Table of contents
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
KNOWN ISSUES
Agent-maintained. Each issue is one line. Criticality governs urgency.
Active Issues
[ISSUE-008] low open area:tests flaky Playwright test — albums.spec.ts:1771 "keeps suggestion review context opt-in" + timeline.spec.ts:1595 "Rerun triggered" dialog wait. Both fail ~50% on scoped test run when web scope triggered. [PR#541 merge CI] [audit 2026-10-24: line refs drifted — albums.spec.ts:1996, timeline.spec.ts:2647; flake class unchanged, no hydration gate added] [audit 2026-10-30: CI evidence pipeline landed via PR#1444 — Playwright traces preserved on failure (upload-artifact + durable host cache fallback) and a nightly flake-detection job (retries=0, repeat-each=10, desktop, the six known-flaky specs) now watches this flake class; --shm-size=1g on all CI Playwright containers removes the Chromium-in-Docker /dev/shm amplifier; attribution pending a trace-bearing red run] [audit 2026-10-30: test surface split — PR#1454 split timeline.spec.ts (3704 lines → timeline.spec.ts + timeline-media-detail/live-insert/hydration/scroll-stability; rerun-toast tests now timeline.spec.ts:1199 + timeline-media-detail.spec.ts:1512+), albums gate landed in PR#1445; hydration subset mitigated — the rerun-toast toast-race subset still awaiting nightly flake-detection evidence; stays open]
[ISSUE-010] high open area:sharing IDOR: POST /api/contributions acts on arbitrary ownerId from request body, principal never compared — any authenticated user force-contributes another owner's private media into a group (SEC-1). [audit:2026-09-04]
[ISSUE-011] high open area:sharing IDOR: contribution-policies create/evaluate for any ownerId, no principal or group-membership check — auto_all + auto_contribute auto-contributes victim's entire library (SEC-2). [audit:2026-09-04]
[ISSUE-012] high open area:sharing IDOR: candidate accept/reject trusts body ownerId, candidateForOwner filters by owner_id only — attacker accepts/rejects victim's pending candidates (SEC-3). [audit:2026-09-04]
[ISSUE-014] high open area:export ExportController no role check; ADR-0002 assigns export to Super Admin — any user triggers full-instance metadata dump, no audit (SEC-5). [audit:2026-09-04]
[ISSUE-015] high open area:auth SessionAuthenticationFilter trusts session attr without DB revalidation — disabled user/downgraded admin keeps working session up to 30d (SEC-6). [audit:2026-09-04]
[ISSUE-016] high open area:export ExportService entire export in one @Transactional → pool exhaustion, repeatable DoS by any user via SEC-5 (SEC-7). [audit:2026-09-04]
[ISSUE-019] med open area:discovery DiscoveryService (String) cast on JSONB provenance column → PGobject ClassCastException → 500 on live PATCH markers + label endpoints (ENG-2). [audit:2026-09-04]
[ISSUE-021] med open area:media Long tx during file I/O: full-file copies, SHA-256 of 500MB+ model files, multi-file resets inside @Transactional → connection held minutes, pool exhaustion (ENG-4). [audit:2026-09-04]
[ISSUE-023] med open area:workers ML workers missing exif_transpose — bbox computed against unrotated dims → face/plant boxes wrong on rotated photos (ENG-6). [audit:2026-09-04] [audit 2026-10-24: face + ml lanes now apply exif_transpose (face_model.py:158, ml jobs.py:606); plants worker still raw — plants half remains]
[ISSUE-027] high mitigated area:people Pending-embedding purge deleted face embeddings >24h old even while their faces were still unassigned → permanently lost embeddings (owner DB: 38 assigned faces, 0 stored embeddings, 5867 purged); purge now orphan-only (face gone/superseded) [fix/face-suggestions-empty]
[ISSUE-028] med mitigated area:people Face-assignment dialog "Suggested people" showed empty state despite named people — blank-query path had no fallback when class_occurrence_embeddings is empty; added name-ordered fallback list in suggestSimilarInstances [fix/face-suggestions-empty]
[ISSUE-030] low open area:workers Face-crop enqueue guard is SELECT-based (no unique constraint) — staggered calls race → duplicate queued jobs per face (observed 9.6k jobs for 4.4k faces on first backfill); self-heals on completion (faces disqualify via asset/pointer), wasted queue slots only; fix: unique index on (job_type, media_item_id, faceAnnotationId-payload) or advisory-lock the enqueue [PR#709]
[ISSUE-032] med mitigated area:workers workers/people clustering holds all embeddings + NxN distance matrix in RAM; ceiling ~40k faces before O(n²) RSS reaches container limits (NumPy unbounded growth reported earlier as worse); vectorized clustering frozen-branch only — chunk/cluster-stream later if libraries grow past the ceiling [PR#718 window, ISSUE-035 memory claim corrected]
[ISSUE-037] med mitigated area:people Zero-candidate recognition retained pending embeddings and repeatedly re-armed the existential guard; completion cursor now bounds retry evidence, and recovery restores structured owner_id before historical jobs re-enter the active queue [spec:018]
[ISSUE-039] med open area:people Pre-deploy recognition jobs with stale assignment snapshots fail 400 "Candidate existing class instance is not active for this owner/domain/class" and burn 3 retry attempts each; fix options: refresh assignment snapshot at claim, or treat permanent validation 400s as terminal without retry [PR#724, diagnosis 2026-09-15]
[ISSUE-041] low open area:discovery H2 test fixture diverges from PostgreSQL in insertOccurrenceNoSteal: H2 MERGE KEY replaces a superseded occurrence row (automated re-accept revives dissolved keys in tests) while PG ON CONFLICT DO NOTHING keeps it insert-only; production semantics documented on the method, H2-only quirk [PR#728]
[ISSUE-042] low open area:tests web Playwright scoped PR suite fails 8 pre-existing tests on clean main (5bec1277): albums.spec.ts:49/1778 + map.spec.ts:83 (desktop+mobile) and accessibility.spec.ts:123/152 (mobile hamburger drawer/search); reproduced identically on detached origin/main checkout — unrelated to changed files [PR#732 run#1620, run#1621]
[ISSUE-052] low open area:discovery Tentative candidate with only unresolvable refs can be accepted with zero occurrences: DiscoveryStandardResultModels.occurrenceRefs does not require a resolvable id, OccurrenceRefCanonicalizer.canonicalizeStored silently drops malformed refs, and the confirm primitive's zero-write guard is skipped when the canonical ref list is empty → accepted audit without occurrence (narrow edge, requires malformed worker input); fix: treat empty canonical refs as revalidation failure or validate resolvable ids at worker ingestion [spec:041 review]
[ISSUE-053] low open area:discovery Tentative auto-accept assigned-ref filter falls back to mediaItemId for bare media refs while stored occurrence keys use mediaItemId + ":" + index, so the pre-assigned check misses that ref shape; harmless today (insert-only path + unique constraint still prevent a steal, preassigned test passes) but filter is ineffective for media-only refs; fix: align key derivation with occurrenceKey() or skip media-only refs [spec:041 review]
[ISSUE-056] low open area:ci Concurrent web-test CI jobs collide on the Playwright preview port: containers run --network host, the Playwright image does not set CI, so reuseExistingServer: !process.env.CI is true and jobs starting within seconds share 127.0.0.1:4173 — a suite can silently run against another branch's preview build (run 2577 badge test failed against run 2578's older app, "passed" runs are not always testing their own head); mitigated in #779 via per-container preview port + reuseExistingServer: false inside containers, but collisions inside the same host still abort the second job with port already used (run 2595) and there is no Forgejo rerun API for scoped jobs — retrigger requires a real file change because detect-scopes.sh skips empty pushes [PR#779, run 2595]
[ISSUE-059] med open area:upload Upload-batch completion can be missed when the last outstanding job of a completed session is terminalized by WorkerJobRecoveryService (expired lease / retry exhaustion) instead of a worker-submitted result: recovery marks media jobs failed/expired without publishing WorkerJobCompletedEvent (it only publishes owner_face_recognition terminal failures), and the commit-path check ran before that, so the session stays unnotified and keeps showing as active batch until some other job of the session completes; fix options: scheduled sweeper for completed unnotified sessions whose jobs are all terminal, or recovery-published completion events for media jobs [feat/upload-batch-notifications]
[ISSUE-060] low mitigated area:upload A browser that never sends the final commit leaves its upload session in status 'receiving' forever (no session expiry exists), so the batch-scoped activity keeps showing the session's final counts and never produces an upload_batch_completed notification; mitigated: the owner can now discard such sessions from the media activity dialog (POST /api/upload-activity/{sessionId}/cancel marks them 'cancelled', receive/commit paths reject terminal sessions), but automatic session expiry is still missing — sessions cancelled only by hand stay in 'receiving' until a user intervenes [PR#952, PR#955]
[ISSUE-069] med mitigated area:ci upload-artifact backend unreachable from memories-deploy runner: job containers run with host networking, so the Forgejo artifact backend hostname gitea is unresolvable (getaddrinfo EAI_AGAIN) and upload-artifact@v3 fails after retries; android APK delivery therefore goes through the persistent host mount instead (/home/chris/git/.ci-cache/android-apk//app-debug.apk + BUILD_INFO.txt, 14-day pruning), with the artifact step kept best-effort via continue-on-error — starts working automatically once runner network config resolves gitea [feat/android-core-sync-engine, run 2882 job 36503]
[ISSUE-074] low open area:api Upload-dedup re-uploads keep committing multiple upload_session_files rows per media item (skipped_duplicate rows with committed_media_item_id set); the maintenance-path crash was fixed via scalar subqueries in #943, but batch-feed queries in SearchService (count/semantic backfill), WorkerJobEnqueueService (6 backfill feeds), DiscoveryReprocessScheduler and UserDiscoveryService still plain-join usf → duplicates can double-enqueue jobs per pass (query-time NOT EXISTS guards evaluate before enqueue); no crash, self-limiting, fix by same pick-one-row pattern [PR#943]
