                             the AFTER shape does not announce is an orphan
+#                             whatever the build's own history says.
+#
+# ⚠️ OMITTING `--live-entities` IS A REFUSAL (exit 2), NOT A QUIETER RUN. It
+# used to be the other way round - an unreadable file refused, a missing flag
+# passed with exit 0 and said nothing about the half it had not checked - and
+# the owner runs this ONCE, at commissioning, under time pressure.
+# `--no-live-entities` is the deliberate opt-out and it prints a standing
+# banner in the report header and again on the exit line.
+#
+# PROVENANCE: these eight are what `release/1.0` announces at
+# `meter_source: firmware`, roster of one, recorded 2026-09-08. They stand in
+# for the export so the checks run in CI. ⚠️ THE REAL INSTALL CARRIES ROUGHLY 49
+# `goodwe*` ENTITIES (card ARCH-02, read 2026-09-06); the other ~41 belong to
+# the ESPHome firmware, to template sensors and to helpers, and this add-on
+# announces none of them. The tool separates the ids it owns from the ids it
+# does not rather than claiming orphans it cannot cause - which is the half of
+# this file the eight-line stand-in cannot exercise, and the reason DOCS.md's
+# owner step asks for the real export.
+sensor.goodwe_battery_power
+sensor.goodwe_battery_soc
+sensor.goodwe_circulating_power
+sensor.goodwe_controller_status
+sensor.goodwe_grid_power
+sensor.goodwe_maintenance_phase
+sensor.goodwe_p1_sample_age
+sensor.goodwe_setpoint
@@ -0,0 +1,42 @@
+# ARCH-02 fixture: THE SHAPE AFTER INVERTER B IS COMMISSIONED.
+#
+# ⚠️ THE FIRST ROSTER ENTRY IS **NOT** CALLED `inverter-a`, AND THAT IS THE
+# WHOLE FIXTURE. `MUTATION.md` habit 4's seventh instance is MAINT-05 shipping
+# `migrate_maint_keys(store, UNIT_A, ...)` under a comment saying it must not:
+# every two-pack fixture in the suite named its first entry `inverter-a`, so
+# `UNIT_A` and `inverters[0].name` were the same string in every check that
+# existed and the defect was invisible. Names that are not `inverter-a` are what
+# make a misdirected namespace something a rehearsal can SEE.
+#
+# `roster-inverter-a.yaml` beside this file is the same roster with the first
+# entry named `inverter-a`, so the two can be compared - see DOCS.md.
+setpoint_entity: number.goodwe_setpoint
+soc_entity: sensor.goodwe_battery_soc_pct
+batt_entity: sensor.goodwe_battery_power_w
+batt_invert: false
+max_w: 5000
+bms_soc_floor: 10
+meter_source: 'off'
+meter_entity: sensor.p1_active_power
+gain: 0.6
+slew_w: 0
+deadband_w: 15
+target_grid_w: 0
+step_w: 10
+heartbeat_s: 10
+stale_input_s: 15
+maintenance_discharge_w: 2000
+maintenance_charge_w: 2000
+inverters:
+  - name: inverter-a
+    setpoint_entity: number.goodwe_setpoint
+    soc_entity: sensor.goodwe_battery_soc_pct
+    batt_entity: sensor.goodwe_battery_power_w
+    max_w: 5000
+    bms_soc_floor: 10
+  - name: inverter-b
+    setpoint_entity: number.goodwe_b_setpoint
+    soc_entity: sensor.goodwe_b_battery_soc_pct
+    batt_entity: sensor.goodwe_b_battery_power_w
+    max_w: 4600
+    bms_soc_floor: 10
@@ -0,0 +1,42 @@
+# ARCH-02 fixture: THE SHAPE AFTER INVERTER B IS COMMISSIONED.
+#
+# ⚠️ THE FIRST ROSTER ENTRY IS **NOT** CALLED `inverter-a`, AND THAT IS THE
+# WHOLE FIXTURE. `MUTATION.md` habit 4's seventh instance is MAINT-05 shipping
+# `migrate_maint_keys(store, UNIT_A, ...)` under a comment saying it must not:
+# every two-pack fixture in the suite named its first entry `inverter-a`, so
+# `UNIT_A` and `inverters[0].name` were the same string in every check that
+# existed and the defect was invisible. Names that are not `inverter-a` are what
+# make a misdirected namespace something a rehearsal can SEE.
+#
+# `roster-inverter-a.yaml` beside this file is the same roster with the first
+# entry named `inverter-a`, so the two can be compared - see DOCS.md.
