Three surfaces built the label key by interpolation — the doctor-report
PDF's visit table, the daily digest's upcoming-visit line and the
clinician share view. `EncounterKind` is `ROUTINE` and the bundle leaf is
`routine`, so all three rendered `encounters.kind.ROUTINE` verbatim: into
a document a practice files, onto a lock screen, and onto the page.
Route them through `encounterKindLabelKey`, a literal switch beside the
existing `encounterKindLabel`, so the mapping is written down once.
The enum-derived i18n guard could not have caught this: it only knows
key spaces of the form `prefix + member`, and this one is not. It gains a
second arm for spaces whose mapping is a function, with `encounters.kind`
as its first entry. Break-proof: point one arm of the resolver back at
the uppercase key and six locale cases go red.
The authenticated shell wraps every route and read one destination
predicate off the `@/lib/record-settings` barrel. That barrel also
re-exports the managed settings-patch schemas, so Zod, the module
registry and three validation modules rode into the eager client
bundle of every authenticated page for a function that inspects a URL
slug. The shell imports `record-settings/classification` now; the
barrel loses the re-export nobody else used.
Measured over all 585 client-rendered routes: none grew, 100 fell, by
4.7 KB gz at the median and 8.1 KB at the most. /insights had been
sitting on its 445 KB ceiling for three releases and went over it on
this branch's CI run; it measures 437 now, /measurements 435, / 450,
/insights/mood 450.
The whole-build sum moved the other way, 3451 to 3476, because the
aggregate counts each emitted chunk once: a graph that leaves the
chunk group every route shares gets counted again wherever it lands,
here inside the deferred settings chunks that use it. Bytes on disk
rose, bytes a reader downloads only fell. The aggregate cap moves to
3500 with the same 24 KB headroom the four earlier re-measurements
kept, and the reasoning sits next to the number.
The two halves of the ECG fix were built apart. One taught the parser
Apple's full English verdict set; the other taught it the metadata keys and
verdict wording a German watch writes. Each was correct alone, and each
left "Sinusrhythmus" unresolved for a different reason.
Together they meet: the alias table translates the wording onto Apple's
own English, and the classifier decides what it means. One place still
owns the meaning, which is what stops a per-language verdict table from
drifting the first time Apple adds a case.
The test that expected null was pinning the gap between the branches
rather than a decision about German. It now expects NOT_DETECTED, and a
French verdict takes over the null case, because four of the six
languages genuinely have no observed key map yet.
The CSV is written in the device's language, and the keys are translated
along with the values: a German watch writes `Aufzeichnungsdatum` and
`Klassifizierung`, not `Recorded Date` and `Classification`. Indexing on
the English keys found none of them, so the parser never located the
waveform and rejected the recording outright — a lost import, not a
missing field.
Two alias maps close it, both taken from real exports rather than from a
dictionary: nine metadata keys and two verdicts, consistent across every
German file observed. Adding a language is a data addition to those maps
and nothing else.
The verdict map is an indirection onto the English wording rather than a
second verdict-to-enum table, so there stays exactly one place that
decides what a verdict means.
The other four shipped languages are absent on purpose. A wrong key
silently mis-files a value and a wrong clinical verdict is worse than
none, so an export this parser cannot place is refused — which an operator
can see — rather than read into the wrong column, which they cannot. Two
gaps are written down at the map: no observed file carries an
average-heart-rate row in any language, and the paired layout's column
header has only ever been seen in English.
The rate matcher demanded "Hz". Observed exports spell the unit out and
lower-case it — "511,422 hertz" — and the German ones separate it from the
number with a no-break space. The fixtures had only ever used "512 Hz", so
the mismatch never showed.
Rate and heart rate now split into a magnitude and a unit word, with the
magnitude read through the same decimal helper the waveform uses. "512 Hz",
"512 hertz" and "511,562 Hertz" all resolve; a missing unit, the wrong
unit, a grouped number and a rate past the bound are all still refused.
The unit word is matched loosely and the magnitude strictly. That is the
right way round: the unit is presentation and varies by region, while the
number is the measurement.
A watch set to a comma-decimal region writes its samples as "-180,596".
The single-column reader rejected any row containing a comma outright, so
every one of those files failed on its first sample row.
The two waveform layouts disagree about what a comma means, which is why
this cannot be decided per line. The parser now names the section it is in
— header, single-column, paired — and both crossings are commented at the
point they happen: the paired layout announces itself with a `Lead,Voltage`
column header, while the single-column one begins at the first blank line
AFTER Lead and Unit have been declared, because real exports also carry
blank lines mid-header.
Scalars in both waveforms now read either dialect through one helper, so
"-180.596" and "-180,596" give the same microvolts and neither is
privileged. A number carrying thousands grouping is refused rather than
guessed at, since resolving it needs the file's region and the file never
states it. The one genuinely ambiguous shape is a separator followed by
exactly three digits; it is read as a decimal, because observed waveforms
put two- and three-digit fractions in the same file and grouping is always
three, and because a machine writing one bare value per row has no reason
to group at all.
Every real export begins with a bare `Name` when the name field is blank —
Apple omits the trailing comma rather than writing an empty value. The
parser required a comma on every header row, so it rejected the file on
line 1, before it had read anything.
A valueless header row is now skipped. It contributes no metadata entry
rather than an empty-string one, because `Lead` and `Unit` presence is what
switches the parser into the single-column waveform mode and a key with no
value must not be able to trip that switch.