[ISSUE-075] med mitigated area:web-tests mobile accessibility.spec.ts:145 nav-drawer test races aria-expanded state after PR#947's AppHeader redesign (drawer toggle aria state settles a tick later on mobile viewport), ~1 failure per full mobile suite at retries=0, passes on retry, reproduced on clean 08aaba65; fix by asserting on the drawer panel visibility instead of the aria-expanded attribute race [PR#947 attribution, run 41153 repro by eng-1] [audit 2026-10-30: prescribed fix landed in PR#1445 — root layout stamps html[data-app-hydrated] in onMount and the test gates on hydration (gotoHydrated) AND asserts drawer panel visibility instead of the aria-expanded race; mobile suite green at retries=0; nightly flake-detection (PR#1444) confirmation pending before resolved] [audit 2026-10-30: mitigation survived the full nightly + PR wave (incl. the #1453/#1454/#1455 test-consolidation and tentative-refresh fixes) — mobile suite still green; stays mitigated pending nightly confirmation]
[ISSUE-076] med open area:ci Scoped PR test jobs intermittently run :apps:api:test without a usable Docker daemon — every Testcontainers suite fails with IllegalStateException (DockerClientProviderStrategy.java:229) cascading into NoClassDefFoundError from static PgVectorPostgresImage init (run 3281 job 41478 on PR#954: 180 failures, all infra-init errors; identical suite green locally and in the neighboring API run minutes earlier); fix is runner-side docker health/retry in the workflow, not code [PR#954, run 3281]
[ISSUE-083] low open area:people Clustering reset (POST /api/admin/people/clustering-reset) removes system-assigned occurrences whose face mirror is dr.state != 'accepted', so a strong-band auto-ACCEPT (policy ON) on a face whose mirror is 'accepted' survives even though the user never confirmed it individually — deliberate conservative trade: mirror-accepted faces are treated as user intent; full unassign requires resetFaces (destroys detections) [feat/clustering-abstraction-and-reset]
[ISSUE-086] med mitigated area:workers worker class_instance_candidate results for full-pool reclusters scale with pool size (5,832 candidates ≈ 20 MB JSON rejected by Jackson default maxStringLength 20M → job failed); mitigated: API limit raised to 64 MB; real fix is chunked result submission [PR#992]
[ISSUE-091] med open area:ci Scoped PR tests silently skip ALL JVM tests whenever detect-scopes.sh's catch-all branch fires: any changed path that matches no rule (e.g. .gitignore, a generated .spec-index.json) hits *) → mark_all_services + run_fallback_smoke=true, and build-deploy.yml gates the "Run API tests" step on run_fallback_smoke != 'true' — so the fallback branch only runs gradle :apps:api:compileJava :apps:worker-media:compileJava (compile-only) plus web-check/CLI/ML checks, never :apps:api:test, while the job still reports success (verified: run 3540, commit 7c4ec643, Task :apps:api:test count = 0); a PR touching e.g. .gitignore therefore merges with zero Java tests run; workaround: Daily Full Tests or a local :apps:api:test run; fix: make the fallback path run the JVM suites too, or stop letting bookkeeping-only paths suppress the scoped test step [PR#1002]
[ISSUE-093] med mitigated area:ci build-web is gated on build-api (needs: [changes, smoke-web, build-api], if: build-api success|skipped) → when the API build fails in the same run that carries web changes (run 3528 / PR#1000), Build+Deploy web are skipped and scoped detection never re-triggers them in later runs (run 3532 had no web scope) → production served the pre-spec-051 web bundle (old TimelineSkeleton day-header bars, no layoutDayCount) until this PR forced a web rebuild; systemic fix pending: decouple build-web from build-api or add a deploy-parity check [run#3528, run#3532]
[ISSUE-094] med open area:timeline With a favorite filter active the days/dayItems/timeline plan keeps a row-wise OR over a metadata_overrides EXISTS-style predicate that inflates plan cost above jit_above_cost=100000 → 5-10x slower filtered requests via JIT compilation; unfiltered path fixed in this PR (join-backed visibility + anti-join format gate + capture-range layout) [this PR]
[ISSUE-099] med open area:ci The JVM-test skip documented in ISSUE-091 is NOT limited to detect-scopes.sh's catch-all branch: the deliberate path rules apps/api/src/main/java/dev/memories/api/{auth,security,sharing,workers,worker}/* + web/WorkerController.java (:154), db/migration/*|application*.yml|api/Dockerfile (:143) and deploy/compose*.yaml|build.gradle.kts|settings.gradle.kts|gradle/* (:127) also call mark_fallback_smoke, and build-deploy.yml gates every JVM test step on run_fallback_smoke != 'true' — so such PRs run only the compile-only fallback smoke while run_api_tests/run_worker_media_tests still print true (verified on PR#1028: detect output run_api_tests=true + run_worker_media_tests=true, yet "Run API tests" (:217) and "Run media worker tests" (:255) were both skipped and the job reported success with zero :apps:api:test/:apps:worker-media:test; suites were run locally instead: api 1546 pass, worker-media 29 pass); fix belongs with ISSUE-091 [PR#1028]
[ISSUE-105] high mitigated area:discovery Stash presses were silent no-ops and ghost suggestion cards re-minted every recognition run: occupiedForInsertOnlyStash counted removed clustering-reset orphan rows (3059 live) as occupied so the stash resolved all_faces_already_assigned with no shadow and NO rejectedTargetedPair signature (47 of 194 real stash presses no-oped, identical cards re-emitted ~10 min), and worker job payloads snapshot assignments at enqueue while embeddings are fetched live at job start, minting cards holding faces confirmed elsewhere in the meantime (2 live cards held 3 faces accepted on person "Holger"); fixed API-side: stash now always creates/extends a shadow (removed/missing rows revived in place, live tentatives released+re-homed per ADR 0102 owner arm, stashed pre-linked target absorbs as absorbed_into_existing_stash), accepted-elsewhere is the sole no-steal exception (suppressed_confirmed_elsewhere), every stash outcome writes the rejectedTargetedPair signature, and persistWorkerCandidates mint-time guard skips cards holding a face accepted on a non-target instance; worker/web unchanged (web PR pairs on the new outcome names) [3a6ae176]
[ISSUE-106] med mitigated area:discovery duplicate pending suggestion cards for the same face minted across recognition runs; second decision errored after overlap-rejected sibling [spec:047]
[ISSUE-109] med mitigated area:people-discovery Suggestions carousel overrode a NAMED stored match target with the live single-centroid person-ranking top row (endpoint metric lags worker multi-representants, ADR 0108) — measured: card stored "Rikke" (dist 0.228) titled "Similar to Jutta" (stale 12-face rank-1); title/primary now follow the live ranking only when the stored target is unnamed/missing [PR#1082]
[ISSUE-110] low open area:discovery ADR 0109 API read rankings (personRanking/mergeTargetRankings/mergeSuggestion) score against the persisted multi-representant set of the LAST completed worker run — minutes-stale by design (same freshness as the candidate cards): a person created or heavily accepted AFTER the last run keeps the live single-centroid fallback (or its previous decomposition) until the next run full-replaces the set; accepted as the documented degradation, revisit only if operators report picker/card disagreement windows [ADR 0109]
[ISSUE-111] low open area:discovery mergeSuggestion bucket deferral (latency fix) decides at embedding-space-key granularity from a metadata-only census: a bucket whose rows are ALL unparseable/empty can still make a rep direction return null in the composer while the census reported the space as reachable, loading buckets unnecessarily (safe direction) — but an instance whose buckets were SKIPPED could in principle hit that degenerate case and change a pair's outcome vs the former full scan; requires every row in the shared space to fail embedding() parsing or be empty (zero-norm vectors still enter buckets and are NOT sufficient; the worker never emits malformed rows; no known production or fixture occurrence). mergeTargetRankings only ever over-fetches in this case, so no outcome can change there [PR pending]
[ISSUE-113] low open area:discovery Person-history revert: a reverted rejected candidate's rejection review-decision row stays append-only and still carries the spec 040 rejectedWeakPair/rejectedTargetedPair signatures, so a LATER worker run will NOT re-emit an identical candidate for the same (face, person)/(refs, target) pair even though the owner un-rejected the card — the revived card itself is untouched (suppression only gates new INSERTs), but if the owner then deletes/rejects-differently the revived card the identical pair can only resurface via a changed composition; accepted because review_decisions must stay append-only for the revert audit contract [spec:040, feat/person-history-rejected-api]
[ISSUE-116] low open area:api OpenAPI advertises POST /api/media/{id}/trash and /restore as 200 + MediaDetail, but MediaController is annotated @ResponseStatus(NO_CONTENT) so the running server answers 204 with an empty body (the web client already types both as void) — a client generated from the spec tries to decode a body that never arrives; the Android Phase 1 viewer was implemented against observed behaviour (executeEmpty, any 2xx) and passes tests that pin the real contract, so the spec document is now the only wrong party; fix belongs to the API spec, not the app [spec:055]
[ISSUE-117] low open area:media MediaDetail carries things but no persons/faces collection, so no detail view can say who is in a photo even though faces and person names exist server-side; the Android info sheet renders only fields the payload actually contains, so people rows are simply absent rather than wrong — add an optional persons summary to MediaDetail when the viewer is expected to show them [spec:055]
[ISSUE-122] med open area:search GET /api/search accepts no cursor and caps limit at 100, and the Android Phase 1 screen deliberately shows a single page with no load-more — a broad query ("Strand", "2023") silently truncates at the server cap with no indication that more matches exist; workaround is a narrower query or an extra filter [spec:055]
[ISSUE-124] low open area:albums Android album-list covers fan out one GET /api/albums/{id} per album after the list lands, because GET /api/albums carries no cover URL; fine at family scale (a handful of albums, failed covers keep the placeholder), wasteful at tens+ albums — a cover/summary field or endpoint would collapse the N+1 [spec:055 T007]
