Eight workstreams, two migrations, and a pass by an independent reader that
found four things this release was claiming about itself that were not true.
Two of the fixes change what people receive. Medication reminders went out
every fifteen minutes because the dedup key could not tell one dose slot from
another, and "handled by the app" was silencing every channel rather than the
one it names. Both are called out in the notes, because a user who set that
flag as a general mute will notice.
Two came from reports: an OpenAI-compatible gateway that the provider list
had no room for, after an earlier answer from us said otherwise, and four
places where a timestamp followed the browser instead of the profile.
The backup half is the one worth reading. Three record types were carried and
could not be restored, or could not be carried at all, and each failed in a
different direction: one dropped silently, one stopped the restore outright,
one was discarded while the restore reported success. What a backup holds is
now declared in one file with a written reason for everything left out, and
the roughly thirty types that are still carried by neither end are named there
instead of waiting to be discovered by someone who needed them.
The Del chip added beside the document card measured 4.06:1 against the card's
background, under the 4.5:1 floor, and axe failed the vault on it. So the
change that made the destructive key visible introduced a barrier of its own.
It uses the full foreground token now. The design rule says muted for meta
text, and a key hint is meta — but it is the one kind of meta that has to be
readable at a glance, so here the contrast rule wins and the reason sits next
to the class list rather than in a commit nobody will read again.
Found by running the e2e suite locally before pushing rather than after.
The guard tested whether the argument list contained the substring "timeZone".
`timeZoneName: "short"` contains it. That option appends "GMT+1" to the output
and leaves the zone as the browser's, so a formatter written that way sailed
through while being exactly as wrong as a bare one — and looking more careful
than a bare one, which is worse.
Tightening it to a colon-terminated property then reported nine correct call
sites, because the shorthand `{ timeZone }` is what most of this tree writes.
The honest response to that would have been to allowlist all nine, which is
how a guard stops guarding. It accepts both forms now, and a test pins both
directions rather than leaving the regex to be trusted.
Two exclusion reasons in the backup plan aged inside their own release and are
corrected: push attempts now also hold the reminder dedup anchors this release
introduced, and the thumbnail's "rebuilds from the stored bytes" only holds on
the disaster-recovery path, since a portable export carries no bytes to rebuild
from.
Four things an independent read found, three of them things this release said
about itself that were not true.
A custom mood tag travelled without its label, its icon, its sort order and
its SCALE. The label is the smaller loss: the person who named a factor
"Migräne" would have got a key back. The scale is the real one — the inverse
flag marks a factor where a high value is bad, and dropping it does not lose a
label, it silently reverses the meaning of every rating already recorded
against that tag. Custom cycle symptoms lost their label the same way. Both
now carry the ciphertext verbatim, like every other encrypted column in a
disaster-recovery payload.
The plan file claimed, in the present tense, that a guard requires both a
payload reader and a restore branch for every carried model. No such guard
exists. The commit that introduced the file said so honestly; the file itself
said the opposite, and roughly 31 models have neither end. BACKED_UP means
"must be", not "is", and both the plan and its test now say that plainly. The
wrong sentence is recorded rather than quietly swapped, because it is the same
defect the file was written to prevent.
The export route still told operators that labs, allergies, illness episodes
and workouts could be exported but never restored. The restore route
re-creates all of them, and since this release nutrient day totals too. The
scope now points at the plan instead of repeating it beside the code, which is
how that sentence came to be wrong.
And the export card promised "every record you own". It does not, so it no
longer says so — six locales.
The resolver is online-first, so mounting the databases fills the ASN
and carrier but does not by itself keep login IPs on the host. Point at
the IP_GEO_LOOKUP_DISABLED section that does.
The hint said switching a widget off here turns it off in the mobile
app. This page writes the flag; whether the client applies it on its
next read is the client's business, and there is an open report that it
does not. Say where the setting lives instead of promising a result
this surface does not control. Six locales.
BMI renders on the native dashboard and was configurable from nowhere.
Two individually-correct decisions produced it: `bmi` is a catalogue id
with no web render path, and the web layout editor filters those out
because a toggle over something this page cannot draw would be a silent
no-op. The widgets PUT accepts and persists them all the same, and the
native client writes them, so an account ended up holding rows it never
set and could not change from any surface.
