# Changelog ## [1.5.4] — 2026-05-28 — Medication wizard: modal dialog, compose-mode, ten-bucket taxonomy The v1.5.3 creation wizard shipped as a single-page card with seven inline steps. Patient feedback during the first day of live use was unanimous: the page felt dense, the step-by-step intent was lost, the "every N days from my last injection" cadence was unintelligible without a worked example, and the edit form was visibly wider than its container. The bi-weekly worker bug was closed in v1.5.3 but the surface a patient actually touches was not yet where it needed to be. This release replaces the wizard with a real modal-dialog flow, lands compose-mode so a single medication can carry multiple parallel schedules (the insulin short-acting + long-acting case), widens the treatment-class taxonomy to ten buckets, and retires the flat edit form. ### Added - **Modal-dialog wizard**. `src/components/medications/wizard/MedicationWizardDialog.tsx` is the new entry surface — a shadcn `` on desktop, a `` on mobile, switched through the existing `` primitive at 768 px. The wizard renders its own header inside the body so the iconified step anchor sits where the patient reads, and the footer carries Back / Save with a fixed sticky position on mobile. Eight steps total, one focused question each: name → treatment class → dose → course window → cadence → sub-cadence detail → times of day → review with reminders toggle and per-schedule summary cards. Cadence + sub-cadence + times collapse out of the path when they do not apply (one-shot walks five, daily walks seven, recurring walks eight); the visible counter mirrors the path the patient actually walks. - **Compose-mode**. A medication can now hold multiple parallel schedules, each with its own cadence and times. Steps 5–7 are per-schedule; Step 8 renders the schedule list with an "Bearbeiten" and "Entfernen" action on each card plus a "+ Weiteren Zeitplan hinzufügen" card at the bottom. The last remaining schedule is non-removable so the medication always carries at least one schedule. A medication opened in edit mode with more than one configured schedule lands the patient directly on the list view; single-schedule edits keep the Step 1 entry. The header on Steps 5–7 carries a small "Zeitplan {n} von {total}" caption when more than one schedule is configured. - **Ten-bucket treatment-class taxonomy on Step 2** — Blutdruck, Diabetes, Hormone, GLP-1-Injektion, Schmerz, Allergie, Vitamine, Nahrungsergänzung, Antibiotikum, Sonstiges. Schmerz lands the chronic-pain cohort with its own rolling-rescue-medication pattern; Antibiotikum lands the one-shot course pattern cleanly so neither falls through to `OTHER` for analytics purposes. Each row carries a monochrome Lucide glyph (Stethoscope / Droplet / Activity / Syringe / Flame / Wind / Apple / Leaf / ShieldCheck / Tag). - **`DIABETES` + `ANTIBIOTIC`** as first-class values of the `MedicationCategory` Zod enum so insights routes that filter on the clinical category surface those buckets without falling back to `OTHER`. Additive at the schema layer; existing rows do not need touching. - **Rolling-cadence mental-model copy.** Step 5's "Flexibel ab letzter Einnahme" row carries the example sentence directly under the label (not behind a tooltip) — "Du lässt den nächsten Termin offen, drückst 'genommen' wenn du sie genommen hast — ab dem Zeitpunkt zählt der Counter wieder." The patient-pain research across chronic-illness adherence studies identified this exact gap as the #1 reason patients abandon medication apps in week one; the in-line example sentence closes it. ### Changed - **Edit path routes through the same dialog.** "Bearbeiten" on a medication card opens ``; the header swaps to "{name} bearbeiten", the CTA to "Änderungen speichern", and the entire payload hydrates from the existing medication shape including the schedule `id` so a PUT preserves the per-schedule identity. The flat `` (1314 LOC, surfaced as a `` that rendered wider than its container on the medications list page) retires in the same change. - **`/medications/new` retires** to a redirect that opens the dialog from the list page (`/medications?new=1`). Existing bookmarks survive without losing the entry point. - **i18n.** A clean `medications.wizard.*` namespace replaces `medications.create.wizard.*` so the locale-integrity guard surfaces every dangling key during the cut. German and English carry native copy; Spanish, French, Italian, Polish ship the English string verbatim as a machine fallback per the project convention. Native polish for the four fallback locales follows. - **``** (v1.5.3's inline-stepped Card) and **``** (an unused titration-phase helper that `knip` flagged once the list page swapped over) retire alongside the form. ### Fixed - **Multi-schedule data-loss risk** — the v1.5.3 wizard collapsed a multi-schedule medication to its first schedule on save and silently dropped the rest. Compose-mode is the proper fix; the encoder now emits every schedule, the hydrator reads every schedule, the per-schedule `id` round-trips so the PUT preserves identity. - **The edit form being wider than its container** — Marc's specific complaint from the v1.5.3 hand-walk. The container is now a constrained Dialog at `sm:max-w-md` on desktop and a sticky-footer Sheet at `max-h-[90dvh]` on mobile; the form-as-page surface is gone. ### Tests - `src/components/medications/wizard/__tests__/wizard-payload.test.ts` — 54 pure-helper cases pinning `validateStep`, `buildCreateBody`, `buildUpdateBody`, `summariseCadence`, the treatment-class row → request body mapping, the multi-schedule encoder, schedule-`id` preservation on edit, the hydration that lands on the list view when more than one schedule is present, the remove-guard against the last remaining schedule, and the add-then-active-bump behaviour. - `e2e/medications-wizard-{daily,weekdays,biweekly,monthly,rolling,oneshot}.spec.ts` — six Playwright specs updated to walk the new dialog surface via the German label set; `e2e/medications-wizard-compose.spec.ts` walks a two-schedule create flow (daily Ramipril + weekly-Wednesday addendum) and asserts both summary cards land on Step 8 before save. - Route tests assert POST `/api/medications` accepts a body with `category: "DIABETES"` and `category: "ANTIBIOTIC"`. ### Notes - **No API or wire-format change.** The wizard writes the same `createMedicationSchema` payload the v1.5.3 wizard wrote; iOS clients reading and writing the existing schedule shape are unaffected. The native `treatmentClass` enum already shipped `GLP1`; the wizard's Step 2 maps the Marc-confirmed labels onto the existing enum + the two new `MedicationCategory` values. - **Cadence-engine read-flip is still v1.5.x deferred** — the today-projector, the cadence chart, and the medication card continue to read the legacy `daysOfWeek` / `intervalWeeks` columns. A wizard-minted rolling or one-shot medication will render a daily-looking next-due chip on the dashboard until the read-flip ships. The reminder worker and the engine work with the new fields correctly. - **Test totals.