[ISSUE-125] low open area:albums GET /api/albums/{id} returns every item with no cursor, so a very large album loads its whole item list into the Android detail screen and picker in one response; acceptable for family albums, would need paging (or a cap + load-more) for big libraries [spec:055 T007]
[ISSUE-128] low open area:android The people grid walks at most 5 cursor pages of 200 (1000 people) per load so that client-side named-first/face-count ordering can see the whole family at once — an archive with more than 1000 recognised people would show only the first 1000 with no "more" affordance; workaround is narrowing, or raising PeopleViewModel.DEFAULT_MAX_PAGES once a paged-ordering strategy exists [spec:001]
[ISSUE-173] med open area:workers WorkerJobHistoryPurgeService.purgeExpiredHistoryBatch deletes worker_jobs rows that discovery_results still references — every 5-min purge tick fails with DataIntegrityViolationException (FK discovery_results_worker_job_id_fkey, observed live 2026-10-18 on prod worker_job 1fe305e5-620d-4c1f-ba76-ddc7514c6fd0) so expired job history never shrinks and each tick logs a full stacktrace; fix: only purge jobs whose id has no discovery_results reference (NOT EXISTS subquery or NULL-out the FK column first), then delete [diag 2026-10-18, memories-api logs]
[ISSUE-174] high open area:android-release existing sideload installs signed with the debug key (<= 0.9.x Forgejo APKs) cannot update in place to release-key builds (INSTALL_FAILED_UPDATE_INCOMPATIBLE) because Google Play rejected the debug key's CN and releases/Play now share the dedicated release key (Play App Signing, uploaded via PEPK) — one-time uninstall + reinstall required on each device (login + re-sync; server media unaffected) [PR#1318]
[ISSUE-175] high mitigated area:people The assignment_cv V7 lane (feat/cv-highest-valid-cut-v2) OOM-died twice (SIGKILL/137, >2 GB) feeding the O(n²) full-matrix experiment HAC (weighted_identity_hac) with max_distance=np.inf over giant U×U components — the 13.6k-face production-class component implies a ~1.5 GB per-merge np.where full-matrix copy (production hac_linkage itself was never affected, it partitions via the ISSUE-100 lazy heap); fixed by transcribing _heap_component_partition (with dendrogram recording) into the experiment module — measured 193 MB whole-process peak, 5.2 s full V7 run, and byte-identical fixed-cut output (3112 merges, 667 cards, max |margin delta| 0.0 vs the pre-heap run) [feat/cv-highest-valid-cut-v2 6299e2f5]
[ISSUE-177] med mitigated area:android On Android 12+ (API 31+) periodic backup runs start from true background where setForeground throws ForegroundServiceStartNotAllowedException — the run previously continued with no foreground service AND no notification (the FGS stop removed it, and before API 33 no permission path existed); now the worker gates the doomed call via BackupNotifications.foregroundPromotionEligible, posts the ongoing progress notification directly via NotificationManager (visible once the user grants POST_NOTIFICATIONS, requested on the backup screen on API 33+), and cancels it when the run ends; accepted limitation: the fallback notification carries no FGS process protections — background runs remain killable mid-upload [fix/android-notification]
[ISSUE-226] med mitigated area:android The battery-optimization exemption request button ("Kør altid i baggrunden" / "Immer im Hintergrund ausführen", onboarding + backup tab) did NOTHING when tapped on stock Android 13+: PowerExemptionImpl.requestExemption first checked resolveActivity(), but Android 11+ package visibility hides the system dialog activity unless the manifest declares <queries> for android.os.action.REQUEST_IGNORE_BATTERY_OPTIMIZATIONS — the guard misread "hidden" as "OEM-stripped" and silently returned (no crash, no dialog; workaround: Settings → Apps → Memories → Battery by hand); fixed by starting the dialog intent unconditionally (manifest <queries> block makes it resolvable/visible), an ActivityNotFoundException fallback that opens the app-info battery screen, and a regression test pinning the start-without-resolution contract [feat/settings-power-exemption]
[ISSUE-178] high mitigated area:android Android upload engine reused a persisted open batch with NO server validation and mapped HTTP 403 ("Not the owner of this upload session" — session created under the API credential but presented with cookie auth, or vice versa) to a plain abort, so the poisoned batch row survived app restarts and every WorkManager retry re-opened the same rejected session forever (uploads permanently stuck); fixed client-side per SEC-4 intent (no server-side ownership fixing): the engine now probes GET /api/upload-sessions/{id} before reusing an open batch (403/404/409/410 or terminal status → discard batch + create fresh session, live status → reuse with the original idempotency key) and the client maps 403 to SessionGone on every session-scoped call (clears batch, next run starts fresh) [fix/upload-engine]
[ISSUE-179] med mitigated area:android The Android backup engine had no in-process single-flight guard: periodic work (memories-backup-periodic) and immediate work (memories-backup-now) have distinct WorkManager unique names and could run concurrently in one process, both streaming the same pending files into the same upload session (observed interleaved duplicate uploads) with racing commits on the batch row; fixed by a process-wide BackupRunGate (kotlinx Mutex on a module-level singleton shared by every engine instance, wait-not-skip so the periodic run's work is only deferred, not dropped) inside BackupEngine.run [fix/upload-engine]
[ISSUE-180] med mitigated area:android A transient 5xx (e.g. reverse-proxy 502) on any file upload or the commit aborted the whole Android backup run immediately, so a large pending set restarted from scratch on every WorkManager retry against a struggling server; fixed with bounded in-run exponential backoff + jitter (UploadRetryPolicy: 4 retries, 2 s → 60 s cap, ±25% jitter) around file upload and commit — replayed file uploads are server-side idempotent (upsert on (upload_session_id, client_file_key)) and 4xx rejections, session-gone and auth failures are never retried in-run [fix/upload-engine]
[ISSUE-181] med open area:web The web timeline's "Familiegruppe" scope chip sends scope=group WITHOUT sharingGroupId (verified in apps/web/src/routes/+page.svelte timelineFilterParams/load), but the server pins group scope to the requested id (SHARED_MEDIA_JOIN_GROUP: c.sharing_group_id = :sharingGroupId; the controller binds :sharingGroupId straight from the query param with no default resolution), so the chip silently renders an EMPTY timeline for any account in a group — only accounts in NO group happen to "work" (equally empty). The Android timeline deliberately ships the chip only with a resolved single active group (GET /api/sharing-groups probe); web needs the same resolution (or a group picker) before the chip can be trusted [feat/parity-android-usability, apps/api TimelineQueryService]
[ISSUE-182] med mitigated area:android The backup running notification never showed progress although the POST_NOTIFICATIONS rationale (backup_notification_rationale_body) explicitly promised "einen Hinweis mit dem Fortschritt": runningNotification built title/icon/ongoing only and neither the FGS path nor the NotificationManager fallback ever refreshed it during a run — fixed by a progress-aware runningNotification(processed, total, currentFile) ("3 von 17" contentText + determinate progress bar + in-flight file as sub-text) that the worker refreshes on every engine file-start progress event (FGS: re-setForeground with the same id → replace in place; fallback: re-notify), with the counts mirrored into WorkManager progress (memories.backup.processed/.total) for the backup screen; de/en/da progress format strings added [fix/backup-progress]
[ISSUE-183] med mitigated area:android The SICHERUNG status screen gave zero feedback while a backup ran: pressing "Jetzt sichern" showed nothing (app "felt stuck"), the last-run card showed only the static "Läuft …" text with no progress, and the count cards stayed on "wartend" even after uploads had succeeded server-side (the ViewModel collected work state but never re-read Room on terminal transitions) — plus a double-tap during a run enqueued with ExistingWorkPolicy.REPLACE, cancelling the live worker mid-file (observed in prod logs: second run 21 s after the first, first killed mid-video, next run re-uploaded from file 1); mitigated by: BackupWorkSnapshot gains processedCount/totalCount read from the worker's progress Data (keys memories.backup.processed / memories.backup.total, null when absent → indeterminate bar), LastRunCard renders a determinate/indeterminate LinearProgressIndicator + "X von Y" while running, "Jetzt sichern" disables and relabels to "Sichert …" while a run is queued/active, backupNow()/retryFailed() skip while active (no REPLACE of a live run), the ViewModel re-reads Room when observed work transitions from active to a terminal state (counts update without pull-to-refresh), and BackupScheduler.enqueueNow additionally guards against enqueueing while the immediate work is non-terminal. Residual: the progress counts depend on the parallel worker-lane progress writes; until that lane ships, the bar shows indeterminate and the labels/stale-count fix already work [fix/backup-feedback]
[ISSUE-185] med mitigated area:android Test media pushed into an AVD with adb push stays is_pending=1/_size=NULL in MediaStore, so the device-album scan (_size>0 AND is_pending=0, StorageCompat.scanSelection) finds nothing until a volume rescan; workaround: adb shell content call --uri content://media --method scan_volume --arg external_primary after every push, documented in docs/dev/android-emulator-testing.md [spec:055]
[ISSUE-188] low open area:sync The client-intent channel (spec 065 PR-C) ships as ADDITIVE capacity: the server accepts POST /api/sync/intents (delete/restore, per-item JMAP-style outcomes with the S5 reason vocabulary, clientIntentId idempotency ledger client_intents V177) and the Android app RECORDS a queued intent for every delete/restore action (Room client_intent queue, DB migration 2→3, replayed with verbatim ids by the engine drain after the delta pull; only the server echo makes an outcome authoritative, S1), but the app's actual mutations still go through the legacy trash endpoints (POST /api/media/{id}/trash|restore) — the queue is bookkeeping for the future intent-mandated channel, so today a replayed intent can surface conflict_state_changed if the legacy call already changed the item (cosmetic: the queue marks it rejected with plain-copy guidance, the library state is correct either way); legacy callers (web + app) stay grandfathered until the spec's "intents-only mutation channel" mandate (if any) lands with PR-D [spec:065, PR-C]