The sortable list stays exactly as it was. Below it, the ids the web
does not draw but the layout carries get their own labelled group with
the same two flags and no reordering — placement belongs to the client
that draws them. An account with no native client has nothing there and
sees nothing. The module gate applies as it does above, fail-open.
Refs #581
A focused card deleted on Backspace as well as Delete. The timeline's
roving tabindex gives a card focus the moment someone tabs into the
grid, Backspace is the reflex "go back" key, and a card carries no
delete control at all — so the reflex press removed a medical document
with nothing but a transient undo toast behind it.
Delete alone now deletes. The key resolution moves into a pure helper
in vault-utils so the contract is testable without a render harness,
and the card stops keeping that contract in a code comment: the overlay
button carries aria-keyshortcuts, and a kbd chip appears while the card
holds focus. The label is per-locale (Entf / Suppr / Supr / Canc).
The switch was described three times as applying to what you upload
next. A genuine off-to-on flip also enqueues a pass over the documents
already stored without a summary, capped at 200 per flip — scanned
medical documents going to a third-party provider, from copy that reads
as forward-looking only.
Drop "newly uploaded" from the sub-label and give the disclosure the
sentence it was missing, including the cap. Six locales. The behaviour
is right and stays as it is; only the sentence changes.
A guard pins the number in the copy to MAX_ENQUEUES_PER_RUN, so moving
the cap without touching the copy fails.
The runtime has always read GEOLITE2_DIR, but the variable was absent
from the compose `environment:` whitelist, from both env examples, and
from the docs. A variable that is not on that whitelist never reaches
the container even when the operator's `.env` sets it, so the override
existed in code and was unreachable from the documented path.
Put it on the whitelist, document it in both examples and the manifest,
and add `docs/self-hosting/geolite2.md`: fetch the databases with a free
MaxMind account, mount the directory read-only, point the variable at
it. The compose file carries the bind-mount line commented out beside
the app service.
The admin notification told the reader to set MAXMIND_LICENSE_KEY in
this repository's Actions secrets — something nobody running the
published image can do, about a fallback that sends their sign-in
addresses to a third party. It now names the mount, names the online
provider it is actually querying, and links the runbook. Six locales.
It also no longer claims the carrier column stays empty; the online
tier has filled the carrier from the ISP field since v1.25.8.
A blank GEOLITE2_DIR now resolves to the default instead of the process
working directory: compose substitutes an unset variable to the empty
string, which is not nullish.
Refs #659
The Bestand tab was 1,139 lines of one client file holding two concerns that
only meet at the row shape: the list of containers a person holds, and the
three dialogs that change it. Registration, per-item correction, packaging
economics, the summary and the row rendering all shared a scope, so reading
any one of them meant scrolling past the rest.
Move-only. The dialogs go to `inventory-dialogs.tsx` with the day-input
conversions their fields produce; the row shape and the cache-invalidation
they both need go to `inventory-shared.ts`; the section keeps the summary and
the list at 458 lines. Every export the tab offered is still offered here, so
no caller changes. No copy changes, no new keys.
Two suites pin these contracts by reading the file's own source — the register
dialog's unit conversion and the correction dialog's sparse patch. Their
`readFileSync` now names the file the code moved to. The assertions are
unchanged, byte for byte; only the path they read is.
The retention-purge comments still cite migration 0277 for indexes that landed
in 0276 — the renumbering moved the directories and left the prose behind. The
schema is trustworthy because its comments are, so a wrong number is not
cosmetic.
The import worker's `already_recorded` reason claimed the skip covers a dose
recorded on any other intake surface. It does not: `source` is part of the
live-row unique, so a hand-logged dose and an imported one at the same slot are
different keys and both are written. Say what the mechanism actually does, and
say plainly that cross-surface dedup is not there.
The per-medication import and the account-wide dose-history import queued
their work through ninety near-identical lines each: row lock, stale-job
sweep, active-job refusal, job row, pg-boss send, and the fallback that fails
the row when the queue is not there. Five steps written twice drift, and the
drift shows up as an import that hangs on one surface and not the other. The
status projection moved out of these routes for the same reason; this is the
write half.
Scope is the only real difference, and it is one argument. What the two
surfaces legitimately say differently — the 409 wording, the envelope, the
kickoff audit line — stays at the call site.
The replay key had the same shape of problem in smaller form: the export
reader owned a function for it and the hand-listed body inlined the same
template. Both now call one function, with a test that pins them equal for
the same medication and instant.
