# Changelog ## [1.4.47.6] — 2026-05-22 — APNs per-channel test endpoint wired The Settings notification-status card's "Test senden" button was unwired for the APNS channel — the `TEST_ENDPOINTS` map at `notification-status-card.tsx:60-64` had TELEGRAM / NTFY / WEB_PUSH but not APNS. Clicking the button on an APNS row resolved to `undefined` and the fetch fired against an empty URL, silently failing. The operator had no first-class way to verify the v1.4.47.5 auto-detect path; only the natural channel-state-machine cron retry could exercise it. This release adds the missing endpoint + map entry. ### Added - `src/app/api/notifications/apns/test/route.ts` (new) — per-channel APNS self-test endpoint. Mirrors the web-push test shape: `requireAuth` + 5/min rate-limit + calls `sendViaApns(user.id, …)` with a localised "test notification" body. Returns `{ ok, reason }` so the UI can surface the actual APNs reason on failure. ### Changed - `src/components/settings/notification-status-card.tsx:60-65` — added the `APNS` row to `TEST_ENDPOINTS`. The button now actually fires. ### Effect - Operator can verify APNs delivery without waiting for the channel-state-machine's exponential-backoff cron retry. The v1.4.47.5 auto-detect path runs inside `sendViaApns`, so clicking "Test senden" on the APNS row will exercise it directly. ### Risk Zero. Additive endpoint + one new map key. Existing channels keep working. ## [1.4.47.5] — 2026-05-22 — APNs gateway auto-detect on `BadEnvironmentKeyInToken` v1.4.47.4 + the Coolify `APNS_PRODUCTION` env-var deletion got JWT signing past Apple's gate, but the next push still failed with `BadEnvironmentKeyInToken` — the server's per-device gateway routing didn't match the actual token environment. The iOS client picks the gateway env when it registers, but in practice the client's reported env can mismatch the token's true environment (e.g. a DEBUG build that registered itself as "production" by accident, or a TestFlight-then-Debug installation chain). This release adds gateway auto-detect: on `BadEnvironmentKeyInToken`, retry exactly once with the opposite gateway. If the retry succeeds, persist the correction to `Device.apnsEnvironment` so subsequent sends go straight to the right gateway. ### Changed - `src/lib/notifications/senders/apns.ts:sendViaApns` — per-device send-loop now walks an env-sequence (`[initial, opposite]`). Breaks on success OR on any non-`BadEnvironmentKeyInToken` failure. On retry-success, updates `Device.apnsEnvironment` in DB and emits a wide-event warning so the operator sees the correction. ### Effect - Future pushes hit the right gateway for each device, even if the iOS app misreported the env at registration time. - The first delivery after deploy may take 2 RTT-to-Apple round-trips for any device that needs correction; subsequent ones go through directly. ### Risk Low. The retry only fires on the specific `BadEnvironmentKeyInToken` reason. Other failures (`BadDeviceToken`, `Unregistered`, network) take the existing single-attempt path. Two new unit tests cover the retry-and-correct path + the no-retry-on-BadDeviceToken path. ## [1.4.47.4] — 2026-05-22 — APNs key escape-free env var (APNS_KEY_B64) The v1.4.47.2 / .3 chain established that the production `APNS_KEY` env-var is mangled along the `docker-compose env_file` → `process.env` pipeline. Defensive normalisation cannot recover it because the base64 body itself is corrupted on arrival. This release adds an escape-free alternative: `APNS_KEY_B64`. The operator base64-encodes the raw `.p8` file (real newlines intact) and stores that single ASCII-safe blob; the app decodes it back to a PEM string. No `\n` escape gymnastics, no character class transformations along the way — base64 chars survive every env-var pipeline. ### Changed - `src/lib/notifications/senders/apns.ts:loadApnsConfig` — read `APNS_KEY_B64` first; if present, decode via `Buffer.from(value, "base64").toString("utf-8")` and pass directly to `crypto.createPrivateKey` for verification. Existing `APNS_KEY` and `APNS_KEY_FILE` paths kept as fallbacks. Precedence: `APNS_KEY_B64 > APNS_KEY > APNS_KEY_FILE`. ### Operator action To enable on Coolify (or any other env-var-block): ```bash base64 -i AuthKey_.p8 | tr -d '\n' ``` Paste the output as `APNS_KEY_B64`. The legacy `APNS_KEY` can stay set or be removed; the B64 variant overrides it when present. ### Risk Zero. Additive — the new env-var has its own try/catch + parse-verification with one-warning-per-process disable semantics. Existing operators keep their setup unchanged. ## [1.4.47.3] — 2026-05-22 — APNs PEM diagnostic dump on parse failure v1.4.47.2 added defensive PEM normalisation but the production env-var was still rejected at `crypto.createPrivateKey` with `error:1E08010C:DECODER routines::unsupported`. The base64 body itself appears to be corrupted somewhere along the Coolify → docker-compose → `process.env` path; normalising the wrapping cannot recover it. This release adds a one-line diagnostic dump on parse failure so the operator can compare the in-container shape against the source `.p8`: ``` APNs key did not parse as PEM: [diag raw_len=… norm_len=… escaped_newlines=… real_newlines=… begin_markers=… end_markers=… base64_chars=… other_chars=… sha256_prefix=…] ``` No secret bytes leave the warning — only character-class counts, structural marker counts, and a 16-char SHA-256 prefix that the operator can compare against `shasum -a 256 AuthKey_*.p8` locally. ### Changed - `src/lib/notifications/senders/apns.ts:loadApnsConfig` — on the existing `crypto.createPrivateKey` failure branch, emit a structured diagnostic warning. Channel disable behaviour unchanged. ### Risk Zero. Behaviour-preserving — only the warning message gets richer when the parse already fails. ## [1.4.47.2] — 2026-05-22 — APNs JWT signing repair (defensive PEM normalisation) Same-day follow-up to v1.4.47.1. Coolify runtime logs surfaced ``` Notification sender threw for APNS: Failed to generate token: secretOrPrivateKey must be an asymmetric key when using ES256 ``` on every push attempt. The channel had auto-paused after 3 consecutive failures (admin UI showed `sender_threw · Versand pausiert · Nächster Versuch 11:10`). Root cause: the .p8 PEM stored in the Coolify env-var arrived at `process.env.APNS_KEY` without parseable newlines around the BEGIN / END markers — likely a `docker-compose env_file` round-trip artefact. `openssl pkey` could still parse it, but `jsonwebtoken@9` (used by `@parse/node-apn`) is strict and refused to sign. ### Changed - `src/lib/notifications/senders/apns.ts:loadApnsConfig` — after the `.replace(/\\n/g, "\n")` 12-factor unescape, normalise the PEM body: strip whitespace between markers, force a 64-char line wrap of the base64 payload, and rebuild a canonical PEM. Idempotent on already-correct PEMs; recovers bare-base64 inputs without markers too. Then verify the result parses as an asymmetric EC key via `crypto.createPrivateKey` before handing to node-apn. On verification failure, return `null` with a one-time warning — beats `sender_threw` on every push. ### Effect - APNs channel returns to active on the next push attempt (the channel-state machine clears the cooldown on the first success). - No env-var change required; the existing `APNS_KEY` value is normalised in-process. - Operators with already-correct multi-line PEMs see no behavioural difference. ### Test plan - [x] 3 new unit tests cover the v1.4.47.2 normalisation paths: collapsed single-line PEM, bare base64 body without markers, unparseable garbage returns null + warning. - [x] Updated `VALID_ENV.APNS_KEY` fixture to a real EC P-256 key so the existing tests exercise the verification path end-to-end. - [x] `pnpm typecheck` clean, `pnpm lint` clean, full suite green. ### Risk Low. Adds defensive normalisation + a parse check. Bad PEMs that previously caused `sender_threw` on every push now cleanly disable the APNs channel with a single warning at load time. Good PEMs are passed through unchanged after the canonical re-wrap (which produces an identical PEM for already-correct inputs). ### Operator notes Standard image roll. No `prisma migrate deploy` step required. ## [1.4.47.1] — 2026-05-22 — Slim summaries slice 9 s → ~0.5-1 s cold Same-day hotfix on top of v1.4.47. Dashboard cold mount on power-user accounts was firing `/api/analytics?slice=summaries` at ~9 s TTFB; the time was almost entirely in one `$queryRaw` against the `measurements` table. The `narrows` query inside `computeFromRollups` was scanning the user's full measurements partition without an outer `measured_at` cap. Every `FILTER` expression inside the SELECT already restricts to 7/30/90 days, so the additional rows were aggregated to NULL and discarded — but the planner still had to read all of them. Adding the matching 90-day outer WHERE turns the read into an index range scan on `(user_id, type, measured_at)` and returns the same output bit-for-bit. Effect for a ~450 000-row tenant: slim slice cold ~9 s → ~0.5-1 s, warm cache hits stay at the sub-50 ms Map-lookup. The default slice keeps its earlier path; a separate split of `computeFromLiveAggregate` is queued for the next release for tenants who still rely on the no-rollup fallback. - **Changed:** `src/lib/analytics/summaries-slice.ts:255-306` — `computeFromRollups.narrows` query gains `AND m."measured_at" >= NOW() - INTERVAL '90 days'`. - **Risk:** zero. Inner FILTER clauses already discarded rows outside the 90-day window; the new WHERE just narrows the scan to the same set. - **No migration. No schema change. No env-var change.** - **Operator notes:** standard image roll; no `prisma migrate deploy` step required. ## [1.4.47] — 2026-05-22 — Audit-backlog closure: drag-to-reorder, Coach disable toggle, OAuth state nonce table, legacy column drop, primitive sweep v1.4.45 closed the v1.4.43-audit follow-up; v1.4.46 caught a same-day server reconcile (PR worker, intake auto-skip, APNS admin test). v1.4.47 is the dedicated follow-up that lands every deferred v1.4.45 audit Medium/Low plus the W14 legacy-column cleanup the v1.4.45 release scheduled for "one release later". Eight changes landed on `develop` before this release commit: - **W0** wall-clock pin on the two idempotency tests (`/api/dashboard/summary` + `/api/medications/intake`) that hard-coded `2026-05-21` and broke on the 22nd - **W1** drop the legacy `consecutive_failures` column on `integration_statuses` — the v1.4.45 W14 per-kind bucket migration carried the legacy integer for one release as a fallback; now removed, alert ladder reads `Math.max(...buckets)` - **W2** extract `