[ISSUE-189] low open area:sync Spec 065 PR-D's C2 backup-path name-conflict auto-resolve (V179 resolved_name_conflict marker, UploadCommitService numeric-suffix disambiguation, resolvedNameConflicts count in the upload-activity feed, S4-C2) detects owner-scoped collisions against original_assets.filename at COMMIT time only, so the interactive web-upload MAY-reject path (duplicate_name_differs_bytes) is NOT shipped (no web UI change in PR-D scope) and an interactive-web upload followed by an API upload of the same name still silently suffixes rather than surfacing a notice on the earlier (web) session's feed; the ratified product rule's web-side half (interactive reject) remains a follow-up if the product owner mandates it [spec:065, S4-C2]
[ISSUE-131] low open area:android The map viewport request caps at markerLimit=500 with includeClusterMembers=true (the web display path leans on the server default of 5000 without members; its member-carrying viewer path uses 240); the server clusters only the fetched rows, so cluster counts undercount and very large viewports lose markers beyond the cap — acceptable for a family library, revisit with a paging or count-only mode if it ever shows [spec:055]
[ISSUE-132] low open area:android The Karte tab's osmdroid surface (MapView lifecycle, cluster badge drawing, drill-down camera animation, tile-source swap) has no JVM test coverage — only MapViewModel behaviour and the map DTO shapes are unit-tested; the canvas itself would need an instrumented/emulator lane, which the project does not have yet [spec:055]
[ISSUE-134] low open area:android GET /api/upload-activity does not expose clientType (it is set only when a session is created, browser/api/cli, and is never echoed back), so the upload-activity screen cannot label a session "Desktop"/"Browser"/"Telefon" as the task brief hoped; it labels only sessions this device can prove it started (its own manual uploads plus the backup engine's open batch) as "Dieses Gerät" and leaves every other row unattributed rather than guessing — a per-device label would need the server to include clientType (or a device id) in the activity payload [spec:049]
[ISSUE-136] med open area:api The OpenAPI contract still documents the retired /api/person-assignments/* paths (people list, detail, rename, tentative, confirm) while no controller serves them — the live people surface is DiscoveryController's generic discovery endpoints (GET/…/api/discovery/domains/people/class-instances, PATCH /api/discovery/class-instances/{id}, GET/POST /api/discovery/domains/people/class-instance-candidates…); the Android app now speaks the live paths, but the stale contract keeps misleading integrators and must be regenerated or corrected [spec:055]
[ISSUE-139] low open area:web The share fixes on explore/shared-people/[collaborativePersonId] (wire-correct policy, describeApiError, user-audience option with the share-target autocomplete) land on a route the app never renders: its +page.ts 307-redirects to /person/:id, which 308-redirects to /discovery/domains/people/instances/:id, so that page is only visible if the load runs before the redirect chain; users actually share from the instance-detail Share dialog, which got the same autocomplete [fix/share-dialog-autocomplete]
[ISSUE-141] low open area:android osmdroid 6.1.20 exposes no HTTP connect/read timeout API anywhere (TileDownloader sets none, IConfigurationProvider has no setters), so a stalled tile socket can hold one of the map's download slots until it dies on its own — bounded only by tileDownloadMaxQueueSize (64), never by a configurable timeout; the map tile-cache config documents this honestly instead of faking one, and the durable fix is upstream or a self-hosted tile server [PR#1151]
[ISSUE-144] low open area:android The map sidebar's truncation banner infers "capped" from the flattened, de-duplicated item count reaching markerLimit=500, so a viewport whose raw marker rows hit the cap but carry cross-cluster duplicates (flattened < 500) shows no banner even though items are missing — a real "total" would need a count-only viewport mode server-side (related to ISSUE-131); worded as "first N loaded" so the banner never overstates what it knows [ux/map-media-sidebar]
[ISSUE-145] low open area:android Android suggestion cards can apply a whole-photo face bbox to an already server-cropped face URL: SuggestionUi carries both faceImageUrl (which prefers the /faces/ crop) and faceBox, and SuggestionCard passes both into FaceImage, so a ref carrying both thumbnailUrl and scope/boundingBox gets the crop cropped again — the web carousel guards this ("A bbox drives the zoom, so the face crop must not win"); align the strip with the same stand-down rule when it is next touched [ux/face-thumbnails-density]
[ISSUE-147] low open area:i18n de/da message translations (apps/web/messages/, Paraglide wave 1: app chrome, setup/register, MaintenanceView, profile tabs) are agent-authored, not native-reviewed — flag non-trivial copy in review; the locale switcher is functional-but-plain pending a designer pass [PR#1178]
[ISSUE-148] low mitigated area:web tests/backup-export.spec.ts:5 flakes under parallel E2E load: clicking "Turn on safe mode" does not surface the On text within 5s (CI run 4506 + local baseline; passes on retry) — looks like a swallowed first click/race rather than a wrong expectation; needs a load-repro before touching either the test or the toggle [PR pending] [audit 2026-10-30: mitigated in PR#1445 — gotoHydrated hydration gate + the missing processing-summary mocks added to backup-export.spec.ts; traces survive CI + --shm-size=1g (PR#1444); nightly flake-detection evidence pending before resolved] [audit 2026-10-30: mitigation survived the full nightly + PR wave — stays mitigated pending nightly confirmation]
[ISSUE-149] low open area:discovery tests/discovery-domain-instance-detail.spec.ts merge tests (783/829) flake under load: after clicking the "…actions" button the Merge menuitem never appears within the 30s test timeout (menu fails to open; passes on retry) — same swallowed-click/race class as ISSUE-148 [PR pending] [audit 2026-10-30: PR#1445 applied gotoHydrated + openActionsMenu fail-fast on aria-expanded to this spec; oracle reclassification — the ⋮ menu trigger is client-only ({#if} hand-rolled, no portal/rAF), so this is NOT the swallowed-click class; primary suspect is runner starvation bursts, now dampened by --shm-size=1g + surviving traces (PR#1444); stays open pending trace attribution from the next red run] [audit 2026-10-30: PR#1453 split discovery-review.spec.ts into four focused specs + discovery-review-shared.ts (describe name preserved) — split verified the rename/merge/dissolve suites green at retries=0 locally, 74/74; ⋮-menu class still attributed to runner starvation pending trace attribution; shm + trace evidence in place since PR#1444; stays open]
[ISSUE-150] low mitigated area:ci The android job's best-effort removal of the local git.wiwie.dk/chris/memories/android-ci:latest tag after a rebuild (build-android-ci-image) can fail when a concurrent job container on the shared docker daemon is using the image — act_runner then keeps reusing the stale local :latest instead of pulling the rebuilt one, shadowing it for up to 168h until the docker-cleanup prune drops it; mitigated by an explicit ::warning:: on rmi failure (was silently swallowed) and by removing the unqualified build tag too, and a failed rmi never fails the job [feat/android-ci-image]
[ISSUE-151] low open area:i18n two multi-part template-literal confirm() strings stayed hardcoded through spec 058 waves 1-5 (census + guard blind spot — the scanner only matches single-literal args, not \…${x}…` + '…'concatenation):src/routes/+page.svelte trash confirm (Move N selected item(s) to trash? …, English) and src/lib/components/profile/DecisionHistorySection.svelteper-entry revert confirm + notice (Danish). Extract in a follow-up micro-pass and teach the guard the concatenated-literal shape [spec:058] [audit 2026-10-24: DecisionHistorySection confirm now i18n (m.decisionHistoryRemoveConfirm); +page.svelte trash confirm still hardcoded EN — half remains] [ISSUE-152]low mitigated area:discoverytests/discovery-domain-rename.spec.ts:240 (mobile) flakes under parallel E2E load:rename-triggerinside the plants_image instance card never appears within the 30s test timeout (CI run 4633; passes on retry and on clean main) — same swallowed-click/race-under-load class as ISSUE-148/149 [PR#1200] [audit 2026-10-30: mitigated in PR#1445 — gotoHydrated hydration gate on this spec; --shm-size=1g removes the /dev/shm amplifier (PR#1444); nightly flake-detection evidence pending before resolved] [audit 2026-10-30: mitigation survived the full nightly + PR wave — stays mitigated pending nightly confirmation] [ISSUE-153]low mitigated area:discoverytests/discovery-instance-tentative.spec.ts:700 (mobile, da-pinned block) flakes under parallel E2E load: the instancedisplayName heading (Alice) is not visible after page.goto(5s timeout; CI run 4633; passes on retry and on clean main) — first-assertion-after-goto variant of the ISSUE-135 load-race class [PR#1200] [audit 2026-10-30: mitigated in PR#1445 — gotoHydrated hydration gate on this spec; --shm-size=1g (PR#1444); nightly flake-detection evidence pending before resolved] [audit 2026-10-30: root cause found & fixed in PR#1455 — loadTentativeCandidates blanked tentativeError at request start, unmounting the {#if tentativeError} retry button mid-flight; fix decouples banner ownership from list freshness (retryTargetStillPending derived) so retry-button DOM identity is keyed to card presence; CI run 5979 flake-detection failure (discovery-instance-tentative.spec.ts:2135, element detached from DOM) reproduced + regression-tested, 220/220 at repeat-each=5 retries=0; :700 ref intact — the spec was extended, not split] [ISSUE-194]med open area:discoveryDiscovery day sections still render hard-English server labels: DiscoveryService.toDiscoveryDay hardcodes "Without date"/"Missing capture date" +EEEE, d MMMM yyyy/MMMM yyyyLocale.ENGLISH (same bug class as /api/timeline/days, fixed there by fix/timeline-day-labels-locale which deliberately left this