** 5500 unit + 1 skipped, 261 integration + 3 skipped, 14 Playwright wizard instances. Typecheck, lint, `openapi:check`, i18n call-site coverage, locale-integrity all green at HEAD. ## [1.5.3] — 2026-05-28 — Medication scheduling: RRULE cadences, rolling intervals, one-shot lifecycle, creation wizard The medication surface in v1.5.2 covered daily and weekday-subset schedules cleanly, but everything else — bi-weekly, monthly, quarterly, yearly, "every N days from my last injection", single-dose appointments — either drifted or was unreachable from the UI. The reminder worker also carried a quiet pre-existing bug where `intervalWeeks` was ignored: a schedule meant to fire every other Wednesday fired every Wednesday. This release lands the full cadence surface, closes the worker bug behind a regression test, and adds a step-driven creation flow patients can walk without consulting a manual. ### Added - **Four new cadence shapes**, modelled at the schema level with explicit columns instead of an overloaded legacy string: - **Calendar-anchored RRULE patterns** (`rrule TEXT`) — `FREQ=WEEKLY;INTERVAL=2;BYDAY=WE`, `FREQ=MONTHLY;BYMONTHDAY=1`, `FREQ=MONTHLY;INTERVAL=3;BYMONTHDAY=10`, `FREQ=YEARLY;BYMONTH=1;BYMONTHDAY=1` all expand correctly. Powered by the `rrule` npm package; the engine adds an `UNTIL` suffix derived from the medication's `endsOn` automatically (and skips that suffix when the user's RRULE already carries `COUNT` or `UNTIL`, so the two never collide). - **Flexible-rolling cadence** (`rollingIntervalDays INT`) — "every N days from the last logged intake". The next-due date re-anchors when an intake is logged; skipped doses pause the schedule until the next real intake. The driving use case is the GLP-1 weekly injection where the calendar Wednesday does not match the user's actual cycle. - **One-shot single-administration** (`oneShot BOOLEAN`) — vaccines, post-op final doses, anything with one scheduled occurrence and an auto-deactivate after the dose is logged. - **Course window** (`startsOn DATE`, `endsOn DATE`) at the medication level. Required for one-shot; optional everywhere else. The reminder worker stops minting slots past `endsOn`. - **`POST /api/medications/extract`** — natural-language extraction for the wizard's first step. A user types "Mounjaro 5mg weekly Wednesday morning starting next Monday" and the route returns a structured payload that pre-fills the wizard. Rate-limited per user, budget-gated, and carries a citation-coverage guard so the extraction cannot return a name or dose the user did not write. - **`/medications/new`** — the seven-step creation wizard (compressed to five on the one-shot path) replacing cold-start entry into the legacy flat form. The summary step interpolates the actual picked weekdays / day-of-month / yearly date so the patient confirms the specifics rather than a category label. - **Reusable picker primitives** under `src/components/medications/scheduling/`: `CadencePicker` (eight cadence kinds, mode-aware via `allowedKinds`), `TimesOfDayChips` (one or more `HH:mm` entries with morning / noon / evening / night presets), `CourseWindowRow` (start + optional end with `lockEndsToStart` for one-shot). Composed by both the wizard and the refactored edit form so any future cadence tweak reaches both surfaces at once. - **Canonical recurrence engine** at `src/lib/medications/scheduling/recurrence.ts` exposing `occurrencesBetween`, `nextOccurrenceAfter`, and `matchesInstant`. The reminder worker routes through the engine via a narrow `worker-helpers.ts` adapter; the today-projector, the cadence chart, and the medication card continue to read the legacy fields through v1.5.x and switch over in the v1.5.4 read-flip. - **Migration `0081_v15_medication_scheduling`** — adds the new columns, backfills `rrule` and `timesOfDay` for every existing schedule (closed-enum regex on the legacy `daysOfWeek` shapes, ELSE NULL fallback), and stages a CHECK constraint forbidding both `rrule` and `rolling_interval_days` populated on the same row. - **OpenAPI coverage** for every medication route — `GET / POST / PUT / DELETE /api/medications`, `GET /api/medications/{id}`, `POST /api/medications/{id}/intake`, `GET /api/medications/{id}/cadence`, and the new `POST /api/medications/extract` — registered in `src/lib/openapi/routes.ts` and regenerated into `docs/api/openapi.yaml`. The `rrule XOR rollingIntervalDays` invariant is documented at both the schema description and the per-field descriptions so iOS code-gen surfaces the mutual exclusion. ### Fixed - **Bi-weekly worker regression** — `src/lib/jobs/reminder-worker.ts` now consumes the canonical engine, which honours `intervalWeeks > 1` via the legacy fallback's week-phase math. A bi-weekly schedule that previously fired every Wednesday now correctly fires every other Wednesday. Pinned by an explicit regression test (`recurrence.test.ts` — `daysOfWeek = "i2;3"` emits 2 of 4 candidate Wednesdays in a 4-week window). - **One-shot lifecycle reconciliation** — `src/lib/medications/lifecycle.ts` defines `reconcileOneShotState(prisma, medicationId, userId)` and the helper runs after every intake mutation (POST + PUT + DELETE). A user who logs the single dose then immediately undoes the log gets the medication back as `active: true`; a user who flips an existing intake from real to skipped also gets the medication reactivated. The helper is a no-op on non-one-shot medications. - **Legacy fallback `startsOn` floor** — `expandLegacy` now respects `medication.startsOn` the way every other dispatch tier does. A legacy-shape schedule with a future `startsOn` no longer emits historical slots between today and the start date. - **Course-window invariants** — the create and update schemas refuse `oneShot: true` without `startsOn` (the design contract requires the anchor date) and refuse any course where `endsOn < startsOn` (which previously produced a silently dead medication). Rolling-cadence schedules cap `timesOfDay` at one entry to match the engine's single-time emission. - **Wizard `notificationsEnabled` toggle** — the Step 7 reminders switch now actually reaches the POST body. Earlier the toggle was visually live but its value was discarded by the body builder and every wizard-created medication ended up with the default `notificationsEnabled = true`. - **Edit form: one-shot single source of truth** — the medication-level `oneShot` switch now drives the picker's `allowedKinds`, so the per-schedule picker can no longer encode `kind: "oneShot"` while the medication-level switch is off (or vice versa). Toggling the