The supply tab carried its own copy of the two container kinds whose first
opening starts a degradation clock, with a comment saying it mirrored the
state machine's list. Nothing enforced the mirror — the state machine types
itself against the generated Prisma client, which the browser bundle cannot
carry, so a copy was the only way to say it.
Move the list into a client-safe module with no imports and let both sides
read it. The state machine re-exports it under the Prisma enum, so a renamed
or dropped container kind is now a compile error rather than a member that
quietly stops matching.
The resolver reads the profile's sex for the body-fat and pulse bands, and
nothing asserted it. A resolver that stopped reading the field, or read it and
defaulted to one side, passed the whole suite. Pin both recorded sides and the
neutral band that OTHER and an unanswered profile share.
`binaryReferenceSex` was already the declared owner of the shrink from the
three stored values to the two a sex-split reference table has rows for, but
eight sites still narrowed inline — including a byte-for-byte private copy in
the effective-range resolver and a ternary 37 lines above a `toProfileSex`
call in the same file. Each of those is a place the fallback can drift.
Delete the copy, replace the ternaries, and name the pair once as
`BinaryReferenceSex` so a signature that needs it stops re-listing the two
literals. No behaviour changes: every site already resolved OTHER and "no
answer" to the same neutral or absent value, and the existing suites pass
untouched.
A migration container starts empty, so applying 0281 there proves the
statements parse and nothing else — the backfill's WHERE matches no rows and a
broken expression passes green. Every self-hoster runs it against rows that do
carry a leg, and those are the only rows it exists for.
Seed the pre-migration shape, run the migration's own UPDATE read out of the
committed file, and assert an attributed row comes out still knowing which leg
broke, an unattributed one is left alone, and a second run does not nest the
array.
The batch route has accepted a `syncTrigger` since v1.32.8 and put it on the
wide event, where no human reads it. Half of what was promised got built: the
value is written and appears in three places in the codebase, none of which is
a screen anyone looks at.
Persist it on the account, alongside the instant of the most recent batch that
arrived without the app being opened, and publish both from
`GET /api/integrations/healthkit`. The card now names how the last sync arrived
and when data last came in on its own. `healthKitLastSyncedAt` alone reads
identically whether the phone delivers by itself or only while someone has the
app open, which is why "is background delivery working" could not be settled
from either side.
A batch that declares no trigger writes null rather than leaving the previous
one standing — the field describes this sync, and an older client's silence is
not evidence about it. A foreground batch leaves the background timestamp
alone; overwriting it would erase the only answer the card has.
Also here: WHOOP, Fitbit and Google Health join the workout-freshness table.
All three write workouts, and without an entry their cards could report every
measurement type fresh while the workout leg had been silent for weeks. And the
status pill drops a generic `aria-label` that replaced its own accessible name,
so a screen reader announced the same words for a healthy connection as for a
broken one.
`failing_leg` was one slot, so two legs of the same provider failing in
overlapping windows overwrote each other. The row remembered whichever failed
last, and that leg recovering then cleared an error the other one was still
causing — the ledger forgot a live failure and painted the card green over it.
The slot becomes a set. A failure adds its leg, a success removes its own, and
the error state, the per-kind buckets and the streak anchor clear only once the
set is empty. The array is ordered oldest failure first, so the last element is
the leg that wrote the message `last_error` holds; when that leg recovers while
others are still failing, the message is cleared rather than left describing a
leg that is now healthy.
An unattributed failure adds nothing and removes nothing. A single-leg provider
therefore behaves exactly as before, and a multi-leg provider no longer loses
what its other legs already reported.
The column never reached a client — it is absent from the status serialiser and
from the OpenAPI contract — so it is replaced rather than shimmed.
The module-gate branch cited the client-managed skip as its precedent for
advancing past a cycle. That skip no longer exists in this file, and the
reason the gate advances is its own: a disabled module has nothing left
to deliver.
`apns-collapse-id` is a replace directive: a push carrying an id
supersedes the earlier undelivered push with the same id. Every
medication reminder carried the literal string MEDICATION_REMINDER, so
three medications due at 08:00 arrived on the phone as one notification
naming one of them. The user takes that one and misses the others with
nothing on screen to suggest anything was dropped.
Key the collapse on the per-slot tag the Web Push sender already
coalesces on, falling back to the event type when a payload carries none.
Grouping stays on `threadId`, which is what grouping is for. Discreet
mode keeps the generic id — no record-identifying string on the lock
screen — and a tag over Apple's 64-byte ceiling falls back rather than
costing the send.