lane untouched), so discovery domain/instance day walls and the self-person gallery show English headers under de/da locales until the key-driven web derive is applied to those consumers too [fix/timeline-day-labels-locale] (renumbered from ISSUE-155: id collision) [ISSUE-119]low open area:android lintDebugfails on the branch baseline (verified on 2362ef77, 16 errors): 8 xUnsafeOptInUsageErrorinui/timeline/ViewerPage.kt, 7 x LocalContextGetResourceValueCallinui/timeline/MediaViewerScreen.ktand 1 xMissingTranslationfor theapp_namebrand string — all introduced by the Phase 0/1 viewer PRs #1112/#1116.lintDebugis not a CI gate (workflow runs only:app:testDebugUnitTestand:app:assembleDebug, both green), so nothing user-visible is affected; neither viewer file is touched by the timeline lane, and this branch adds no new lint errors [spec:055, PR#1112, PR#1116] [audit 2026-10-24: LocalContextGetResourceValueCall errors gone; MissingTranslation (app_name missing in en/da) remains, UnsafeOptIn not re-verified without a lint run] [ISSUE-121] high mitigated area:androidDebug APKs were not updatable: the CI android job runs in an ephemeralgradle:8.13-jdk21container, so AGP regenerated$HOME/.android/debug.keystoreon every run and each CI APK shipped a different random signing key (verified: 3 cached CI APKs → 3 distinct signer certs, none matching the local dev keystore) — installing any new build over an existing one failed on-device withINSTALL_FAILED_UPDATE_INCOMPATIBLE("App not installed"). Mitigated by one persistent debug key provisioned from the repo secretANDROID_DEBUG_KEYSTORE_BASE64into$HOME/.androidbefore every CI build (same pattern as wishlist-nextcloud;.keystorebacked up in Bitwarden, CI also fingerprint-gates the built APK); devices with an old random-key install need one final uninstall [PR#1152] [ISSUE-031]med mitigated area:peopleSuggestions dialog rendered literal "0% match" for name-fallback rows while real similarity was structurally impossible for unassigned faces (embeddings pending until acceptance); fallback rows now carry null similarity ("No face match yet") and source pending embedding feeds real scores [fix/suggestions-zero-match] [ISSUE-195]high mitigated area:peoplediscovery_people candidate refs carried only mediaItemId → occurrences stored with fallback keys/type (mediaItemId:index/generic_evidence) → all embedding-flush paths (listener + reconciliation) filter face_annotation/<faceAnnotationId> and never matched (production: 16,411 buffered, 24 flushed); API now derives face_annotation type/key at occurrence accept time, flush event fires on UPDATE path too, and reconciliation re-keys unambiguous legacy rows [fix/face-pipeline-embedding-flush] (renumbered from ISSUE-032: id collision) [ISSUE-033]high mitigated area:peoplepersistWorkerCandidates auto-create passed in-memory candidate (camelCase occurrenceRefs) to acceptCandidateOccurrences which reads the DB column name (occurrence_refs) → 0 occurrences created → people page empty (instances() requires an accepted occurrence); now re-reads the persisted candidate row [fix/face-pipeline-embedding-flush] [ISSUE-034]high mitigated area:peopleInfinite recognition loop: owner_face_recognition jobs submitting 0 candidates left faces owner_tentative while PeopleDiscoveryPostCompletionListener + PeopleDiscoveryRecognitionScheduler re-enqueued endlessly; both now consult FaceRecognitionProgressService and only arm when tentative faces hold unconsumed embeddings (pending or unassigned accepted) [fix/face-pipeline-embedding-flush] [ISSUE-036]med mitigated area:discoveryexisting_person_match candidates (worker-submitted, non-people-cluster types) never had an auto-create path and no review target exists for non-existent instances → stuck pending forever; auto-create now covers all unmatched people/person candidates [fix/face-pipeline-embedding-flush] [ISSUE-073]med open area:uploadAPI/Android uploads store client-relative-path-derived names double percent-encoded in both original_filename and client_relative_path (e.g.primary%253ADCIM%252FCamera%252Fx.jpg, where %253A = encoded %3A); upload-activity display mitigated by single decodeOnceForDisplay in UploadActivityService, but stored values remain double-encoded and other consumers of these columns can render garbage; upstream fix should stop the client/API layer from double-encoding on receive [spec:049, PR#932] [ISSUE-155] med open area:videosExisting videos keep no posters/rendition until reprocessed:video_previewis only enqueued on upload commit, so every video ingested before spec 059 stays on the "Preview pending"/placeholder tile until an operator triggers reprocessing (MediaMaintenanceService already lists video_preview as resettable, but nothing auto-backfills) — needs either a one-shot backfill job or maintenance-run guidance for self-hosters [spec:059] [audit 2026-10-24: backfill tooling landed (enqueueVideoPreviewBackfill + POST /api/media/metadata/video-preview-backfill); 1002/1768 videos still rendition-less on owner DB — operator run pending, stays open] [ISSUE-156]med mitigated area:videosHover-to-play and viewer playback stream the original file whenever no video_rendition exists yet (web pickViewerVideoSource fallback): correct but heavy for large 4K originals (no range-limited transcode, full bitrate over LAN/WAN); acceptable while renditions are missing post-merge, self-resolves per item once video_preview completes — revisit if users report stutter on originals-only playback [spec:059] [ISSUE-158]med mitigated area:castAndroid cast + album autoplay (spec 060 T010-T014) is JVM/build-verified only — the design §8 Chromecast device matrix (photo/video swipe-sync, mixed-album autoplay, ?st= token-expiry self-heal, TV power-off/session loss, phone screen-off session survival) still needs manual on-hardware verification; the slideshow is deliberately foreground-scoped (leaving the album screen stops it; no foreground service yet) and every LOAD depends on the parallel backend lane's POST /api/media/token being deployed (mint failures surface as a skip-with-snackbar until then) [spec:060] [ISSUE-159]low open area:peopledecideCandidate stash with an EXPLICIT existingInstanceId target still sets stashed_at on a NAMED instance without clearing display_name — violates the stash=unnamed invariant (stashed named person would re-enter default views and ADR 0107 keeps showing candidate cards against it) on a non-UI-reachable path; endpoint stash clears the name per ADR 0112; legacy decideCandidate arms deliberately untouched (guard + explicit-target semantics byte-identical), fix candidate: also discard display_name in the explicit-target arm [PR#1225] [ISSUE-162]low open area:peoplesemisupervised-v3 fail-closed message says "semisupervised-v2 refuses to cluster" when a U component exceeds SCALABLE_MAX_HAC_NODES: partition_hac_components raises a shared static error string that names only v2 — cosmetic misnomer on the v3 path, fail-closed behavior itself unchanged and pinned by test (ADR 0113); fix candidate: parameterize the message with the calling method [ADR 0113 / feat/semisupervised-v3-worker] [ISSUE-163]low open area:assignment-cvassignment_cv confusion/hub tables index the winner CANDIDATE row directly into the universe name list, so person labels shift in folds where a 1-face person drops out of the candidate universe (6 such persons on the owner library; counts/rates unaffected) — surfaced by the G1 gate report which recomputes attribution with the corrected candidate-row→person-id mapping; both conventions agree on all G1 verdicts [PR #1259 protocol / docs/adr/0113 G1 / assignment_cv/make_g1_report.py] [ISSUE-166]med open area:album-suggestionsa single day with more than MAX_MEDIA_PER_WINDOW (500) photos produces no candidate window at all — oversized same_day/weekend/rolling windows are dropped, not trimmed (and >500-media adaptive_gap segments likewise), so very dense days can never become suggestions; interleaved anchor selection fixed the 500-window cap starving recent media but intentionally does not address oversized-window dropping [ref:PR#1273] (2026-10-17) [ISSUE-167]high mitigated area:peopleThe divergent representative-thumbnail ladder bug class (ISSUE-095/ISSUE-140 family) is structurally mitigated per ADR 0114: ONE shared resolution (ClassInstanceRepresentativeService — manual pin via manual_selection provenance → medoid of the largest cluster [persisted medoid_face_annotation_id V170, else one embedding cluster] → deterministic terminal ladder) now backs every surface (with generic and listing paths that ignored pins, ranking endpoint SQL fragment copies, person-relation chips, collaborative bridge, self-person avatar, face-picker marker), the auto-pin of the first accepted occurrence is removed, re-pinning re-enqueues the canonical crop, and thumbnail_derived_asset_id accepts fail closed unless the crop's payload occurrence is the current canonical representative (last-write-wins removed) [ADR 0114 / feat/rep-thumb-api] [ISSUE-168]med open area:albumsthe single-item album endpoints (addItem/update/delete/reorderItem/removeItem) resolve album/media visibility with JdbcClient.single()inside requireActiveAlbum/requireOwnedActiveMedia, so an unknown or soft-deleted album — or an unknown/inactive media item — surfaces as EmptyResultDataAccessException, which ApiExceptionHandler does not map (no @ResponseStatus on that Spring exception, only NoSuchElementException→404 and AccessDeniedException→403) and which therefore leaves the endpoint as an uncaught 500; found while building the spec 064 batch endpoint, which normalizes exactly these cases to NoSuchElementException→404 inside its own path and leaves the legacy single-item behavior untouched (out of scope for that PR) [spec:064] [ISSUE-169]low open area:peoplePersonRelationServicePostgresTest.theChosenRepresentativeWinsOverTheLatestOccurrence failed once in a full :apps:api:test run (2026-10-17, branch spec/064-api) at the pinned representative-thumbnail URL assertion (~line 256) and passed again in an isolated re-run — load/order-dependent flake in the representative-thumbnail ladder, unrelated to the album work in that branch [spec:064] [ISSUE-196]high mitigated area:syncTwo adversarial-review findings fixed together [fix/sync-intentscope-reinstate-copy]: (1) POST /api/sync/intents required the READ scope media:read, so a read-only API credential could trash/restore/reinstate library items — now gated on media:upload (v1's existing write scope, no new scope value; read endpoints GET /api/sync/changes + /removed-history keep