The relaxation is confined to the header. Inside the single-column
waveform a comma-less row is a sample and is read as one; inside the
paired waveform every row is lead and voltage, so a missing comma there is
still malformed.
CadenceKind and WeekdayToken were imported only to be re-exported. The
re-export went in the previous commit but the import survived, because
the doc comment above it still named both types at the time.
SENSITIVE_LEAVES and COACH_SUGGESTED_ACTION_TYPES each aliased a
constant defined elsewhere and were never imported. Neither is a
re-export the way the barrel lines were — the definition itself is the
dead thing, and both doc comments describe a consumer that does not
exist: the purity guard reads SENSITIVE_LEAF_IDS from ./catalogue
directly, and nothing mirrors the suggested-action allowlist.
The imports that fed them go with them.
Eighteen modules carried a convenience re-export of a symbol defined
elsewhere. In each case the callers the comment describes now import the
concrete module directly, so the line had no importer left. Where the
whole re-export block went, the comment explaining it went with it —
doctor-report-pdf-core, trend-card and medication-extract-prompt each
claimed to be a single import surface that nothing was using.
Definitions are untouched; imports that existed only to feed a removed
re-export go too.
Same shape as the derived and MCP barrels: the line in the barrel had no
importer because the consumer imports the concrete module. Affected:
cycle (15), coach tools (9), record-settings (8), links (4), data-list,
workout-detail and analytics/score (1 each).
analytics/score also carried CompositeValue on an explicit export line
above an export * of the same module, so the name stays exported either
way.
Definitions are untouched.
The barrel mirrored the tool, resource, prompt, write and scope registries;
its only consumer is the /mcp transport route, which needs three of them.
The registries are imported from their concrete module by ./server and by
their own tests, so the other 31 lines had no importer.
The barrel mirrored every export under src/lib/insights/derived. Of the
124 lines, 101 had no importer: consumers and per-metric tests reach for
the concrete module (./sleep-score, ./trajectory, …) directly. Keep the
cross-cutting surface — registry, dispatcher, baseline loader, and the
value shapes outside consumers read — and say in the header that the
barrel is not a mirror, so the next addition needs a consumer first.
Definitions are untouched; only the re-export lines go.
The classification map held three values. Apple's set is larger, so a real
verdict was being thrown away on English devices: "Heart Rate Over 120"
and "Heart Rate Under 50" both became null, and the column cannot tell
that apart from a recording that carries no classification at all.
The wordings are observed rather than remembered. The CSV writes a display
string and it does not match the framework's case names — the case
documented as InconclusiveHighHeartRate reaches the file as "Heart Rate
Over 120". Across a corpus of 128 real exports the field held five
distinct values; three of them mapped and two were dropped. The threshold
belongs to the watch generation, so the two heart-rate arms match the
number instead of pinning it, and the rest of the map takes its wording
from Apple's ECG instructions for use.
Classification is now split from the column write so the parser can tell a
verdict it has never heard of from one it knows it cannot represent.
Unrecognized is the second kind: the device did not fail to classify the
waveform, it received a verdict it does not know, so recording it as
"could not classify" would assert something Apple never said. Both still
store null, because that is the column's whole vocabulary, and no column
is added for a distinction nothing reads yet.
The old table pinned one row for "Localized or unknown verdict". That
assertion is still true, but it fused three different facts — a localised
verdict, an unknown one, and the two documented verdicts being silently
dropped beside them — and it read as "everything else is unknown" when two
of those were verdicts Apple documents. It is split into the four outcomes.
The language half stays open and is now named at the mapping. A German
export writes the KEY in German too, so a non-English verdict never
reaches this function; it fails the parse earlier.
The grounding tokenizer read English unit words and English or German
month names, while the reply is written in the reader's language. Two
silent failures came out of that, pointing in opposite directions.
A missing unit loosens the guard. "7,4 Stunden", "45 minuti" and "8000
pasos" came back unkinded, and an unkinded token clears the kind gate
against every ledger entry, so a fabricated night's sleep could ground
itself on a glucose reading that happened to share its magnitude. Kind
scoping is the thing that stops a weight grounding a blood-pressure claim;
for five languages it was not running.
A missing month corrupts correct prose. "Le 21 juillet" left the day
untyped, the plain-magnitude pass graded it, nothing matched, and the
soft-strip replaced it — the reader was shown "Le […] juillet". Spanish
and Italian also put a particle between day and month, so the pass admits
one.
The unit vocabulary had been written out three times and the copies had
already drifted; only one of them knew about "%". One list now feeds the
classifier, the range pass and the measurement pass. Polish months are
listed in the genitive, which is what a date uses and also the only form a
trailing word boundary can match, since that boundary is defined over
ASCII and "ń" is not part of it.
The bot's real commands are slash commands, and `telegram.helpBody`
documents exactly those in all six bundles, so nothing a user is TOLD to
type depended on their language. The undocumented courtesies did: the bare
word "help" and "hilfe" opened the help text and five greetings were
answered, everything else fell through to the intake keyword, failed it,
and returned without sending anything. Someone typing "aide" or "pomoc"
got silence, which is the worst possible answer to a request for help.
Both lists now cover the six shipped languages and are compared against
the folded message, so "cześć" and "czesc" are one entry and a capital
cannot miss. The match stays anchored to the whole message: "help me log
my weight" is prose the later paths should still see.
These are ordinary words rather than app labels, so there is nothing in
the bundles to derive them from.