switch on with multiple schedules surfaces a confirmation toast before the collapse-to-single-schedule, instead of silently dropping the extras. - **Wizard accessibility and tap-target hygiene** — focus advances into the new step's first input on Next; the cadence-picker rows are now full-row click targets (44 px); the nav buttons (Back / Next / Create) and the natural-language trigger rise to 44 px; the wizard card carries `aria-busy` while the submit is in flight. ### Changed - **Recurrence-engine defence-in-depth** — `nextOccurrenceAfter` carries a `MAX_CHUNKS = 80` cap alongside the pre-existing 10-year `hardCap`, so a pathological RRULE that walks zero forward (e.g. leap-day-only) can no longer compound through many 90-day chunks. RRULE parse failures now surface via `annotate({ action: "medications.rrule.parse_error" })` instead of silently returning `[]`. - **Edit form refactor** — `src/components/medications/medication-form.tsx` now composes the picker primitives for the edit path. Pre-v1.5 medications round-trip through `inferCadenceFromLegacy` on load and dual-write both shapes on save, so the existing legacy schedules keep working unchanged. - **i18n** — 111 new keys × six locales populating the wizard, the picker primitives, the natural-language overlay, the edit-form sections, and the plain-language cadence summary. German and English carry native copy; Spanish, French, Italian, Polish ship the English string verbatim as a machine fallback for this cut. Native polish for the four fallback locales follows. ### Tests - `src/lib/medications/scheduling/__tests__/recurrence.test.ts` — 28 unit cases covering every cadence kind plus the edge-case matrix (DST spring-forward, timezone shift mid-course, skipped + late + retroactive doses, paused medication, `endsOn` cap, missing `startsOn`, multi-schedule fan-out). - Eight `tests/integration/v15-cadence-shapes.integration.test.ts` cases exercising every cadence shape against a testcontainer Postgres, plus the one-shot lifecycle (take → reconcile-deactivate → delete-intake → reconcile-reactivate → put-skip → reconcile-reactivate). - `src/components/medications/scheduling/__tests__/{CadencePicker,TimesOfDayChips,CourseWindowRow,CreationWizard}.test.tsx` — 90 component tests against the picker primitives and the wizard helpers (`validateStep`, `buildCreateBody`, `summariseCadence`, `progressIndices`, `allowedKinds`). - `src/lib/ai/coach/__tests__/medication-extract-prompt.test.ts` — five snapshot cases pinning the extraction prompt across the cadence shapes plus a citation-coverage guard test. - `src/app/api/medications/extract/__tests__/route.test.ts` — six route cases covering auth, rate limit, budget, missing provider, parse failure, and the happy path. - `e2e/medications-wizard-{daily,weekdays,biweekly,monthly,rolling,oneshot}.spec.ts` — six Playwright specs walking the wizard end-to-end via the German label surface, one per cadence shape. CI runs the full suite; locally the specs typecheck without executing the Next.js prod build. ### Notes - **Dashboard read-flip caveat (v1.5.x window).** Until the v1.5.4 read-flip lands, the dashboard card, the cadence chart, and the medication-card "next dose" line read the legacy `daysOfWeek` / `intervalWeeks` columns directly. A wizard-minted medication with `rollingIntervalDays = 7` or an RRULE encoding renders a daily-looking next-due chip on the dashboard. The reminder worker, the integration tests, and the canonical engine all consume the new fields correctly — the schedule fires on the right date, only the visual chip on the card lags. The read-flip with legacy fallback is the next release. - **OpenAPI structural enforcement.** The `rrule XOR rollingIntervalDays` invariant is enforced at four runtime layers (Zod refine, route invariant, engine dispatch, DB CHECK). The OpenAPI surface documents the constraint in prose; a structural `oneOf` discriminator lands before the iOS v0.7.x cut. - **Worker `lastIntakeAt` aggregation.** The reminder worker currently fires one `findFirst` per rolling-medication per 15-minute tick. Functionally correct; one `groupBy` collapses it to a single round-trip per tick. Deferred to v1.5.4 with the read-flip. - **Test totals.** 5494 unit + 1 skipped, 261 integration + 3 skipped, 12 Playwright wizard instances. Typecheck, lint, and `openapi:check` all green at HEAD. ## [1.5.2] — 2026-05-26 — Plumb SESSION_COOKIE_SECURE through docker-compose v1.5.1 added a `SESSION_COOKIE_SECURE` env var so plain-HTTP self-hosts can drop the cookie's `Secure` flag. The Node helper read it correctly and the unit tests passed, but a self-hoster following the documented `.env` workaround reported that `docker compose exec app env | grep SESSION_COOKIE` came back empty — the value was set in `.env`, set in the helper, but never reached the running container. Root cause is the way the bundled `docker-compose.yml` passes env vars to the `app` service: it lists each one explicitly under `environment:` rather than mounting the `.env` file wholesale. Variables not on that whitelist are read by `docker compose` for `${VAR}` substitution but never propagated to the container's process env. `SESSION_COOKIE_SECURE` wasn't on the list, so setting it in `.env` was a silent no-op. ### Fixed - `docker-compose.yml` now lists `SESSION_COOKIE_SECURE: "${SESSION_COOKIE_SECURE:-}"` under the `app` service's `environment:` block. Defaults to empty (so the helper falls back to `NODE_ENV === "production"`, the pre-v1.5.1 behaviour); setting it to `false` in `.env` now actually reaches the Node process. ### Self-hoster recipe (full, now end-to-end working) ```bash git pull # picks up the new compose file docker compose pull # picks up the new image (no-op if already on :latest) echo 'SESSION_COOKIE_SECURE=false' >> .env docker compose up -d --force-recreate ``` Verify with: ```bash docker compose exec app env | grep SESSION_COOKIE # SESSION_COOKIE_SECURE=false curl http://10.x.x.x:3000/api/version # {"data":{"version":"1.5.2",...