media:read); (2) the S11 removed-history screen offered "Wieder aufnehmen" on permanently_deleted entries where the reinstate intent either rejects (override_rejected_explicit_deletion, C6) or accepts WITHOUT effect (re-authorization only, ISSUE-190) — the button is gone for purged entries; the card states the actual state+path as plain de/en/da copy instead (no fake navigation), trashed entries keep the working Wiederherstellen [ISSUE-198]med mitigated area:sync The app delete/restore screens recorded an intent into the Room queue for EVERY successful legacy mutation (spec 065 PR-C's "record-first"), so the replayed intent's CAS was always stale (expectedLifecycleState=activeagainst an already-trashed item) and each drain seeded the ledger with meaninglessconflict_state_changed REJECTED rows — reject copy surfaced for actions that had succeeded (the F1 noise ISSUE-188 described as cosmetic, but user-visible and misleading); mitigated per the ratified record-on-failure amendment (spec 044 S3): the legacy endpoints stay the immediate mutation path, a screen queues an intent ONLY when its legacy call FAILED — with NO CAS (expectedLifecycleState=null, the landing is unknown, so the server's idempotent C4 goal table judges the replay; a Forbidden/403 stays unqueued since a replay would be rejected not_owner) — success records nothing, the queue is now the truthful durable RETRY channel, existing bogus REJECTED rows are left as honest history (drain/purge infra already ages them out); removed-history reverts are untouched (their queue write IS the action, S11); historical note: the task was pre-assigned ISSUE-194 but main merged ISSUE-194/195/197 first, so this entry takes the next free number [fix/sync-intent-ledger-hygiene] [ISSUE-200] high mitigated area:syncThe S11 removed-from-library history derived its entries from an INNER JOIN on the V178 media_deletion_decisions provenance table, so every deletion made BEFORE the sync server shipped (V178 deployed 2026-10-24; this instance's 18 bulk-purged items of 2026-10-23 have no decision rows) was invisible: the backup screen correctly showed "18 bewusst nicht in Memories" (change feed = pure lifecycle_state) but the view history answered "Momentan ist nichts aus der Mediathek entfernt" — plus, emptyTrash recorded decided_by='user' with actor_user_id NULL, so even NEW bulk deletions rendered the false "Automatisch" for a human action and the payload carried no WHO discriminator; fixed server-side (history base set = item lifecycle_state trashed/permanently_deleted with provenance LEFT JOINed as attribution enrichment, COALESCEd removal timestamp, add-only attribution vocabulary user|sweep|legacy with legacy = neutral no-actor copy; emptyTrash now records the authenticated session user as actor — NULL reserved for genuinely automated paths, of which none exist today) and client-side (attribution→copy mapping with unknown/absent-field fallback, never a crash); OpenAPI + spec 065 S2/S3/S5/S11 amended [fix/sync-history-legacy] [ISSUE-201]high mitigated area:apiThe V176 change-feed allocation trigger (trg_media_items_change_seq) fired only for a fixed 15-column UPDATE OF list, making that list a silent SECOND schema-maintenance point: a future writer or migration updating ONLY unlisted media_items columns (the schema already carries two — metadata_status, preview_status) would skip allocation AND the feed silently — the row's committed change never delivered to any client (the ADR 0117 invisible-gap failure mode entering through the schema instead of the writer; oracle F8); fixed by V180 re-creating the trigger unqualified (BEFORE INSERT OR UPDATE — every update allocates; the allocator's pass-through guard keeps explicit-seq writes verbatim), after a writer inventory showed NO current path touches only unwatched columns (widening changes zero behavior today; extra no-op entries are absorbed harmlessly by the client applier's direct re-bind), with a schema-level guard test (pg_get_triggerdef must stay column-unrestricted) + behavior tests (unwatched-column update allocates and emits, carried-over seq allocates, explicit assignment passes through) [fix/sync-trigger-guard-amendments] [ISSUE-202]high resolved area:android0.10.7 launch crash: the app-scoped ExclusionsViewModel (MemoriesApp creates it eagerly at startup) had no branch in AppViewModelFactory, so every launch died in the factory'selsewith "Unknown ViewModel class" — the per-VM unit tests construct the ViewModel directly and structurally cannot catch a missing factory branch; fixed by registering the branch (excludedMedia/mediaStates from the container, PreferenceExcludedAlbumsSource(albumPreferences), TrashViewModel-style session/api/intentRecorder wiring) plus a factory class-coverage canary test that drives AppViewModelFactory itself for every registered ViewModel class [fix/exclusions-viewmodel-factory] [ISSUE-204]med resolved area:search 0.10.8 user report "filter bar unusable once a search/filter result set is empty": the search screen itself verified clean — its input and chips are composed unconditionally above the weighted content box in every state and an empty answer never clobbers the uiState feeding them (now locked by two SearchViewModel regression tests), and the filtered empty body now names the cause (search_no_results_filteredde/en/da); the report's literal mechanism — a filter row living inside the branch the empty state replaces — did exist in the Photos timeline, where the screen-levelwhenswapped the whole TimelineGrid (header + search entry + filter chips) for a full-screen MessageBox whenever a filter/scope reload emptieddays`, leaving only a Refresh that re-ran the same filter: header and filter rows are now hoisted above the state branch (chrome always composed; loading/empty/error bodies fill the content area only) [fix/search-filter-empty-state]
[ISSUE-203] low mitigated area:ci Build converter worker image failed on merge c225d67e with received unexpected HTTP status: 504 Gateway Timeout during a base-image layer pull (docker hub/registry side, ~2 min of layer retries then abort) — unrelated to the change set (converter untouched); retry semantics: the deploy workflow does NOT auto-retry failed image builds, so a transient registry 504 skips that service's deploy until a later relevant change rebuilds it; mitigation: re-run the failed job or merge a triggering commit; watch for recurrence before blaming workflow changes [PR#1376]
[ISSUE-205] med resolved area:android 0.10.9 "Nicht in Memories" listed a device image once per byte-identical copy (3 duplicate files → 3 shadowed state rows with distinct mediaStoreIds but the same contentHash → 3 identical ServerRemovedEntry rows, e.g. 18 entries for 6 images) — shadowed rows now group by (contentHash, serverLifecycleState) into ONE entry per image with a copies count rendered as "· N Kopie(n)" (de/en/da) and all grouped server ids restored together (one idempotent restore intent per server twin); null-hash rows fall back to (serverMediaItemId, lifecycle) and never merge [fix/exclusions-shadow-dedup]
[ISSUE-206] med resolved area:api 0.10.10 upload stats counted server-side dedupes as uploaded: 3×6 byte-identical photos showed "18 hochgeladen" although the commit (S4-C1 skipExistingOwnerDuplicate) created 6 items and deduped 12 (deduplication_state='skipped_existing_owner_media', storage reuse reused_stored_object) — the commit response carried no per-file verdict and the client counted every file entering the session as uploaded; fixed additively: commit response now carries files:[{fileId,deduplicationState}] (OpenAPI amended, older clients ignore the field), the Android engine adds BackupRunResult.dedupedCount (ruling: skipped_existing_owner_media = deduped, reused_stored_object/new_stored_object = uploaded), deduped files STAY marked UPLOADED locally (they are on the server; only the run statistics change) and the backup stats row appends „Duplikate übersprungen N" / "duplicates skipped N" / "dubletter sprunget over N" (its own bucket, never merged into uploaded or the reconciler's skipped) [fix/upload-dedup-count]
[ISSUE-207] med resolved area:timeline After a deep jump (timeline axis or ?date=/?y= restore) scrolling back UP never auto-loaded newer days: the top reveal sentinel renders inside the reveal observer's 900px rootMargin but ABOVE the viewport, and an IntersectionObserver fires only on threshold transitions — the single callback fired at the -900px crossing was correctly rejected by sentinelWithinRevealMargin (ISSUE-074 guard), and while the user scrolled the remaining distance to the top the sentinel stayed continuously intersecting (no new callback), so the user hit the top with only the manual "Neuere Tage anzeigen" button; secondary stranding: pendingPreviousDaysReveal set mid-jump had no consumer because every hydration-pump reveal check early-returned while jumpingDayKey was set and nothing re-checked after it cleared; fixed with two wiring changes, no new machinery: schedulePreviousDaysRevealCheck now runs from the scroll handler (tick-coalesced, strict-geometry-gated so the ISSUE-074 prepend drift loop cannot resurrect — its regression test stays green ×2), and jumpToTimelineDay's finally consumes the state once at settle with the guard lifted; new E2E test proven failing on the branch base (start index stuck at 20, 3/3 attempts), passing after the fix including --repeat-each=2 [fix/timeline-auto-reveal-newer-days]
[ISSUE-208] high resolved area:android Backup tab's primary button could stick on „Sichert …“ (disabled, presses no-ops) forever while nothing ran: BackupStatusUiState.isWorkActive treated BackupWorkState.ENQUEUED as active, and the UI snapshot (WorkManagerBackupStatusSource.observe) falls back to the PERIODIC WorkInfo when no one-off row exists - a periodic job sits ENQUEUED virtually forever waiting for its next 6h window (JobScheduler confirmed: constraints satisfied, only TIMING_DELAY unsatisfied, no immediate-work job registered), so the fallback mapped 'schedule waiting' to 'run active', disabled the button and made backupNow()/retryFailed() early-return no-ops; resolved by marking the periodic-fallback snapshot (BackupWorkSnapshot.isPeriodic) and counting only RUNNING (either source) plus one-off ENQUEUED as active, with the LastRunCard honestly showing the idle label instead of „Geplant“ for the never-ran periodic-ENQUEUED state [fix/backup-button-stuck-enqueued]