The gate is "looks off-topic AND does not look health", and both halves
were English/German word lists. Two things followed. An off-topic ask in
French, Spanish, Italian or Polish was never recognised. And a genuine
health question in those languages that happened to brush an English deny
word was refused outright: "Ho una serie di misurazioni strane" trips
"serie", and nothing on the allow side knew "misurazioni", so a question
about the reader's own data came back as a refusal.
The allow side is no longer transcribed. Every metric label key and the
health nav labels already name these concepts in all six bundles, and
those labels are the nouns people type, so the vocabulary is read from
`messages/<locale>.json` and a seventh language is understood the day its
bundle lands. Matching is whole-word over a shared fold applied to both
sides. A subtractive list keeps label scaffolding out; French "bien-être"
alone would have contributed "bien" and "etre" and marked nearly every
French sentence as health.
The deny side stays hand-written — these are not app labels — and its new
arms match the folded message. A pattern anchored with \b cannot match a
word starting with "ż", because neither the preceding space nor the letter
is an ASCII word character, so there is no boundary between them.
The prompt-injection bank is untouched. It reads the person's own message
and the foreign-text path is fenced elsewhere; this is defence in depth,
not an open door.
The normative-claim bank held English and German words. Insight
recommendations are generated in the reader's locale, so for French,
Spanish, Italian and Polish every recommendation scored non-normative:
the annotation reported zero normative claims and zero uncited ones, and
the admin quality dashboard read that as full coverage rather than as no
measurement at all. A check that cannot fail is indistinguishable from a
check that passes.
The bank now has one arm per shipped locale and the caller passes the
generation locale; the reader's arm and the English one both run, because
a fallback provider answers a non-English reader in English. Matching
folds both sides through the shared label fold and anchors the leading
word boundary, so inflections still hit while the everyday Italian
"soprattutto" no longer counts as a comparative.
These words are the model's prose, not app labels, so they are not in the
message bundles and cannot be derived from them. The structural test
refuses an empty arm instead: a seventh language fails the suite rather
than quietly scoring nothing. An empty arm also used to compile to an
alternation that matched every string, which would have flipped the same
blindness to the opposite extreme.
The gate landed with a softer mode for accounts holding no second factor,
on the premise that they could not clear it. That premise was wrong. A
passkey is a possession factor here and the codebase already says so in
three places: a primary passkey login stamps Session.mfaVerifiedAt,
passkey sits in FRESH_FACTOR_METHODS so a passkey-proved elevation reaches
the fresh-factor routes, and the enrolment policy counts a primary passkey
as satisfying it outright. The only thing refusing such an account was the
reachability pre-check inside requireFreshMfa, which counts second factors
and nothing else.
So the mode is gone and the route asks for a full freshFactor: true with
proofSource: "any-possession" — a second knob that widens which credentials
count as "could ever prove this", and changes nothing about the gate. A
passkey-only account now gets a real gate with the re-provable
auth.stepup.required rather than no gate at all, and the dead-end
auth.stepup.mfa_not_enrolled is reserved for an account holding nothing.
hasSecondFactorEnrolled is deliberately NOT widened. It answers a different
question, and requireFreshMfaIfEnrolled asks it before deciding whether
account deletion, the data reset, the password change, the encrypted export
and key rotation gate at all — counting a passkey there would change five
destructive actions in passing. The two predicates sit beside each other
with docblocks saying why they differ, and collapsing them turns eight
existing tests red across three files.
Recovery is named rather than assumed. On Bearer: mint an elevation with
method "passkey". On the web: sign in again, which is what stamps the
session — there is no in-page re-proof for a cookie session because the
mint endpoint is Bearer-only by design. That limit is written into the
route docblock instead of left to be discovered, and the removal's refusal
now renders its own sentence in all six locales rather than the server's
English "second-factor" prose, which was wrong for the account this card
mostly serves.
Four surfaces assembled the correlation-discovery matrix and three carried a
different channel set. The Coach `get_correlations` tool had neither the
environmental-exposure channels nor the custom-metric ones, so asking it
whether air pressure relates to your sleep met a matrix the pair had never
been in. Its own file header said the opposite - "the SAME full-matrix scan
the route runs" - and had said it since the weather channels arrived in
v1.25. The period narrative was missing those two families and the
compliance / symptom pair on top.
`src/lib/insights/discovery-matrix.ts` is now the one place a matrix is
built. v1.30.3 hoisted the fetch discipline into a shared module for exactly
this reason and left the assembly duplicated behind; this finishes it. The
route, the per-metric card, the Coach tool and the period narrative all call
`assembleDiscoveryMatrix` and get the same families. What still differs is
declared as an option with its reason at the declaration: the measurement
read mode (a read-cost choice, left per surface), the extra metric types the
narrative's delta beat needs, and the RATED mood factors.
Mood factors stay on the narrative alone. That capture kind is frozen, the
two surfaces that persist a discovered pair write its channel key into a
plaintext column, and a custom factor's key is an opaque id - so admitting
them elsewhere is a decision about what to ask, not a parity fix.
Restore the per-metric filter on the assessment card's relations. It was
written in place of the pattern-dismissal filter rather than chained with it,
so a weight card's prompt had been receiving every surviving pair in the
record while the function name, the docstring and the early-return guard all
still said one metric.
`discovery-matrix-guard.test.ts` freezes who may fetch a channel, who may
construct one, and who may scan. Each matcher asserts a non-zero match count,
so a matcher that has stopped matching fails rather than agreeing with the
allowlist.