+setpoint_entity: number.goodwe_setpoint
+soc_entity: sensor.goodwe_battery_soc_pct
+batt_entity: sensor.goodwe_battery_power_w
+batt_invert: false
+max_w: 5000
+bms_soc_floor: 10
+meter_source: 'off'
+meter_entity: sensor.p1_active_power
+gain: 0.6
+slew_w: 0
+deadband_w: 15
+target_grid_w: 0
+step_w: 10
+heartbeat_s: 10
+stale_input_s: 15
+maintenance_discharge_w: 2000
+maintenance_charge_w: 2000
+inverters:
+  - name: garage
+    setpoint_entity: number.goodwe_setpoint
+    soc_entity: sensor.goodwe_battery_soc_pct
+    batt_entity: sensor.goodwe_battery_power_w
+    max_w: 5000
+    bms_soc_floor: 10
+  - name: shed
+    setpoint_entity: number.goodwe_b_setpoint
+    soc_entity: sensor.goodwe_b_battery_soc_pct
+    batt_entity: sensor.goodwe_b_battery_power_w
+    max_w: 4600
+    bms_soc_floor: 10
@@ -0,0 +1,11 @@
+{
+ "_comment": "ARCH-02 SYNTHETIC FIXTURE - a store that has ALREADY run the flat config under 0.7.0+, so DEV-02's `schema_version` is stamped and the SAFETY-02 latch is not tripped. NOT a copy of the live box's /data/state.json. This is the CLEAN CONTROL for the MAINT-05 half of the rehearsal: the flat->roster transition over it must come out with NO findings, which is what makes each deliberate mutation's RED mean something. A harness that is red on every input has not measured anything.",
+ "schema_version": 1,
+ "phase": "idle",
+ "phase_started": "2026-08-07T10:00:00+00:00",
+ "last_completed": "2026-08-07T18:42:00+00:00",
+ "last_start_attempt": "2026-08-07T10:00:00+00:00",
+ "sign_tripped": false,
+ "sign_reason": "",
+ "auto": true
+}
@@ -0,0 +1,3 @@
+{
+ "_comment": "ARCH-02 SYNTHETIC FIXTURE - a FRESH INSTALL. No maintenance history, no latch, no schema markers: what /data/state.json looks like before the add-on has ever completed a cycle. It is a clean control, and it earned its place: the first draft of the harness reported this transition RED, because the DEV-02 migration writes both SAFETY-02 defaults under `inverter-a` before a roster exists and the tool called them stranded. Losing a default loses nothing."
+}
@@ -0,0 +1,12 @@
+{
+ "_comment": "INSTR-01 SYNTHETIC FIXTURE - `state-legacy.json` with a SECOND roster unit already latched. Not a copy of any live state file. It exists to restore M12 END-TO-END on the STAGED leg, which is the arrangement SAFETY-44 and INSTR-01's first draft both missed: `snapshot_over()` starts the after-leg over the file the BEFORE-leg wrote, so a transition-made destination is available there and the two questions the provenance clause and the narrowing ask are NOT incompatible. Every same-step arrangement makes them incompatible; that is what was measured and over-generalised to 'no fixture can'.",
+ "phase": "idle",
+ "phase_started": "2026-08-07T10:00:00+00:00",
+ "last_completed": "2026-08-07T18:42:00+00:00",
+ "last_start_attempt": "2026-08-07T10:00:00+00:00",
+ "sign_tripped": true,
+ "sign_reason": "SIGN_DISAGREEMENT: commanded -1800 W against a meter reading -1450 W for 190 s",
+ "auto": true,
+ "unit/shed/sign_tripped": true,
+ "unit/shed/sign_reason": "SIGN_DISAGREEMENT: shed, latched before the switch"
+}
@@ -0,0 +1,10 @@
+{
+ "_comment": "ARCH-02 SYNTHETIC FIXTURE - a PRE-DEV-02 flat state file. NOT a copy of the live box's /data/state.json; this session cannot read that one and did not try. It is built to carry, deliberately, the two shapes that state built by the happy path does not have: a LATCHED SAFETY-02 failsafe under the flat spelling, and a real maintenance history under the flat spelling. Both are LEGACY-SHAPED: there is no schema_version and no maint_schema_version, so this file was last written by a controller that had never heard of unit namespaces. Losing either one is silent.",
+ "phase": "idle",
+ "phase_started": "2026-08-07T10:00:00+00:00",
+ "last_completed": "2026-08-07T18:42:00+00:00",
+ "last_start_attempt": "2026-08-07T10:00:00+00:00",
+ "sign_tripped": true,
+ "sign_reason": "SIGN_DISAGREEMENT: commanded -1800 W against a meter reading -1450 W for 190 s",
+ "auto": true
+}