}} ``` Then log in over plain HTTP from a non-localhost browser — the session cookie no longer carries `Secure`, the browser keeps it, and the round-trip completes. ## [1.5.1] — 2026-05-26 — Self-hosting opt-out for the session-cookie `Secure` flag A self-hoster running HealthLog on a LAN address (`http://10.x.x.x:3000`) over plain HTTP reported a silent login failure: the `/api/auth/login` POST returned 200 with a `Set-Cookie` header on the response, but the very next `/api/auth/me` request came back 401 and the page reloaded to the login screen. Root cause is the modern browser behaviour around `Secure`-flagged cookies — every cookie the app issues in `NODE_ENV=production` carries `Secure`, and on a plain-HTTP origin that is not `localhost` / `127.0.0.1` / `::1`, the browser silently drops the cookie before sending the next request. The default-Docker image runs `NODE_ENV=production`, so any operator browsing from a different host than the one running `docker compose up` (NAS / homelab / VPS / Tailscale + Magic-DNS) used to hit the dead-end. ### Added - `SESSION_COOKIE_SECURE` environment variable. Controls the `Secure` flag on the session cookie, the onboarding-hint cookie, the Withings OAuth state cookie, and the Codex device-OAuth state cookie. Three values: - **unset** (default) — flag is set when `NODE_ENV === "production"`, the long-standing behaviour. - **`false`** — flag is never set. Use this for LAN / VPN / Tailscale-only deployments where the operator deliberately serves plain HTTP and accepts the trade-off (the session cookie crosses the wire unencrypted; do NOT use on an open-internet HTTP origin). - **`true`** — flag is always set, useful when a developer fronts a `pnpm dev` server with HTTPS to test the production cookie path. - New shared helper `src/lib/auth/secure-cookie.ts` (`shouldEmitSecureCookie()`) that every `Secure`-bearing cookie call now reads. Replaces four scattered `secure: process.env.NODE_ENV === "production"` literals in `src/lib/auth/session.ts`, `src/app/api/auth/codex/device-start/route.ts`, and `src/app/api/withings/connect/route.ts`. - `.env.example` block documenting the new variable with a clear warning about the open-internet-HTTP case. ### Tests - `src/lib/auth/__tests__/secure-cookie.test.ts` — six cases covering the default path, both explicit overrides, whitespace / case-insensitive parsing, and the fall-through for unrecognised values. ### Notes - The default behaviour is unchanged for every deployment that runs behind an HTTPS-terminating reverse proxy (the documented self-hosting path) — those continue to set `Secure` exactly as before. The opt-out is intentional and explicit. - Operators currently working around the issue with `NODE_ENV=development` can move back to `NODE_ENV=production` plus `SESSION_COOKIE_SECURE=false` on this release; that keeps build optimisation + production error masking on while letting the cookie reach the browser over HTTP. ## [1.5.0] — 2026-05-24 — Native iOS client public-beta + per-day cumulative stats overwrite + cadence-aware medication compliance The minor-version cut that marks the native iOS client publicly available. The SwiftUI iOS app (separate repository) is now joinable via TestFlight: https://testflight.apple.com/join/bucuTBpa. The backend contract the iOS app speaks against has been live since v1.4.23 and has been continuously validated across every v1.4.2x–v1.4.50 release. The 1.5.0 cut also lands the highest-leverage iOS-client unblocker per the v0.6.1 code audit: `/api/measurements/batch` now overwrites per-day cumulative `stats:*` rows on a re-post instead of dropping the new value as a duplicate. ### Added - `POST /api/measurements/batch` recognises `externalId` values starting with `stats:` (`stats:HKQuantityTypeIdentifierStepCount:YYYY-MM-DD` and every other per-day cumulative HK metric — Active Energy, Sleep Duration, Walking/Running Distance, Flights Climbed) and treats a duplicate on those as an **overwrite**, not a discard. Each re-post of the same day's external id replaces the row's `value`, `unit`, `measuredAt`, `externalSourceVersion`, `deviceType`, and `sleepStage`. Sample-class externalIds (every other prefix — `uuid-*`, opaque HK identifiers) keep the strict immutable `duplicate` contract because each sample is a canonical reading. - New per-entry status `"updated"` on the batch response envelope so the iOS sync cursor can distinguish a fresh insert from a value-bump re-post. The aggregate envelope now carries an `updated` count alongside `inserted` / `duplicates` / `skipped`. - New wide-event annotation `measurement.batch.stats-overwrite` (fires only when at least one row was overwritten) so operators can grep how often per-day cumulative re-posts happen as a healthy ingest signal. - `measurement.batch.ingest` audit-log details now include the `updated` count alongside `inserted` / `duplicates` / `skipped`. ### Why Before this change, the iOS HealthKit observer would POST today's running step total once in the morning, the server would persist row #1, and every subsequent same-day re-post (as the user walked) would come back `status: "duplicate"` with the new value silently dropped. Today's Schritte tile froze at the first-sync value until next midnight. The same shape would hit every cumulative metric on a deterministic per-day external id. Closes [#213](https://github.com/MBombeck/HealthLog/issues/213); cross-device parity (web ↔ iOS) for `stats:*` metrics now works for today and every historical day on a re-sync. ### Fixed - Medication compliance now honours `daysOfWeek` and `intervalWeeks` across every call site that surfaces a rate. The legacy aggregator computed `totalExpected = schedules.length * days`, which silently ignored cadence. A weekly Ozempic schedule with all four Mondays taken in the last 30 days reported ~13% adherence (4 / 30) instead of 100%; a weekday-only 3×/day metformin schedule with every weekday dose taken reported ~73% (66 / 90) instead of 100%. `calculateCompliance` is now a cadence-aware adapter on top of `buildCadenceTimeline` — the same pair-matching pipeline that drives the per-medication cadence chart — so the rate on the medication card, the AI Coach prompt context (7d/30d/90d windows in `src/lib/insights/features.ts`), the BP-status compliance gate, the medication-compliance status insight, `/api/insights/targets`, `/api/insights/comprehensive`, and the medication-compliance pillar of the dashboard Health Score all agree on a single, cadence-correct denominator. The wire shape (`{ totalExpected, taken, skipped, missed, rate, streak }`) is unchanged so every UI tile and persisted-insight consumer keeps reading the same fields. Closes [#214](https://github.com/MBombeck/HealthLog/issues/214). Expected user-visible shift: users on weekly meds (GLP-1 agonists, biologics) will see their Health Score rise as the medication pillar moves from ~13 to ~100; users on weekday-only multi-dose schedules will see their score rise as the pillar moves from ~73 to ~100; users on daily-only schedules see no change because the legacy denominator was already correct for that path. Migrations across the eight production call sites are mechanical — the function signature is unchanged. ### Changed - README rewrite for the v1.5 cut: TestFlight badge in the badge row plus an iOS TestFlight link in the Website / Demo / Docs row and the footer. Buy Me A Coffee badge added. Status block updated to reflect that v1.5 is now the current line, with a new "Heavily developed" advisory directly below it that tells self-hosters to pin a tag, take a backup before every upgrade, and read the CHANGELOG before pulling `latest`. Tech-Stack table