The statistics are untouched - same Pearson, same exact p-value, same
Benjamini-Hochberg, same floors. The Coach's family is larger now, which is
the point: it is corrected over the same family the insights page is.
A restore's first act is to delete the class of records it is about to
rebuild, so a payload whose documents section is simply absent did not
"restore everything else". It emptied the vault, reported success, and left
the operator to find the hole later.
The whole file is refused instead, above the transaction that wipes the
account, and the refusal names the section and says nothing was changed —
an operator holding the only copy of an account cannot act on "restore
failed". The sections ride the error envelope as meta.errorCode
backup.section.missing, and the admin console holds them in a panel beside
the skip report rather than a toast that outlives nothing.
The manifest is the discriminator, and only the manifest. A portable export
leaves the mental-health screeners and the consent receipts out on purpose
and declares both as omitted; those files still restore. So does a file
whose section is present and empty, which is what an account with no
documents honestly writes, and so does every backup written before the
manifest existed. The integration test proves the distinction with the same
builder and the same deleted keys, differing only in what the manifest said.
The route is deliberately outside the published contract — requireAdmin is
cookie-only — so the OpenAPI document is unchanged.
DELETE /api/auth/passkeys/{id} took a plain cookie session or a wildcard
Bearer, while DELETE /api/auth/me/mfa/webauthn/{id} beside it demanded a
fresh factor proof. The softer gate sat on the primary sign-in credential:
the last-method check stops an account locking itself out, but nothing
stopped a hijacked session stripping the credentials the owner signs in
with.
The route now goes through requireMfaManagementAuth, the same gate the
security-key removal uses, in a new "if-enrolled" mode. An account with a
second factor pays the full price on either transport — a freshly verified
cookie session, or a factor-proved X-Step-Up elevation on Bearer, so the
native client keeps the capability rather than being sent to the website.
An account with no second factor keeps the contract it had, because a gate
it can never satisfy is not a safeguard but a locked door in front of its
own credential list.
The elevation is spent immediately before the delete, so a foreign id or a
refused last credential does not burn a proof that then has to be minted
again. The route is added to the elevation allowlist with its reason, and
the either-factor enrolment test that was written out three times in the
API kit is now one function.
DELETE /api/user/thresholds without ?metric= erased the account's whole
override set on an omitted query parameter alone: no confirmation, no dry
run, and no rate limit of its own while the PUT beside it had one.
The wide form now wants { "confirm": "RESET_THRESHOLDS" } in the body — the
same typed-confirmation shape /api/settings/account and /api/settings/data
already use — and refuses 422 without writing anything. The ?metric= form
is untouched: it resets one band the person can retype, and the targets
sheet fires it in a loop. Both forms now run the PUT's 30 / 5 min limit.
The settings card's "reset all" control already opened a confirm dialog;
it now sends that answer to the server too. Its token is compared against
the route's declaration in the tests, so the two ends cannot drift apart
into a permanent 422.
Both named /api/doctor-report, a family removed some releases ago. The
contributor doc sent a reader to a POST endpoint that answers nothing,
and the renderer's own header described itself as serving an endpoint
that is gone.
Found by the sweep behind the retired-route registry, which asks what
this repository still references rather than what a client calls. Two
hits, both prose, neither a call.
The three locale matchers were built in parallel and each grew its own
fold. Two of them are the same function: the side-effect tag fold and
foldForMatch agree character for character, so one now calls the other.
The third stays. foldLabel decomposes with NFKD, maps sharp s to ss and
joins with underscores; foldForMatch decomposes with NFD, leaves sharp s
alone and joins with spaces. That is a disagreement about who typed the
text, not an accident: one side of a lab comparison is a shipped label
retyped or prose a lab printed, where tolerating a different spelling
claims more than the fold can prove, while someone naming a metric to an
assistant is composing rather than reproducing. Both docblocks now say
so and name the other.
Three resolvers turned a user-authored word into an internal code by
comparing it against a hand-typed English list. The app ships six
languages, so five of them resolved nothing: `Fer` and `Żelazo` came back
as `unknown_nutrient` while the settings card rendered exactly those
words, and `poids` and `sommeil` reached no series at all.
Stop transcribing. `nutrients.names.*` and `measurements.type*` already
label these concepts in every bundle, so the match index is derived from
`locales` instead — a seventh language works the day its bundle lands,
with no edit at any call site. The nutrient tool's English labels now come
out of `messages/en.json` too, so the wire wording cannot drift from the
screen.
The MCP contract is unchanged in shape and stays protocol-level English.
What moves is which words are understood: `unknown_nutrient` /
`unknown_metric` still mean "could not place this word", distinct from a
placed name answering `no_data`. Nothing turns absence into data.
Two findings the sweep did not predict:
- `Puls` resolved to pulse-wave velocity, because the English
display-name pass matches substrings and "pulse-wave velocity" contains
it. An exact name now beats a loose English fragment.
- French and Italian use one phrase for body fat and fat mass, and three
bundles use one for fat-free mass and lean body mass. A word two
metrics claim resolves to neither; answering kilograms to someone who
asked for a percentage is worse than saying the word was not
recognised. The known set is pinned so a new collision surfaces.
The medication needles are the exception: they are drug names, with no
bundle to derive from. They fold to a normalised INN stem rather than to
a language, so `Metformin`, `Metformine`, `Metformina` and `Metforminum`
are one entry, and `Semaglutyd` and `Atorwastatyna` cost nothing to add.