[ISSUE-210] high resolved area:api Production HikariCP connection-pool exhaustion on the upload receive path: UploadReceiveService.receive() was @Transactional while streaming the ENTIRE request body to staging (MB to GB, minutes on slow mobile links), so one pooled connection per in-flight upload was held from method entry to return (36 ProxyLeakTask leak detections/hour all pointing at DataSourceTransactionManager.doBegin ← receive ← UploadFileController.uploadFile), concurrent/retrying Android BackupWorker uploads drained the 50-connection pool, every other API request blocked on getConnection and surfaced as Traefik 502s; fixed by splitting into three phases — (1) fast transactional ownership+status checks before any attacker byte (SEC-4 order unchanged), (2) body streaming + SHA-256 with NO transaction/connection held, (3) one short transactionTemplate transaction for the file upsert + WHERE-guarded status UPDATE (atomicity unchanged, cancel still wins) with the unsupported-extension audit joined inside it and the afterCommit activity event binding to it; payload-too-large audit semantics untouched (recordUploadFailed was already REQUIRES_NEW), and two deliberate cleanup extensions: staged bytes are now deleted when phase 3 loses the cancel race or when streaming fails mid-body, where the rollback path previously left them on disk with no reaper [fix/upload-receive-connection-leak]
[ISSUE-209] med resolved area:android First map open with an empty/geolocation-less library showed the viewport error banner instead of the "no located photos" hint: the world-camera fallback (zoom 2, centre 0/0) makes osmdroid's boundingBox extrapolate BEYOND the Mercator poles on a tall phone screen (worldPx 1024 at zoom 2 < view height ~1600px → latNorth/latSouth outside ±90), and the emitted viewport was sent unclamped — server validateBounds rejects |lat|>90 (and |lon|>180 from a date-line wrap) with an error answer whose hasViewportError suppressed the empty pill; fixed by clamping every emitted viewport to the projection range (clampViewportBounds: latitudes to ±85.05112878 Mercator limit, keeping the box tile-consistent; only out-of-range longitude endpoints wrapped into the server's antimeridian form so a full-world −180..180 box does not collapse) plus belt-and-suspenders: hasLocations=false is a library fact, the empty pill now coexists with the error banner in one stacked top-centre column instead of hiding behind it [fix/map-empty-viewport-clamp]
[ISSUE-211] med mitigated area:android Server-side changes landing while the Fotos tab stays continuously open do not appear live — they surface on the next re-entry (>10 s away) or when a backup run finishes; GET /api/sync/changes is owner-scoped so a group-wide change/epoch signal (server work) would be needed for true live refresh; workaround: pull-to-refresh [PR#1404]
[ISSUE-212] med resolved area:android Backup tab's primary button locked the user out during retry backoff: after repeated upload failures (server-side 502s from the web-proxy hop during big video uploads) the one-off BackupWorker climbed WorkManager's EXPONENTIAL backoff (device-verified: attempt 9 → next run +2h08m, JobScheduler 'Minimum latency: +2h7m59s'), during which the button showed disabled „Sichert …" and BOTH gates treated the backing-off ENQUEUED request as untouchable (BackupStatusUiState.isWorkActive included one-off ENQUEUED; BackupScheduler.enqueueNow skipped on ENQUEUED/BLOCKED), so presses did nothing and the tab looked hung for hours; resolved by splitting the semantics - the button/retry gate (isRunInProgress) and the scheduler's shouldSkipImmediateEnqueue now protect only RUNNING work (a REPLACE cancels only non-running work, so a press during backoff starts a fresh attempt now and resets the backoff; the label reads „Jetzt sichern" again while waiting), plus a LINEAR 30s backoff on the one-off request and a retry cap (BackupWorker MAX_AUTO_RETRIES=5, attempt 0-based: retries on 0..4, the 6th failing run returns FAILURE carrying the last cause in the output keys) so the last-run card shows an honest „Fehlgeschlagen" with the cause instead of an endless ladder [fix/backup-retry-backoff]
[ISSUE-213] med open area:web The web-proxy upload hop returns 500/502 (TypeError: fetch failed) for multi-100MB uploads over flaky connections - the device backup retries then climb the backoff ladder (see ISSUE-212); root cause in the proxy hop not yet isolated (body streaming vs. timeout vs. connection drop), being debugged separately from today's device session [fix/backup-retry-backoff]
[ISSUE-214] med open area:api Scheduled worker_jobs retention DELETE recurringly violates discovery_results_worker_job_id_fkey (ERROR in API logs on every retention sweep): the retention job deletes worker_jobs rows that discovery_results still references (no ON DELETE action / ordering between the two tables), so the sweep never completes cleanly for those rows; needs the retention order (or FK ON DELETE CASCADE) reviewed [fix/backup-retry-backoff]
[ISSUE-215] med open area:android Timeline live updates via SSE (ISSUE-211 follow-up): the Fotos tab now consumes GET /api/events/stream (new core ServerEventStream, app-lifetime loop, 1-60 s jittered reconnect, auth-401 hard backoff) so new media and finished thumbnails appear while the grid is open — but the coverage is timeline-scope only: timeline_visibility_changed/thumbnails_ready drive the timeline's days/items, while backup-status, albums, trash, search and the delta-pull cursor still learn about server changes only via re-entry/backoff triggers (ISSUE-211's owner-scoped GET /api/sync/changes point stands for those surfaces); also the 400 ms-debounced visibility refresh re-reads the whole day list on non-preview reasons (window refresh), not the web's visible-window-only slice — acceptable at current day-list sizes, revisit if day counts grow [feat/android-timeline-live-updates]
[ISSUE-221] med open area:ci E2E Curation-Actions flake family blocks merges: discovery-domain-instance-detail.spec.ts Rename/Merge/Dissolve dialog tests (desktop AND mobile) intermittently time out at 30 s waiting for the ⋮-menu items, plus backup-export.spec.ts:5 and discovery-instance-tentative.spec.ts:2119 — verified 2026-10-29 the SAME failures hit the Daily Full Tests on clean main (run 5911: 2 failed/29 flaky) BEFORE the unrelated PR #1432 existed, then flapped that PR's CI across reruns (5913: 1 failed → green on attempt 3; 5917: 16 ✘ desktop-only → green on attempt 2), so merges currently need rerun-until-green or --no-require-ci with manual triage; pattern smells like per-runner resource contention (menu portal/rAF starvation) rather than app regressions; fix: isolate why the ⋮ menu never opens under load (trace/screencast on the runner) and stabilize or retry-quarantine the dialog tests [runs 5911/5913/5917] [audit 2026-10-30: mitigation pipeline landed — PR#1444 preserves Playwright traces on failure (upload-artifact + durable host cache), adds a nightly flake-detection job (retries=0, repeat-each=10, desktop, the six known-flaky specs), and sets --shm-size=1g on all CI Playwright containers; PR#1445 stamps html[data-app-hydrated] and gates the specs via gotoHydrated (validated 144/144 at --repeat-each=3 --retries=0 on the two flakiest specs); oracle reclassification: the ⋮ menu trigger is client-only ({#if} hand-rolled, no portal/rAF), so this timeout family is NOT the swallowed-click class — attributed to runner starvation bursts pending trace evidence from the next red run (traces now survive CI); stays open until then] [audit 2026-10-30: PR#1453 split discovery-review.spec.ts into four focused specs + discovery-review-shared.ts (describe name preserved) — split verified the rename/merge/dissolve suites green at retries=0 locally, 74/74; the ⋮-menu timeout class remains attributed to runner starvation pending trace attribution from the next red run; shm + trace evidence in place since PR#1444; stays open]
[ISSUE-222] med resolved area:people Stale pending people suggestions survived owner face-assignment cleanup: cards whose embedding justification died (faces removed from wrong persons, representatives rebuilt) stayed visible in the carousel and BLOCKED the correct re-mint via the spec-047 pending-overlap guard (William faces: stale Alice card suppressed the correct Holger card); resolved by end-of-run invalidation in persistSemisupervisedCandidates — targeted pending cards are superseded with an owner-audit row (actor NULL, semisupervised_run_invalidation, no rejectedTargetedPair suppression) when a face is live on another instance (face_reassigned_elsewhere) or the min cosine distance of all-unassigned faces to the target's current representatives exceeds suggestMaxDistance (target_representatives_changed, cluster_person_match only); prod DB manually invalidated 586 cards with the same audit convention (2026-10-29) [fix/stale-suggestion-invalidation]
[ISSUE-224] med open area:people Two accepted face occurrences carry Holger's face inside Anette's cluster (face ids 2e9ee758, ff6ea936; age-model reads child ~5–7y/male from BOTH A/B models while Anette's person median is ~59/female): legacy acceptance predates the attribute lane and predates gate demotion, so the wrong-pair faces block a correct re-suggestion through the spec-047 pending-overlap guard — same class as ISSUE-222 but with live accepted occurrences, not stale cards; needs targeted invalidation + re-mint through the methodological path (attribute-based cross-check: occurrence age-read far outside person birth-year profile → flag for owner review) once spec 068 T006 re-backfill lands MiVOLO reads on prod [spec:068]
[ISSUE-225] high open area:people Confirming a tentative face from the MEDIA VIEWER face dialog (?media=… surface) on an ambiguous cluster_person_match card 400'd ("ambiguous cluster_person_match requires an explicit owner-chosen target; use the candidate decision endpoint (never a bulk tentative confirm)") because confirmSelectedTentativeFace posted every confirm to the refs-only bulk endpoint, whose body can never carry the explicit owner-chosen target the Spec 050 guard requires (Spec 052 S3 predates Spec 050; PR#1047 fixed the person detail page but missed the media-viewer dialog); owner confirming a face on this person's dialog IS the explicit owner choice — fix routes ambiguous cards through decideClassInstanceCandidate with the current person's instance as existingInstanceId scoped to the face's ref, non-ambiguous keep the bulk confirm [PR#1465]