The per-phase "already sent today" guard read the Telegram message
ledger, and that table is written only when a Telegram send succeeds. A
user on email, APNs or ntfy never had a row, so the 15-minute tick
re-derived the same overdue slot and sent it again on every pass — around
fifty mails for a single missed dose.
Anchor the guard in `push_attempts` instead, the ledger the daily
briefing and the illness escalation already use. Two properties are
deliberate: the anchor is stamped for every dispatched slot regardless of
what any channel returned, because keying it on a delivered push would
re-open the flood the moment a channel starts failing; and it is stamped
before the dispatch, so a crash costs one phase's notice rather than a
repeat every quarter hour.
The key carries the schedule id and the slot's real HH:mm alongside the
phase and the local date. Both are load-bearing: two legacy
single-window schedules on one medication share the empty first-class
time-of-day, so a coarser key would collapse the evening dose into the
morning one. The tests assert the other direction too — the next phase
still escalates, and the next local day starts clean.
The Telegram ledger keeps its message-tracking duties; it is no longer
the dedup key. Anchor rows carry a sentinel channel and stay out of the
operator's notification health and diagnostic reads, which count
deliveries, not decisions.
`medication.clientManaged` and `measurementReminder.clientManaged` mean
the iOS client schedules the reminder locally, so the server should not
push it. That is what the OpenAPI description, the settings copy and the
iOS contract all say. What the crons did was skip the dispatch entirely,
so a user who flipped the switch also lost Telegram, ntfy, webhook,
email and Web Push — channels they may well read on a desktop while the
phone holding the local schedule sits in a drawer.
Move the gate into the dispatcher's channel loop, where it applies to
APNs and nothing else. The skip is not counted as an attempt, so
`dispatched` keeps meaning "something reached the user", and it stays
countable: the same wide-event keys the crons emitted are emitted here,
plus a `skipped` row in the push-attempt ledger.
The Vorsorge path is the sharper half. It suppressed every channel and
then advanced `nextDueAt`, which dropped the reminder out of the
preventive-care card and out of the daily digest: a reminder that never
reached anyone was recorded as handled. The advance now happens only
where it always belonged, after a dispatch that a channel actually
delivered. When nothing delivers, the reminder stays overdue and visible.
The restore resolves a mood entry's rated factors by key and asks only for the
seeded catalogue. A tag the user created is not in it, so the lookup came up
short and the whole restore threw 'Unknown mood factor keys'. The file was
complete; it simply could not be read back.
That is the same cause as the cycle symptoms two commits ago, with the
opposite ending. There, an unresolvable key was filtered out and the loss was
silent. Here it stopped everything. One shape of missing definition, two
answers, neither of them the right one.
The account's own tag definitions travel now, get written before the factors
resolve, and the lookup accepts the seeded catalogue or the owner's own. The
seeded rows stay out of the file for the same reason as before: a restore must
not rewrite another instance's reference data.
Four more test mocks needed the new delegate, which is again the compiler and
the suite naming the call sites rather than me hunting for them.
The export has written nutrient day totals since v1.29. The canonical schema
never declared the field, so `.passthrough()` carried it through parsing and
no reader ever looked at it: every water and vitamin total in the file was
discarded on restore, and the restore reported success.
Declared in the contract, cleared and re-created beside its siblings, and
counted in both places that count — what the file holds and what the restore
replaced. A count that stops at the domains someone remembered is how this
stayed quiet.
restoreCycleData was loud about one missing reference and silent about
another, three lines apart. An unknown cycle reference threw; an unknown
symptom key was filtered out of the list, so the day-log came back with one of
its symptoms quietly gone and the restore reported success.
The cause was upstream: the seeded catalogue is reference data every instance
has, but a symptom the user created lives only in their account and was never
carried. The key had nothing to resolve against, and the filter turned that
into silence.
Both ends now. The account's own definitions travel with the backup, the
seeded rows deliberately do not — shipping those would let one instance's
restore rewrite another's reference data. And the resolve step throws on an
unresolvable key rather than filtering it, so the same function answers the
same question the same way twice.
Widening the builder's client type made four export tests fail on a missing
delegate, which is the sweep working: those are the call sites, named by the
compiler instead of found by reading.