A guard asserts no two molecules collapse onto one stem. Iron is the one
class with no stem to fold to — it is named for the element — so its word
list is hand-tabulated and says so.
Every English resolution that worked before returns the same answer.
The derived names extended the analyte class by list index. Reordering the
signal list would have re-pointed them at the units or the report header with
nothing to notice, so the class is a named constant and the check compares
against it.
The gate that decides whether a transcribed document is worth an extraction
call scored four classes of evidence and needed two. Three of them — the
reference-range label, the analyte names, the report header — were written in
English and German, so a French, Spanish, Italian or Polish report scored at
most one: the units, which are the same everywhere. Auto-staging never fired
for it. The document was lost from the flow silently, because the manual
extract button stays available and nothing reads as broken.
The text is folded once — lower case, accents dropped — and every pattern is
matched against that, which on its own recovers a good deal: "Cholestérol",
"Créatinine" and "Emoglobina" reach the existing stems as soon as the accents
are out of the way. The reference-range labels and report headers of the four
locales are added by hand, because they are prose a lab printed and the app
owns no string for them.
The analyte names ARE derived, from `labs.catalog.<slug>` across every shipped
bundle: those are names the app itself displays, and a new locale should not
need an edit here to be recognised. Names under four folded characters are
dropped — "ALT" is German for "old", and a three-letter alternative in a gate
this cheap would make every German letter a lab report. The acronyms worth
having are in the hand-written half, where each was considered on its own.
Which way this errs, stated at the code: a false negative loses a whole
document from auto-staging, while a false positive costs one extraction call
and stages facts a person then declines — nothing commits either way. So the
vocabulary is widened, and the two-distinct-class rule holds the other side.
The derived names extend the analyte class rather than forming a fifth one, so
a document naming one analyte in two languages still scores one. Tests pin
each addition on a case that fails the gate without it, and pin the prose that
must keep failing it.
The lab form offers "negative" and "positive" as quick picks and writes the
label it rendered, so the column holds the bundle's own string in the language
the reading was recorded in. The SNOMED mapper matched that string against a
list transcribed in English and German, so "négatif", "negativo", "ujemny",
"positif" and "dodatni" all exported without a code — text-only on the one
document meant to be machine-readable.
The negative and positive arms read `labs.form.qualNegative` and
`labs.form.qualPositive` out of every bundle now. The form writes those exact
keys, so the write side and the read side cannot drift, and a new locale is
coded the day its bundle lands.
The detected / not-detected pair stays hand-written English and German, and
that is the honest limit of this change: the app ships no string for those
terms, they are prose a lab printed and a person copied in, and coding "non
détecté" or "niewykrywalny" would mean inventing clinical vocabulary in four
languages nobody here can check. This module's own rule is that an unverified
code is worse than an honest `.text`. When they need coding they need a quick
pick and a bundle key, which is the route the other two arms took. A test pins
that they stay uncoded rather than leaving it to be discovered.
Borderline stays uncoded as before; a test now pins that the derivation did
not sweep its bundle key up by accident. Comparison folds accents as well as
case, so a result typed without them still resolves.
A biomarker minted from the labs catalog stores `t("labs.catalog.<slug>")` —
the name in the language of whoever minted it — and that stored name is what
reaches the FHIR exporter as free text. The exporter matched it against an
alias table written in English and German. So an account working in French,
Spanish, Italian or Polish exported "Cholestérol total", "Glucosa en ayunas",
"Glicemia a digiuno" and "Cholesterol całkowity" as text-only concepts with no
LOINC code, on the one document whose whole purpose is being read by another
system.
The names are DERIVED from `messages/<locale>.json` now, not transcribed into
this file. Both sides of the new slug map are app-owned identities with no
language in them, which is what makes it the one edit a new locale does not
need: `locales` grows, `messages/xx.json` lands, and the analyte names it
ships resolve the same day. Transcribing four more languages would have fixed
today's six and reopened the hole at the seventh.
The hand-written aliases keep precedence. They cover what the app owns no
string for — clinical abbreviations, the ways a person shortens a name — and
they were verified one at a time over several releases. A derived name landing
on the same key is a harmless duplicate; one landing on a different key would
be a translation typo re-pointing a verified mapping at another analyte, so
the build refuses it and the guard suite asserts the refusal held.
The lookup fold now drops accents before stripping punctuation. Without it a
name carrying diacritics lost them entirely rather than reducing to its base
letters — "Cholestérol total" folded to `cholestroltotal`, and "Hämoglobin" to
`hmoglobin`. That last one means the German umlaut spellings never resolved
either: the hand-written `hamoglobin` and `nuchternglucose` aliases only ever
caught the un-umlauted typing. Every pre-existing alias is ASCII, so the fold
is a widening and what resolved before resolves to the same key.
The free-text reference-range parser decided what a printed window meant from
an alternation of German and English words. The app ships six locales, and a
lab prints in the language of the country it stands in, so `jusqu'à 5,0`,
`hasta 5,0`, `fino a 5,0` and `do 5,0` all fell through to `low: null,
high: null`. The string was stored, the reading was then judged against the
biomarker's own band or against nothing at all, and no surface anywhere said
that the lab had printed a window the app could not read.
The vocabulary is TABULATED here, not derived from `messages/<locale>.json`
the way an app-owned word list would be. These words are prose a lab printed
on paper; the app owns no string they could be read out of, and the language
of a report has nothing to do with the language of the UI. So the table is
hand-written, keyed by the shipped `Locale` union so that adding a locale
fails to compile until someone has answered how that language prints a range,
and every locale's words are tried against every input at once. The cost is
written down at the table: a report in a language absent from it stays
text-only.