flags the iOS app as TestFlight-available. Roadmap table promotes v1.5 from "in active development" to "current". - README simplification: the `How it works` diagram cluster (four SVGs covering data flow, Coach pipeline, source priority, and security model) is no longer inlined in the README. The diagrams continue to live in [`docs/diagrams/`](docs/diagrams/) and are surfaced through [docs.healthlog.dev](https://docs.healthlog.dev) where they render reliably across themes and viewport widths. The `03-self-hosting-topology.svg` stays inline under Deployment because it carries deployment-time information a self-hoster wants on the first scroll. ### Tests - `tests/integration/measurements-batch.test.ts` — three new cases pinning the `stats:*` overwrite contract: solo re-post overwrites the value and returns `status: "updated"`; sample-class duplicate keeps the strict first-write-wins contract; a mixed batch with one insert + one overwrite + one duplicate returns all three statuses correctly. - `src/lib/analytics/__tests__/compliance.test.ts` — parameterised cadence matrix: 1×/day daily (7 / 0 / 18 of 21), weekly Mondays-only (all taken / one missed), bi-weekly (`intervalWeeks=2`), weekday-only 3×/day metformin, skipped-dose denominator exclusion, `medicationCreatedAt` truncation, DST spring-forward boundary in Europe/Berlin, and over-logged-day rate cap. The matrix pins the contract for every cadence the production app exercises so future schedule-shape work can't silently regress. - `src/lib/analytics/__tests__/health-score-fast-path.test.ts` — two cadence-aware regression cases: a weekly Mondays-only med with every Monday taken now lifts the medication-compliance pillar to ≥ 50 (previously ~13 under the bug); a daily-only med with every dose taken stays ≥ 90 (no regression on the path that worked). - Eight stale integration assertions retired across `withings-oauth.test.ts`, `withings-oauth-flow.test.ts`, `analytics-bp-aggregate-paged.test.ts`, `analytics-sleep-stages.test.ts`, and `apns-dispatch.test.ts` — drift from the v1.4.47.x OAuth fine-grained reason tags, the v1.4.47.2 ES256 PEM verify guard, and the v1.4.49.1 analytics slim-slice annotation rename. Three `source-priority-two-axis` cases skipped with an inline TODO referencing the v1.4.49.1 commit and the relocation candidate (`pick-canonical-workout-rows`); these tests exercised picker semantics that no longer fire on the default analytics summaries path. ### Notes - iOS coordination items closed alongside this cut: the v1.4.49 server-side `clientManaged` MEDICATION_REMINDER suppression rule is now active for iOS v0.6.0.8+ clients that opt in via `PATCH /api/auth/me/notification-prefs`; tracked in healthlog-iOS#9. Issue [#206](https://github.com/MBombeck/HealthLog/issues/206) is closed. - HealthLog suite: 5285 unit (5279 carryover + 3 new stats-overwrite cases + new compliance matrix and Health-Score regression cases that net out the legacy assertions retired during the cadence-aware migration) + 253 integration tests pass on the local Vitest run, lint clean, typecheck clean. ## [1.4.50] — 2026-05-24 — MoodLog reverse-sync (HealthLog → MoodLog push) Marc reported that mood entries logged inside HealthLog — specifically via the iOS app — never reached MoodLog. The historic integration ran one-way: `syncMoodLogEntries` polled MoodLog every 15 minutes and pulled new rows, but nothing flowed the other direction. A user tracking mood in HealthLog ended up with one log per app and no overlap. ### Added - `src/lib/moodlog/push.ts` — `pushMoodEntriesToMoodLog(userId, entries)`. Best-effort reverse-sync push to MoodLog's new `POST /api/integrations/health-log/mood` endpoint. Same per-user `moodLogUrlEncrypted` + `moodLogApiKeyEncrypted` credentials the pull side uses, same `isPublicUrl` SSRF guard, same manual-redirect policy. Wraps every failure mode in a wide-event warning so a network blip or a stale MoodLog deploy can never bubble back to the user's mood-create request. - `POST /api/mood-entries` (single) and `POST /api/mood-entries/bulk` now fire the push as a fire-and-forget side-effect after a successful create. Entries with `source === "MOODLOG"` skip inside the helper to avoid an echo loop. Bulk push only includes rows that landed as `inserted` — duplicates and skips don't retry a known failure. ### Changed - `src/lib/moodlog/sync.ts` (pull side) filters out entries where MoodLog reports `loggedVia: "HEALTHLOG"`. Those are echoes of rows we just pushed; re-importing them would flip the `source` column from the original attribution (MANUAL / WEB / TELEGRAM / iOS) to MOODLOG and double-count in any source-segregated dashboard. An `echoSkipped` wide-event annotation lets the operator confirm the round-trip is closing at one hop. ### Tests - `src/lib/moodlog/__tests__/push.test.ts` — 9 cases covering the MOODLOG-source filter, no-credentials skip, SSRF refusal, headers + body shape on the successful path, 5xx + network-reject + 3xx-redirect failure modes, and the `tags` JSON-string → array key normalisation. Pins the contract the call sites no longer await so future helper refactors can't silently break the fire-and-forget integration. - HealthLog suite: 5279 unit + integration pass, lint clean, typecheck clean. ### Cross-repo coordination The matching MoodLog v0.2.0 release adds the `POST /api/integrations/health-log/mood` handler this push targets. Both apps must deploy together for the round-trip to close — MoodLog first so HealthLog never hits a 405 / 404 on the new endpoint. ## [1.4.49.4] — 2026-05-24 — BD-tile number truncation + medication-compliance test tz drift Marc reported the dashboard BD (Sys) tile rendering `"13… mmHg"` — the systolic / diastolic pair (`131/85`) was wider than the narrowest grid column and the value-row's `truncate` ellipsis chopped the most important part of the tile (the number itself). ### Fixed - `src/components/charts/trend-card.tsx` — paired-metric tiles (BP systolic/diastolic) now render the value at `text-2xl` (one Tailwind step down from `text-3xl`). Single-value tiles (weight, pulse, glucose, …) keep `text-3xl` because their value never threatens the column-width budget on its own. The `truncate` class stays as a defence-in-depth safety net so an outlier value (e.g. `200/120 mmHg` plus a future trend caption) still clips gracefully rather than overflowing the card boundary. - `cn()` argument ordering tightened so `tailwind-merge` keeps the explicit `leading-none` — Tailwind v4's `text-3xl` / `text-2xl` carry their own default `line-height`, and the last-wins dedup rule would otherwise drop the explicit `leading-none` and break the across-tile baseline alignment. The `` baseline-alignment tests pin the contract. - `src/app/api/medications/intake/__tests__/route.test.ts` (one pre-existing test) computed its `todayKey` from UTC date