The mirror of the wipe plan, for the same reason and against a worse failure:
a list that was right when written and stops being right every time a model is
added. On the delete side that leaves data behind; here it means data does not
come back, and nothing says so until someone is restoring a backup because
something already went wrong.
The list is longer than the wipe list by sixteen models, and that gap is the
point. A wipe may lean on the database — deleting a medication cascades to its
schedules, dose changes and inventory events — but a restore has no cascade,
so every child has to be named here.
Every exclusion carries its reason in prose, and the guard enforces a length
floor on it. That floor is not pedantry: the first draft of this file said
'Same as WithingsConnection' four times, which is a pointer rather than a
reason, and a pointer silently inherits whatever the thing it points at
becomes. Writing the decision where it is made is the only way it stays
reviewable.
The guard found two things in its own author's work on the first run: a model
with no verdict at all, and a reason too thin to be one.
The restore side of the guard — that every carried model also has a restore
branch — lands with the payload work it will police.
A formatter with no timeZone follows the browser, which is a different
question from the one this app answers. Four surfaces had drifted apart that
way before anyone noticed, and one of them printed a shared report a day off
from the page it came from.
The admin host chart went with them rather than onto the allowlist. Server
samples are real instants, and an operator reading their own instance should
not find a spike at 14:20 here and 15:20 two screens away.
Two deliberate limits, both written into the test. It does not match
toLocaleString, which in this tree is overwhelmingly a number formatter, and
burying four real hits under every thousands separator is how a guard gets
silenced. And it cannot tell a Date built from calendar arithmetic — a month
label, a weekday header, no instant behind it — from one parsed out of stored
data, so those four sites carry a written reason instead. The value is the
ratchet: a new bare formatter has to be argued for before it ships.
The bucketer hard-coded Europe/Berlin for day, week and month boundaries, so
a reading taken on the last evening of a month landed in the wrong month's
bar for everyone who does not live there. Two smaller siblings did the same
thing: the Today chip asked the browser for a time instead of the profile,
and the score card's tooltip formatter carried a locale but no timezone.
The boundary timezone is a required option rather than a defaulted one. A
default would have left any caller that forgets to pass one still wrong and
still silent; required makes the compiler name them, which is how the three
call sites and the eight test sites in this diff were found rather than
guessed at. The tests were always asserting Berlin's calendar without saying
so, and now they say so.
Two fixture mistakes worth recording, both caught by running the tests rather
than trusting them: summer time is already in force in Berlin on 31 March, so
the first instant chosen was April on both sides; and 31 March 2025 falls in
the same ISO week as 1 April, so the week case needed its own instant where
one side is still Sunday.
The page at c/[token] resolved the owner's timezone; the PDF download beside
it resolved nothing and fell through to the display default. Same link, same
reading, two dates — and the one a practice files is the one that was wrong.
Resolved in the shared download gate rather than in the route, so the PDF and
the FHIR bundle cannot drift apart the way the page and the PDF did. The
owner's clock preference travels with it for the same reason.
The fixture is deliberately west of UTC: Berlin is the fallback, so a Berlin
fixture cannot tell a resolved timezone from a forgotten one.
The rotation-coverage guard greps the script for a quoted column name.
Listing the new column in the header comment satisfied that grep without
the script rotating anything, which made the guard pass on a script that
would have left the column behind. The comment now names it unquoted and
says why.
Names LiteLLM, OpenRouter and vLLM, states what it does not share with
the OpenAI provider and why, and keeps the Local provider's own section
intact — it still speaks the same wire and remains the right pick for a
model on your own machine.
The properties this provider has to keep: it resolves against the base
URL that was saved for it, the pinned OpenAI arm cannot see that URL and
still posts to api.openai.com, and the OpenAI key never travels to the
gateway. Each is asserted at the wire, not on internals.
The SSRF floor is checked through the same alt-notation classes the
Local provider is checked against, and the pinned tags are shown to keep
requiring a public host even when the operator opened every private one.
The AI settings connect-src is pinned to what it already allowed: the
gateway is called server-side, so no CSP change is needed, and a future
change that needs one will turn this red.
Its own form with its own three fields, its own entry in the provider
dropdown and in the fallback-chain editor, and copy in six locales that
says plainly where the OpenAI key does not go.
The chain editor offers the entry but the default chain does not carry
it: it resolves only against a base URL somebody saved.
The read endpoint reports the gateway's base URL and model, and the
bearer as presence only. OpenAPI regenerated.