Three things had to change beside the word lists. Introducing words are now
stripped only on a second pass, because Italian "da", Polish "od" and French
"de" also state a floor on their own and stripping first read "da 3,5" as a
bare number. Input and table are diacritic-folded and their apostrophes
normalised, so a table row written "jusqu'a" meets a report's "jusqu'à" and
the curled apostrophe a word processor emits. And one-letter separators ("a",
"e", "y") are only accepted with whitespace on both sides, so "3.5e5" is not
read as a window. Decimal commas already worked and are pinned by the new
cases.
An unreadable window now has a name. `isUnreadableRange` separates "the report
stated nothing" from "the report stated something nobody could read" — the two
were indistinguishable at every call site, which is what let the French case
pass unnoticed. The manual form already showed the reader that the text was
kept but not read; the two ingest paths now count it into their wide event, so
an unsupported notation surfaces as a number rather than as silence.
The German and English lists are the previous alternation split word for word,
and the existing suite covering them passes unchanged.
Three surfaces, one source. The registry the proxy answers 410 from now
also emits the tombstones into `docs/api/openapi.yaml` and rides on
`GET /api/meta/capabilities`, so a retirement cannot be honoured at
runtime and invisible everywhere a client could have read it first.
The contract publishes each retired path with 410 as its only response,
the removing version, and the replacement. `deprecated: true` is
imprecise — the word means "still works, stop using it", and these do not
work — but it is what makes a generator warn at the call site, and the
description carries the exact truth the flag approximates. `security: []`
because the answer is produced before any handler and reads no
credential; publishing it as authenticated would tell a client to sign in
to be told a path is gone.
The capability read carries the same list. That surface is the one a
client already polls on launch, so a retirement can reach it before the
failing call rather than through it.
The coverage guard gains a third state. Published and exempted both
describe a route that exists, so removing one used to leave exactly one
legal move — erase it from the contract as well — and that move is
invisible to every client already calling it. A registered path is now
required to be absent from `src/app` AND present in the contract, which
is the one combination the phantom-path case rejected outright. Both ways
out are named in the failure message.
Its docblock says what the state cannot do, because the first draft
overclaimed: a check over one commit's state cannot see a deletion, so it
cannot insist that a removal should have been registered. Proving that
needs the previous release tag, and the job running this suite checks out
at depth one with no tags, which would make the check one that silently
cannot fail. What is held instead is the registration's integrity, in
three directions and with the mutation for each written down.
The sharing-denial guard learns the same state: a retired path resolves
no record fence because it reaches no handler.
A route removed on purpose still leaves a client calling it. As reported
in iOS #94: between 2026-08-08 and 2026-08-20 the native client asked
`/api/auth/me/research-mode` on every GLP-1 detail screen, got a bare
404, and painted a permanent error card with a retry button aimed at the
same removed path. Twelve days, on every reachable instance including the
public demo, which was built after the removal. Deleting the route was
right and the removing commit reasons it out in full; the answer said
nothing. A 404 from a deployment is indistinguishable from a broken
build, a reverse proxy pointed at the wrong upstream, or a typo in the
caller's own path, so reading it as transient and retrying was the
correct reading of the only evidence there was.
One registry, `src/lib/http/retired-routes.ts`, carrying the path, the
release that removed it, the replacement where there is one, and one
sentence on why. Every other surface reads it.
410 over a 404 with an error code because it is the status whose entire
meaning is "this existed here and is gone": a client that understands
nothing else about this API still learns not to retry. The code is not
dropped, it rides in the same envelope with `removedIn` and `replacedBy`
for a client that reads further. The status is the floor, the code is the
detail.
Answered in the proxy, before routing, rather than by a catch-all route
module: a catch-all under `/api` would intercept every unmatched path and
make the ordinary 404 ours to manufacture, which is a far wider blast
radius than the problem. The proxy already owns the exact-path decisions
of this shape. It reads nothing off the request but the path, so an
anonymous caller and an authenticated one get identical bytes and the
tombstone cannot become an authentication oracle.
Seeded from the removed-route set: every route path under `src/app` at
every release tag, unioned and differenced against the current tree.
Twenty-seven paths have been removed over the project's history; the five
registered here are the ones that were published or client-facing, and
each removal is pinned to its release by the route's presence at one tag
and absence at the next. Whether any client still calls one of them is a
separate question that evidence in this repository cannot answer, and the
registry deliberately makes no claim about it.
The GLP-1 side-effect chips in the mood form write the localised label into
the flat `MoodEntry.tags` column, and three readers matched that label against
a hand-written list of English and German words: the therapy timeline, the
Coach snapshot, and the doctor report's GLP-1 block. The app ships six locales.
Tap the nausea chip with the UI in French, Spanish, Italian or Polish and the
day never reached the timeline, the Coach grounded its answers on an empty
side-effect list, and a clinician read an empty table over a record that had
three months of symptoms. Nothing on any of those surfaces said the table was a
filter rather than the truth. The three lists had also drifted apart from each
other.