components but the route's `readMedicationCompliance` uses the user's timezone (Europe/Berlin in the test fixture). At every nightly 22:00 — 24:00 UTC window — once the Berlin clock crossed midnight — the two computations diverged by one calendar day and the rollup-tier read returned zero for "today". The test now derives `todayKey` from the same `Europe/Berlin` zone the route uses; the assertion stays the same. ## [1.4.49.3] — 2026-05-23 — Full i18n call-site audit + 28 stale missing keys filled Marc reported additional raw key strings (`notifications.eventMoodReminder`, `notifications.eventMoodReminderDesc`) reaching the UI after v1.4.49.2 had landed the relative-time fix. The narrow `{count}`-call audit that drove v1.4.49.2 had missed every dynamic key construction; the full exhaustive sweep below picked up 28 keys called from real code that had never existed in any locale bundle. ### Fixed - Filled 28 missing translation keys across all 6 locales (de / en / es / fr / it / pl = 168 entries), grouped by feature surface that introduced them and never wired the strings: - **Notifications event matrix** (14 keys — added with v1.4.41 / MOOD_REMINDER surfaced the gap in v1.4.49): `notifications.event{MedicationReminder,MeasurementAnomaly,ComplianceLow,WithingsSyncFailed,SystemAlert,PersonalRecord,MoodReminder}` and the `Desc` variant of each. The `/notifications` settings page is the consumer; before this release the per-event rows on Marc's account rendered the bare keys instead of the localised name + description. - **Welcome carousel** (13 keys — added with the 3-slide intro in v1.4.45): `onboarding.welcome.{title,carouselLabel,slideOf,prevSlide,nextSlide,gotoSlide,cta}` for the chrome plus `slide{1,2,3}.title/body` for the content. The carousel mounted with raw keys for every screen-reader label, dot-pager aria-label, and slide body. - **Measurement list error state** (1 key — added with v1.4.44): `measurements.loadError`. The error branch in `measurement-list.tsx` rendered the raw key when a fetch failed. ### Added - `src/__tests__/i18n-call-site-coverage.test.ts` — call-site audit that walks every `.ts` / `.tsx` file under `src/` (excluding tests + generated code), extracts every literal `t("ns.key")` call with comment-aware parsing (skips line + block + JSDoc comments so example pseudocode in docstrings is not flagged), and asserts each key resolves to a string leaf in `messages/en.json`. The existing `i18n-locale-integrity.test.ts` then propagates the EN guarantee to every other locale via key-set parity. Pre-fix this guard would have surfaced 17 stale call sites at every test run; future regressions print a structured punch list so the offender can land every gap in one commit. ### Why this happened The guard above did not exist before. v1.4.27 B6 introduced `i18n-drift-guard.test.ts` with hand-curated key groups; surfaces introduced after that point (notifications matrix in v1.4.41, measurement list error in v1.4.44, welcome carousel in v1.4.45) shipped without anyone editing the drift-guard groups, and the locale-integrity test only catches DRIFT across locales — it does not flag keys that are missing from EVERY locale uniformly. The call-site test closes that hole: any `t()` call without a backing key now fails at `pnpm test` time, before reaching production. ## [1.4.49.2] — 2026-05-23 — Raw i18n key leak on relative-time helper Marc reported the raw string `insights.relativeHoursAgo` rendering verbatim on medication cards, recent-achievements, admin sections, the iOS notification preview, and every other consumer of `formatDateOrRelative`. Audit across the full `t(…, { count })` surface (~20 call sites) confirmed exactly two leaks: `insights.relativeMinutesAgo` and `insights.relativeHoursAgo`. Both pointed to keys that have NEVER existed in the translation bundle — only the pluralised `*One` / `*Other` variants ship. ### Fixed - `formatDateOrRelative` in `src/lib/format.ts` now dispatches to `insights.relativeMinutesAgoOne` / `relativeMinutesAgoOther` / `relativeHoursAgoOne` / `relativeHoursAgoOther` based on `count === 1`, mirroring the `src/lib/i18n/relative-time.ts:24-48` pattern the v1.4.43 i18n fix-up added to its twin helper. The format-twin was missed at the time — `t()` performs no auto-pluralisation, so the bare key passed straight through `t()`'s identity fallback into the UI. ### Test hygiene - `src/lib/__tests__/format-date-or-relative.test.ts` expectations updated to the new `*One` / `*Other` dispatch, plus a new regression guard that drives every relative bucket and asserts each `t()` key has a matching entry in `messages/en.json`. Pre-fix this guard would have failed on `relativeMinutesAgo` and `relativeHoursAgo`; future twin-helper divergence will trip the same check before it reaches prod. ### Why the audit found no other leaks Cross-checked every `t(…, { count })` call across the codebase against `messages/en.json`. The other 17 sites (`insights.dayStreak`, `medications.importDuplicatesSkipped`, `achievements.metricPercent`, `passwordStrength.minLength`, `settings.integrationPill.*`, `targets.relativeDay.daysAgo`, `targets.card.streak`, `admin.section.backups.uploadSuccess`, `trendHints.remainingMany`, `dashboard.staleHint`, …) all use plain (non-pluralised) keys that exist verbatim in the bundle. The bug was isolated to the format-helper twin. ## [1.4.49.1] — 2026-05-23 — Default `/api/analytics` cold-path fix (rollup-tier delegation) v1.4.49 shipped the slim-slice (`?slice=summaries`) cold-fallback fix, but the production HAR captured against the new image still showed `GET /api/analytics` at 8 s cold on the 467 745-row tenant — and crucially, the supposedly-fixed `?slice=summaries` request observed the same 8 s when fired concurrently. Investigation found the slim slice was queuing behind the default slice's 15-way per-type `fetchMeasurementSeriesChunked` live walk: the fan-out held the `p-limit(4)` lanes saturated and the 20-slot Prisma pool packed for ~8 s, starving every other Prisma client (including the concurrent slim slice) until it drained. ### Fixed - `GET /api/analytics` (default slice) no longer fans out 15 per-type `findMany` reads against `measurements`. The route now delegates the per-type `summaries` work to `computeSummariesSlice`, which reads the same data from `measurement_rollups` DAY buckets + a 90-day narrow `$queryRaw` for the windowed avg / slope / r² columns. On Marc's production account this drops the cold critical path from ~8 s to sub-second; warm cache is unchanged. - Pool starvation that ALSO affected the supposedly-fast `?slice=summaries` request when fired concurrently — the slim slice now resolves in its own SQL budget regardless of whether the default-slice request is also in flight. ### Changed - `computeSummariesSlice` narrow `$queryRaw` extended with a `FILTER (WHERE