The chain gains a seventh entry. It builds the OpenAI client against the
base URL the user saved, with the bearer they saved for it and the model
they named, and it reads none of the OpenAI columns — the pinned
api.openai.com arm stays pinned.
Host policy is the Local provider's, unchanged: a public host always, a
private one only where the operator allowlisted it, which is what a LAN
LiteLLM or vLLM needs. Egress keeps riding safeFetch.
The JSON-mode dialect the Local client learns per endpoint moves into a
shared module and now serves both clients, so a gateway that rejects
response_format is detected once and retried without it.
The entry stays out of the default chain: it can only resolve against a
base URL somebody configured.
Three dedicated User columns — base URL, optional bearer, model — for a
provider whose endpoint the user chooses. They are separate from the
OpenAI columns on purpose: the OpenAI arm pins api.openai.com, so it must
share no column with a user-supplied host, and a gateway must never
receive the key issued for OpenAI.
The bearer joins the encrypted-column registry and the rotation script;
the three columns join the account-wipe reset set.
Four specs navigated and then read the DOM once, with nothing in between
that waits for anything but the network. An idle network says nothing about
whether React has rendered, so the read can sample a tree that is still
mounting and the assertion then blames the application for a race inside the
test. Two such reads went red on consecutive CI runs, on different specs.
The settings over-scroll guard also slept a fixed 500 ms, which is the wrong
shape twice over: too long when the page is ready, too short on a loaded
runner, and it never says what it is waiting for.
The structural guard runs in the unit suite rather than in Playwright,
because e2e runs at PR-CI only and flaky counts as failed there — one
unguarded read otherwise costs a whole pipeline before anyone sees it. It
refuses to count a wait whose rejection is thrown away, since that reads
like a guarantee and is not one, and would leave the check defeatable by
appending a catch. The one deliberate swallow is on the allowlist with its
reason: a genuine regression there still fails at the width assertion below
it, which reports the measured number instead of a timeout message.
A release asset is served with content-disposition attachment and nosniff, so
an image embedded from that URL renders as a broken image in the notes. The
raw path serves it as an image, which is the only thing that actually works.
A chart shows the shape and hides the numbers. An appointment needs the
numbers, and so does anyone moving readings into a spreadsheet.
The table reads the same expression the chart hands Recharts, after the range
slice, the week/month bucketing and the display scale, and it formats through
the chart's own value and date formatters. Nothing is re-derived, so a cell and
the tooltip cannot print two numbers for one point.
Collapsed by default, with the point count in the trigger so the size reads
while it is closed; the region scrolls in its own box rather than growing the
page, and the table primitive already scrolls a wide row set sideways. The
caption names the grain, because a weekly mean labelled as a daily reading
would be a lie, and the grain now has one source shared with the bucketing.
Derived overlays get no column: a moving average is drawn from the points, not
measured. An empty window renders no table at all, since the chart's own empty
state already says there is nothing there.
Opt-in per mount. The metric sub-pages take it; the dashboard tiles do not,
because a table under every tile buries the dashboard it summarises.
Refs #636
The tag crosstab already ran a with-versus-without Welch test per structured
tag, with a D to D+1 pairing mode whose stated example was an evening tag
against next-day recovery. It only ever had one channel at that lag, and a
recovery score is a wearable composite plenty of accounts never record. The
rated-factor board next to it has tested next-day resting heart rate and HRV
since v1.14.0, but a rated factor is a continuous score, so a binary tag never
reached either vital.
Add both to the tag map. Same engine, same per-side day floors, same
Benjamini-Hochberg family, same card. The fetch filter derives from the map, so
the channels come along without a second query.
The delta read-out goes neutral to match the rated-factor board: with resting
heart rate on the board, a positive delta is worse on one row and better on the
next, and a green plus sign would claim a verdict the association cannot carry.
Refs #616
The preamble already says once what the two states mean and that neither is
absence; repeating it per row would have doubled a manifest that rides every
turn.
The unit suite mocks Prisma, so it proves the discriminant and nothing about the
query underneath it — a wrong column would have left every test green and every
account answering no_data_unconfirmed in production. Seeds the reported case and
the four other shapes: an empty record, another account's rows, tombstoned rows,
and a history that mixes units.
Rounds the out-of-window mean to one decimal, like every other mean the snapshot
surfaces, and annotates a failed manifest probe rather than letting it degrade in
silence.
Marc-André Bombeck