The chips are a closed set with stable keys, so matching the label was the
defect and the key is the fix. `glp1-side-effect-tags.ts` holds the catalogue —
seven keys, each with the i18n key its chip renders — and the mood form now
draws its strip from it, so the write side and the read side cannot diverge
again. `glp1-side-effect-tag-match.ts` builds the label-to-key index by reading
`messages/<locale>.json` for every entry in `locales`, rather than transcribing
the labels into a second list. Transcribing would have fixed today's four
locales and reopened the same hole on the day a seventh is added; deriving
means a new bundle is matchable as soon as it lands, with no edit in the
matcher. Comparison folds case, accents and separators, so the English and
German tags already in the database keep resolving and a straight apostrophe
matches the typographic one the French bundle uses.
All three readers now report the catalogue key instead of the stored string.
The same recorded symptom yields one key whatever language it was tagged in,
which also fixes the mixed-language output an account gets after switching UI
locale; the doctor-report PDF resolves the key through the label map so the
clinician reads it in the report's language.
A tag outside the catalogue is still never counted — a free-text mood tag must
not reach a doctor as an invented symptom, and the rest of somebody's day does
not belong on a therapy timeline. It is no longer dropped without trace: the
timeline route annotates the unresolved count on its wide event, so the
classification rate stays answerable while the tag text stays in the database.
The Coach's copy of the list also carried `vomiting`, `reflux` and `erbrechen`.
Those are removed. No chip writes them, so they only ever matched a hand-typed
tag in two of six languages; keeping them would have rebuilt the defect inside
its own fix. Restoring either symptom means giving it a chip and a label in all
six bundles.
The response schema for `GET /api/insights/glp1-timeline` now enumerates the
keys, and its description no longer documents the bilingual match as if it were
the contract.
Three places where the copy promised more than the handler delivers.
The hero's personal-baseline line rendered unconditionally while both its
siblings in the same meta row gate on their own data. It claims the subtitle
above it was drawn from the reader's own last 90 days, but with no briefing the
subtitle is fixed fallback copy computed from nothing — so an account with no
briefing at all was told its generic sentence rested on a personal baseline.
Gate it on the briefing, like the rest of the row.
The pulse sub-page promised a personalised Karvonen target band. Karvonen is
not computed anywhere: the band comes from CDC/NCHS population percentiles by
age and sex, and falls back to a fixed 60-100 for a reader with no age on file.
The copy now says that in all six locales, and the fallback's own source label
says it is a general adult range rather than a personal one, because that arm
is the same band for everyone. Karvonen itself stays unimplemented; whether to
build it is a product decision, not a copy fix.
The Coach's `allTime` window returns 365 days, identical to `lastYear`, and it
is the default. "year so far" was wrong too — a rolling 365 days is not
year-to-date. Both labels now say what the reads do. The two options remain
functionally identical and one of them should probably go, but removing a
default is a product decision.
A rollup band outlives the rows it was folded from. The read-swap falls back to
live SQL only when a band is EMPTY, so a bucket that is merely wrong is served
as current; `ensureUserRollupsFresh` repairs the DAY tier over the trailing 90
days only, and keys off `Measurement.updatedAt`, which a hard delete does not
bump. The WHOOP body sync hard-deletes a WEIGHT row with no invalidation, so a
MONTH bucket keeps reporting a value the record no longer holds.
The block was handing the model two irreconcilable accounts of the same record
and marking neither: the coarse tail said one thing, the all-time extremes read
live beside it said another, and `asOf` stamped the whole thing current because
it is derived from raw reading age and knows nothing about rollup age.
Reconcile them where they are assembled. A bucket mean is an average of real
rows, so it cannot fall outside those rows' own all-time extremes — when it
does, the band is describing readings the live read cannot see. That is an
invariant rather than a heuristic, so the disputed tail is dropped and
`asOf.coarseHistoryWithheld` records that it went. Narrating a ghost is worse
than narrating less.
`currentForTodayClaims` stays keyed to raw recency, which is what it means. The
surviving readings really are from today, and suppressing a true present-tense
statement to punish a stale history band would trade one dishonesty for another.
The bucket on disk is still wrong, and a wide event now says so. The deeper fix
is to make the tier notice — compare a bucket's `computedAt` against the type's
in-window MAX(`updatedAt`), or invalidate on the hard-delete path — which is a
larger change than this repair.
The numeric grounding guard keyed its activation off `loop.toolResults` alone.
A tool that finds nothing resolves to `{ present: false, reason: "none" }` with
no `available` payload, `runCoachToolLoop` drops that shape, and the no-tools
snapshot fallback is populated only in the non-tool branch — so the guard sat
out every turn where the record was empty, which is precisely the turn where a
fabricated figure is likeliest and does the most harm. Ask about a metric you
have never logged and an invented number was persisted and streamed verbatim,
with no withheld-figure marker to say it had never been checked.
`toolTrace` is the discriminant, because it records the pure misses the results
array drops. Tools ran and every one missed: grade the reply — an empty ledger
there is not missing information, it is the positive statement that nothing can
be cited, so `findUnverifiedCoachNumbersInLedger` takes an explicit opt-in to
grade against it rather than no-opping. The model answered without calling a
tool: still dormant, unchanged, because the tool-mode base prompt carries no
pre-computed figures and flagging a legitimately recalled one was the v1.32.1
regression. The inventory counts ride along as a widener so a plain count
restatement stays grounded.
Eliding the digits is not enough on an empty record. "Your sleep averaged […]
minutes, up from […]" still asserts a series, an average and a direction for a
metric with no readings; only the precision goes, the false claim stays. So a
turn that missed every tool and lost a figure to the guard is replaced with the
honest answer, in all six locales. Scoped so a reply whose figure reconciles
against a prior turn or an inventory count is left alone.