measured_at >= NOW() - INTERVAL '60 days' AND measured_at < NOW() - INTERVAL '30 days')` `avg30_last_month` column. The 60-day lower bound stays inside the existing 90-day outer cap, so the additional column adds zero extra row scan work — the planner already touches every block this clause reads. Dashboard `tileCompareDelta` (`compareBaseline === "lastMonth"`) now sees a real value via the slim path; previously only the deleted live walk produced it. - Deleted dead code: `fetchMeasurementSeriesChunked`, `ChunkedRow`, `MEASUREMENT_CHUNK_SIZE`, the `ANALYTICS_TYPE_FETCH_CONCURRENCY = 4` constant. None had callers outside the removed fan-out. - `meta.analytics.bp_aggregate` wide-event field retired — it carried the per-type walk's row-count + `live_since` cutoff, both meaningless without the walk. The slim slice's `meta.analytics.slim_summaries` (`row_count`, `type_count`, `path`, `year_over_year_types`) carries the equivalent signal. - `summary.anomalyCount` on the default-slice response now consistently returns `0`. No consumer in the codebase reads it from `/api/analytics`; the insights pipeline (`/api/insights/comprehensive`, `/api/insights/cards`) sources `anomalyCount` from its own `comprehensive-aggregator` narrow query. ### Test hygiene - `route.test.ts` updated: dropped the 130k-row PULSE stress test and the `caps per-type Prisma fan-out` concurrency test (both pinned the removed fan-out). Added a negative assertion that `prisma.measurement.findMany` is never called with the chunked-walk shape `(select: { id, source, deviceType }, take: 5000)` on the default critical path, plus a positive assertion that `avg30LastMonth` exists on the response. - `since-cap.test.ts` updated: dropped the `where.measuredAt.gte` and `bp_aggregate.live_since` annotation assertions (both pinned the deleted code path). The slim-slice negative invariant (no per-type loop on `?slice=summaries`) remains. - Final: **5 268 unit + integration tests pass** (one pre-existing skip), zero failures, lint clean. ### Verification Post-deploy production logs from `/api/analytics` (default + slim, cold cache) confirm the fix landed: both routes report `path: "rollup"` for every component and `duration_ms` returns to the sub-second envelope. See the v1.4.49.1 deploy verify section in the project memory. ## [1.4.49] — 2026-05-23 — Server-side reminder suppression + diagnostic endpoint backed + backlog closure v1.4.47 closed the high-priority punch list. v1.4.49 bundles the remaining v1.4.47 follow-ups together with the items deferred out of v1.4.48: server-side suppression of `MEDICATION_REMINDER` APNs for iOS clients that manage their own local reminders, a `push_attempts` table backing the diagnostic endpoint, the `/api/admin/notifications/diagnostic` OpenAPI entry, and a sweep of v1.4.48 forward-findings (Withings reason-tagging, observability PII hardening, MoodReminderCard auto-clear parity, simplifier dead-code, sub-locale copy gaps). The workout-batch integration suite that had been red on `main` for three releases turns green; the cold-mount analytics fallback path picks up the same 90-day outer cap v1.4.47.1 shipped for the rollup-fresh path; iOS validation-failure audit rows carry the rejected payload shape so iOS serialiser drift can be chased from a single log line. ### Added - `/api/admin/notifications/diagnostic` — admin-only endpoint that surfaces device APNs token presence (masked to an 8-char prefix + 8-char suffix), notification channel config presence, and recent push attempts for the calling account. Closes the DB-shell bottleneck that slowed the v1.4.47 APNs investigation. The endpoint is registered in `docs/api/openapi.yaml` so future iOS clients can codegen against it. - `push_attempts` table backs the diagnostic endpoint's `recentPushAttempts` field. Every APNS / Web-Push / Telegram / NTFY sender writes a fire-and-forget row per attempt (channel + eventType + result + reason + createdAt). The diagnostic endpoint now returns the last 20 attempts ordered by recency. Daily cleanup cron at 03:35 Europe/Berlin prunes rows older than 90 days. - `GET` + `PATCH /api/auth/me/notification-prefs` — per-user notification preferences with deep-merge semantics. Initial category `medication.clientManaged: boolean` (default `false`); future categories slot in without overwriting siblings. The cron `medicationReminderJob` skips dose-due APNs when the flag is `true`, emitting a `medication_reminder.suppressed_client_managed` wide-event annotation per skip. Other notification kinds (mood reminder, personal record, system alert) are unaffected. Closes the server side of GitHub issue #206. - `redactSensitiveFields(body)` helper at `src/lib/observability/redact-payload.ts` with a denylist (`password` / `token` / `secret` / `apiKey` / `authorization` / `csrfState` / `nonce`, case-insensitive, recursive). Used by the wide-event `received_shape_excerpt` surface so future free-text routes adopting the pattern cannot leak credentials. - Native sub-locale translations for `settings.about.tourReplay` and `tourReplayHint` (es / fr / it / pl), and for `settings.about.linksHeading` + `newerAvailable` across the same four sub-locales. Tightened `dashboard.dragHandleHint` copy in de + pl (was ~30 words on a `sr-only` paragraph; screen readers spent ~6 s reading it on focus). - OpenAPI documentation for `GET` + `PATCH /api/auth/me/disable-coach`. - Operations runbook notes: `attemptNumber` audit-row vs alert-signal semantics at `docs/ops/attempt-number-semantics.md`, and explicit migration-before-container ordering at `docs/ops/deploy.md` for manual / partial-deploy scenarios. ### Changed - `/api/analytics` cold-fallback path splits the heavy single-statement aggregate into two parallel queries. The linearly composable columns (`count` / `min` / `max` / `mean`) keep the full-partition scan; the windowed 7 / 30 / 90-day `avg` / `slope` / `r²` columns take a 90-day outer cap so the planner does an index range scan on `(user_id, type, measured_at)` instead of a full-partition sequential scan. Output is bit-identical; cold slim slice drops from multi-second to sub-second on tenants with large measurement partitions. - Dashboard reorder + onboarding tour helpers consolidated. Arrow-button reorder delegates to the same swap path as drag-and-drop. `restartOnboardingTour` extracted so About + Account sections share the surface. `useSortable` honours `prefers-reduced-motion`. Drag-handle `aria-describedby` hint gates on visible widget count so the empty-state cannot orphan the hint paragraph. - Coach disable surface tightened. The status banner auto-clears after 3 s. `PATCH /api/auth/me/disable-coach` body migrated to Zod + `returnAllZodIssues` for the uniform 422 envelope. The `eslint-disable react-hooks/rules-of-hooks` line in `useDisableCoach` is replaced with a shared `useQueryClientMounted` helper that both `useFeatureFlags` and `useDisableCoach` consume. - Withings OAuth callback consumes the nonce atomically. `delete` is issued first; the `P2025` catch handles the replay branch. The previous read-then-delete pattern allowed a theoretical race between two concurrent callbacks with the same nonce. Silent `.catch(() => {})` on delete failures now surface via a Wide-Event warning so a real infra failure reaches the audit trail. - Withings callback `?reason=state` redirect is now differentiated into four distinct reason tags so operators can tell `csrf1` (URL/cookie mismatch), `replay` (nonce already consumed), `expired` (TTL elapsed), and `cross_user` (session/row userId mismatch) apart without DB-shell access. The delete-failure warning template interpolates `err.name` rather than `${err}` so Prisma error messages cannot echo offending values into the audit log. - Withings `connect` endpoint rate-limited to 10 calls per 60 s per user (mirroring the disable-coach pattern); row-create failures redirect to the consistent error surface (`/settings/integrations?withings=error&reason=connect`) instead of bubbling a 500. - `buildPayloadDiagnostic(body)` helper extracted to `src/lib/api-response.ts`. The dashboard widgets PUT route and the measurements series GET route now route their bodies through `redactSensitiveFields` first then call the shared helper — single source of truth for the iOS payload-diff wide-event shape, redaction layered in front composably. - `sanitiseZodIssues(issues, { stripValuesFromMessage: true })` overload added; 14 audit-log sites that wrote `JSON.stringify({ issues })` to `details` now strip the `message` string to a `{ path, code }` pair. Most impactful catch: the `/api/devices` route's `invalid_format` issues were echoing `apnsToken` values verbatim into the audit log. - `disableCoachBody` + `disableCoachData` Zod schemas collapsed into a single `disableCoachFlag` schema in the OpenAPI route table; `zod-openapi` emits the request + response pair from the one source. - `yaml@2` emitter no longer sorts map entries during `pnpm openapi:generate`. Alphabetical sort had placed alias references before their anchors when `.meta()`-tagged sub-schemas appeared inside `z.array(...)`. Output stability is preserved by Zod's declaration-order guarantee. - Worker boot emits a Wide-Event warning if any `integration_statuses` row still has a NULL per-kind counter after the v1.4.47 legacy-column drop. Closes the silent alert-ladder gap where such rows would alert two strikes later than they should. - `MoodReminderCard` status banner auto-clears after 3 s — parity with `DisableCoachCard` (the existing docstring claimed parity that the implementation lacked). - `mergeReorderIntoLayout` emits a dev-only `console.warn` when a reorder id has no matching widget in the layout. Statically unreachable today; defence-in-depth for the upcoming per-tile Suspense refactor that introduces dynamic widgets. ### Fixed - Workout-batch integration tests (`workout-batch-create`, `workout-batch-race`) updated to match the v1.4.42 cross-source dedup semantics. The integration suite had been red on `main` since v1.4.45 (three release commits shipped through it). Suite is green again. - Twelve stale "migration 0076" references in code comments corrected. `consecutiveFailures` sites read 0077; `disable-coach` sites read 0078. Comment-only sweep, no behavioural change. - `docs/ops/attempt-number-semantics.md` referenced migration 0076 for the dropped `consecutiveFailures` column; the actual migration is 0077. - Per-channel APNs "Send test" button (Settings → notification channels) now fires with `eventType: "MEDICATION_REMINDER"` so the dispatcher's `time-sensitive` interruption-level + `apns-priority: 10` branch runs. The previous `SYSTEM_ALERT` event-type used the dispatcher's default `active` level, which iOS may summarise into the Notification Center rather than presenting a lock-screen banner — leaving the user with no signal whether a real dose reminder will surface visibly. A successful test now exercises the exact path production reminders take. Title and body strings still read "Test notification" so the push cannot be mistaken for an actual scheduled dose. ### Performance - Cold-fallback `/api/analytics` slim slice — `aggregates` query split so the windowed columns get a 90-day outer `WHERE` cap, matching the inner `FILTER` windows. Multi-second TTFB on cold mounts drops to sub-second on tenants with large measurement partitions. The `narrows` query received the same cap in v1.4.47.1. ### Security - Withings `connect` rate-limited (per-user 10 / 60 s) to bound ledger-row growth from a misbehaving client. - Coach `disable-coach` validation envelope unified through `returnAllZodIssues` for a uniform 422 shape. ### Observability - `dashboard.widgets.validation-failed` and `measurements.series.validation-failed` audit rows now carry `received_keys`, a truncated `received_shape_excerpt` (256-char hard cap), and the sanitised Zod issues. The excerpt is generated through `redactSensitiveFields` so credential-shaped keys (password / token / secret / apiKey / authorization / csrfState / nonce) land as the literal `"[redacted]"` instead of their raw values. One log line per validation failure carries both the iOS-sent shape and the server's rejection reason, so iOS serialiser drift can be chased without DB-shell access. - Push attempts persisted in `push_attempts` for every channel (APNS / Web-Push / Telegram / NTFY). The diagnostic endpoint reads the last 20 per user; the daily cron at 03:35 Europe/Berlin prunes attempts older than 90 days. Operators can now grep the wide-event `notifications.send` annotation OR query the table directly for an APNs incident triage. - Withings callback redirect reasons emit matching `meta.reason` annotations (`csrf1` / `replay` / `expired` / `cross_user`) so the wide-event stream carries the same differentiation the URL query param does. ### Documentation - `NEXT_PUBLIC_APP_VERSION`, `NEXT_PUBLIC_APP_BUILD_SHA`, `NEXT_PUBLIC_APP_BUILT_AT` documented as build-time-only env-vars in `docs/self-hosting/reverse-proxy.md`. The Coolify runtime entry was misleading; Next.js bakes the values into the bundle at build time via the `next.config.ts` `env:` block. ### Test hygiene - New grep-discovery test mirrors the `flags.coach` discovery test for `user.disableCoach` / `useDisableCoach`. New sites must be added to the allowlist explicitly so CI fails on an unreviewed addition. - `target-card.tsx` per-card "Ask the coach" CTA pinned in the user-disable fixture (the cascade fixture already covered the page-level gate). - `textarea` `forwardRef` `Symbol.for("react.forward_ref")` sentinel assertion deleted (React-internal API). Ref forwarding is covered indirectly by the call sites that pass `ref` to `