Three characterisation tests for grounding gaps found while verifying that
every figure the Coach states comes from the record.
The numeric grounding guard rewrites any figure the turn's Grounding Ledger
cannot account for, and it holds: removing the strip reddens the existing
cross-turn ledger test. What is not covered is its activation. On the tool
path the guard runs only when `loop.toolResults` is non-empty, and a tool that
found nothing resolves to `{ present: false, reason: "none" }` with no
`available` payload, which the loop drops. So the guard disarms itself exactly
when the record is empty: ask about a metric with no readings, and an invented
figure is persisted and streamed verbatim, with no withheld-figure marker to
tell the reader it was never checked. The no-tools snapshot fallback is
populated only in the non-tool-mode branch, so the tool path has nothing to
fall back to, and every cloud provider runs the tool path.
Separately, a rollup bucket outlives the rows it was folded from. The read-swap
falls back to live SQL only on an empty band, `ensureUserRollupsFresh` repairs
the DAY tier over the trailing 90 days only, and it keys off
`Measurement.updatedAt`, which a hard delete does not bump. The integration
test reproduces the hard `deleteMany` in the WHOOP body sync against a
MONTH-band bucket: the deleted value keeps reaching the Coach's coarse tail
while the all-time extremes beside it report only the surviving value, and the
block is stamped current because `asOf` is derived from raw reading age and
knows nothing about rollup age.
All three assert current behaviour and carry a docblock saying so. Closing
either gap will redden them, which is the point.
The generated document is the whole visible effect of both halves, so it
conflicted line for line. Regenerated from the merged registry rather
than resolved by hand.
`.meta()` clones the schema in Zod 4 and registers the clone, so a bare
`schema.meta({ id, description })` statement registers nothing at all.
Forty-one of them sat across nine route modules. The ids they named —
`MeasurementType`, `MeasurementSource`, `LoginPasswordRequest`,
`CreateMeasurementRequest`, `CreateBatchWorkoutRequest` and thirty more —
reached `components.schemas` zero times, the descriptions written for them
were published nowhere, and every use site inlined an anonymous copy. A client
generated from the contract got duplicated nameless types where one named
model belongs, which is the part that matters, because the iOS client is why
the contract is published.
Forty of them now bind their clone, and every use site points at it. Where the
schema comes from `src/lib/validations/*` the import is aliased and the clone
takes the name the use sites already spell, so nothing mutates the runtime
validation schema. The seven that cross module boundaries are exported from
`./shared` and the eight consumers import them from there; importing the raw
schema from the validation module instead is exactly what inlined them.
Thirty-four named components appear where inline copies stood. Fully
dereferenced, the document differs by eighty-six added descriptions and
nothing else: no operation's request or response shape moves. `CoachPrefs`
was already force-registered in the index and finally carries its description.
Five ids legitimately stay out of `components.schemas` —
`ListMeasurementsQuery`, `ListAllergiesQuery`, `ListEncountersQuery`,
`SuggestEncounterQuery` and `ListPractitionersQuery`. The emitter expands a
`requestParams.query` object into inline `parameters`, so a query id has
nowhere to land; they stay in the assigning form for consistency.
Three findings the wiring turned up:
`CreateShareLinkRequest` and `HealthRecordExportRequest` were annotated in the
Coach module, which imported the two schemas for no other reason and never
referenced them. Both operations that send these bodies live in the
health-record module, so the annotations move there. Both descriptions
described a `sections` field: the export body has no such field and rejects it
with a 422, and the share-link body freezes scope through `selection` plus
`documentIds` / `documentOnly`. Publishing either sentence verbatim would have
published a lie, so both are corrected to the shape the routes actually take.
`CoachPlansListQuery` — the forty-first — named a component that nothing
referenced and that a query object could never reach. The operation publishes
its two parameters by hand, because the schema carries a cross-field
refinement the expansion cannot express, and every clause of the annotation is
already published on the operation and on both parameters. It is deleted
rather than left registering nothing.
The annotation was already in the assigning form and still never reached
the document, because both consumers extend the base rather than
reference it and .extend() builds a new object. Same shape as Medication
and CoachPrefs, same remedy: name it in the forced-registration block.
Zod 4's `.meta()` clones the schema and registers the clone, so a bare
`someSchema.meta({ id })` statement registers nothing at all. Thirty-nine of
those across eleven route modules published neither the name nor the
description written for it: the schema inlined anonymously at every use site,
and a client generated from the contract got a duplicated nameless type where
one named model belongs. That last part is the cost that matters, because the
native client is the reason the contract is published.
Bind each annotated clone to a const and point the route table at that const.
Twenty-nine ids now emit under `components.schemas`, referenced from their
thirty use sites. Seven stay absent by design: they annotate
`requestParams.query` objects, which the emitter expands into inline
`parameters` rather than a component. `MoodTagLayout` stays absent too, for
the reason `Medication` and `CoachPrefs` already carry a forced registration
in the components block: its only consumer is an `.extend()`, and extending
builds a new object rather than referencing the base.
Two annotations in the illness module named ids that no route ever referenced,
`GetIllnessDayLogQuery` and `ListIllnessDayLogsQuery` — the day-log path
publishes the merged `GetIllnessDayLogsQuery` instead. They were dead twice
over, so they go rather than get wired to nothing.
`docs/api/openapi.yaml` grows the twenty-nine components and their refs. The
only other movement in it is thirty descriptions appearing where the text had
never reached the document. No request or response shape changes.
Marc-André Bombeck