[ISSUE-228] high resolved area:api ALL worker result submissions 400'd since deploy 0.67.421 (2026-11-12 18:14): PR #1493 (commit 9c3d1809, WorkerResultService handler split) created the 15 WorkerResultHandler implementations as plain classes with NO Spring registration, while WorkerResultService now builds its registry from an @Autowired List — Spring injected an EMPTY list, so every result type (class_instance_candidate AND the issue fallback) was rejected "Unknown or unsupported result type"; 0 worker_job_results rows in the ~19h window vs 2613 in the prior 24h; owner_face_recognition attempts failed at submit (clustering effectively down); API stayed healthy so nothing surfaced until workers submitted; tests stayed green because the suite wires the service via the explicit-list fixture constructor, bypassing Spring collection — fixed by WorkerResultHandlerConfiguration (@Configuration, 15 @Bean methods mirroring productionHandlers(...), javadoc cross-ref) + WorkerResultHandlerConfigurationTest (15 beans, no duplicate resultType, class_instance_candidate/issue resolve); incident recovery: 3 pre-reset stale owner_face_recognition jobs cancelled (their results would 409 "stale semisupervised result" per the V151 guard — correct behavior), face_recognition_dirty marker re-inserted, fresh job f2378bad completed 2026-11-13 09:43 with the result ACCEPTED (first accepted submission since deploy) [PR#1498 / fix/worker-result-handler-beans]
[ISSUE-227] med open area:web Admin Discovery "Faces & Clustering" tab hid the ONLY control for the production clustering parameter while v3/v4/v5 methods were active: isSemisupervisedActive hardcoded ['semisupervised-v1','semisupervised-v2'].includes(...), so under semisupervised-v5 (live) the Clustering-Maximaldistanz section (endpoint PATCH /api/admin/people/clustering-max-distance, spec 050/ADR 0101) never rendered and the operator could not change clustering_max_distance from the UI at all (same parameter, same endpoint for v1–v5); additionally the method dropdown (CLUSTERING_METHOD_OPTIONS) and the ClusteringMethod type union still list only v1/v2 — selecting v3/v4/v5 in the UI remains impossible (follow-up) — fix: prefix check startsWith('semisupervised-') + copy suffix dropped in de/en/da + Playwright stored-method-v5 visibility test [fix/admin-distance-visible-v5]
[ISSUE-226] low open area:web Latent inline-dialog overflow pattern in DiscoverySettingsSection (same class as the one PR#1488 fixed in ApiKeySection): the REPROCESS ALL / DEACTIVATE confirmation dialogs are in-flow flex children inside shrink-0 action slots whose max-content row is wider than the card's md:grid-cols-3 grid column, so the expanded dialog overflows ~150px into the neighbouring profile card; currently harmless only because the neighbouring cells are empty — the moment another card lands there, its opaque background paints over the dialog (later DOM sibling wins hit-testing between non-positioned elements) and Confirm/Cancel become unclickable (deterministic, seen as Playwright intercepts pointer events timeouts on the ApiKeySection variant); fix is the same one-liner: relative z-10 on the dialog container so expanded inline dialogs paint above all non-positioned sibling cards [PR#1488]
[ISSUE-225] med open area:api Worker result submit of owner_face_attribute_backfill intermittently 409s (DataIntegrityViolationException) when a semisupervised rerun overlaps the same faces; all 13 observed cases converged operator-side; needs root-cause on the submit TX (repro: reset failed→queued during active discovery run) [spec:068]
[ISSUE-223] med resolved area:discovery A manual_selection-provenance pin whose occurrence later left the accepted state (re-clustering/release/media-delete supersede paths) resolved to NOTHING instead of falling through: ClassInstanceRepresentativeService.resolveChunk put such instances in the pinned set, the validator rejected the dead pin without re-queueing it, and the unresolved.isEmpty() early return skipped rules 2/3 — blanking the thumbnail of that person/instance across every surface (relation chips, listings, self-avatar, crop picker). Uncovered by tests because the PersonRelationServicePostgresTest pin fixture wrote a bare pointer without the manual_selection provenance, so the pin was never honored and the dead-pin branch never executed (that same missing provenance made the test's first assertion a terminal-ladder coin flip: equal confidence/recency + random-id tiebreak → the 2-of-4 flake of ISSUE-169). Fixed: rejected pins rejoin the rule-2/3 candidates (validator now returns the rejected ids); test fixtures pin with the production provenance pair and supersede occurrence+face together; added ClassInstanceRepresentativeServiceTest.deadManualPinFallsThroughInsteadOfBlankingTheThumbnail (red on unpatched resolver) [ISSUE-169 / fix/representative-thumbnail-flake]
[ISSUE-226] high resolved area:people Cardless tentative occurrences invisible in the Tentative review wall: live tentative occurrences whose backing card was decided (verified prod: owner removal decided auto_assigned_named_target_medium_band's card accepted on 25.10, later flows — mint-time same-instance absorb, overlap supersede — decided the re-allocated card; person 32cd0e03 strand) rendered on the All occurrence wall but appeared in NEITHER Tentative (wall renders only pending cluster_person_match cards) NOR Approved — violating the Spec 050 backing-card invariant (ADR 0101), the owner could see a wrong auto-assignment but never review it; prod had 22 stranded rows across 4 instances; resolved read-side: tentativeCandidates() synthesizes a reviewable card per stranded tentative (occurrence candidate_id NULL or not a pending non-face_detection card targeting this instance), synthetic id = the occurrence UUID (collision-free, decision endpoints resolve occurrence-backed ids via decideSyntheticOccurrenceCandidate: accept → in-place confirm + face mirror accepted, reject → removeOccurrences-style removal + pending-card rejection + mirror release, move → re-point + confirm; all state-guarded, owner-scoped), bulk confirm extended for refs selections containing synthesized cards; legacy allocation untouched (Part 1 guarantees surfacing); 4 regression tests (synthesize/no-duplicate/confirm-by-ref/accept-reject-move) [fix/stranded-tentative-cards]
[ISSUE-227] high resolved area:discovery Suggestions carousel permanently empty ("0 Vorschläge") despite ~1.5k deliverable pending cards (verified live 2026-11-01: paged endpoint limit=12 → 0 items while unbounded listing → 1557): the paged path fetched limit+24 raw rows per round with a hard 2-top-up cap (=108 raw rows) but the quality-ranked head of the pool was a rank-aligned DEAD BLOCK — ~115 stale targetless new_person_cluster cards (quality 0.65–0.92, minted Oct 27–31, every face ref already accepted/tentative elsewhere, none backing its own tentative) that the read-time stale filter (rejectUnactionableCandidates) removes AFTER the raw fetch — so page 1 ended empty as soon as the dead block outlasted the row budget (first actionable card at rank ~116); the dead block kept growing because end-of-run invalidation (invalidateStalePendingSuggestions) covered only TARGETED cards; fixed both ends: (1) candidatesPaged top-up now continues while the page is short AND the last raw fetch was a full window (structural termination: finite pool, offset+rawConsumed strictly advances; ISSUE-108 bounded-batch discipline untouched — statements stay ⌈pool/(limit+24)⌉ window fetches + chunked IN probes), regression carouselTopsUpPastDeadBlockOfHighQualityStaleCards (115 dead cards above one fresh card) red on unpatched; (2) new stale class D all_refs_assigned_elsewhere: targetless cards whose every canonical face ref holds a live occurrence and that back no own tentative are superseded at end of run with the standard audit convention (actor NULL, semisupervised_run_invalidation, NO rejectedTargetedPair — never feeds future-suggestion suppression), implemented set-based in the PG arm (third UNION ALL branch, branch-exclusive of A/C, empty-refs guard keeps media-only cards actionable) and in-memory in the H2 arm; 3 new Class D tests + full api suite 2,516 tests green [fix/suggestions-carousel-empty-page]
[ISSUE-228] med open area:workers WorkerJobHistoryPurgeService scheduled tick fails every 10 min with DataIntegrityViolationException (FK discovery_results_worker_job_id_fkey on DELETE FROM worker_jobs): the purge deletes children first (worker_job_results, worker_artifacts, worker_media_access_grants, processing_issues) but NOT discovery_results / class_instance_occurrence_candidates rows referencing the job, so any terminal job still referenced by a discovery result is undeletable and its whole purge batch fails forever; visible in api logs ≥2026-10-28 (60+ occurrences/4 days); impact: worker_jobs retention bloats, no suggestion-path impact observed; fix: exclude jobs with referencing discovery_results/class_instance_occurrence_candidates rows from the batch (or null/delete the references per retention policy) [ref:WorkerJobHistoryPurgeService]
[ISSUE-230] med open area:ci Forgejo runner action-cache directory ownership breaks any job whose workflow defines an upload-artifact step: the runner daemons execute as user git while memories-deploy-runner-*/workspace/db/ (act's action clone cache) is owned by chris mode 755, so git clone actions/upload-artifact dies with "could not create leading directories" ~200ms into the job — BEFORE any workflow step runs (PR #1477 'Run scoped PR tests' failed twice at 21:39/21:45 on 2026-11-01, zero test output); job containers run as root and are unaffected (which is why changes and non-artifact jobs stay green); mitigated 2026-11-01 with chmod -R a+rwX on both runner-2 workspace/db trees (rerun went green immediately) — but the ownership drift will reappear the next time a job runs as chris creates cache dirs; durable fix: run both runner daemons with the same uid that owns the workspace tree, or an init step that re-applies group-write on workspace/db [runs 6134 attempt 1+2 / PR#1477]
[ISSUE-231] med mitigated area:web-tests gotoHydrated hydration gate dies terminally when the browser fails to fetch the app's own static chunks on the shared CI runner: Playwright trace (run 6257 flake-detection, albums compact-preview failures) shows 8 × ERR_NETWORK_CHANGED on _app/immutable/chunks/*.js + nodes/1.*.js, then TypeError: Failed to fetch dynamically imported module — SvelteKit never boots, the html[data-app-hydrated] marker never appears, and the gate waits out its 60s on all 5 known-flaky failures + the 3 flaky rename-dialog tests (full suite); recovery = ONE page.reload() inside gotoHydrated (fresh navigation re-fetches every chunk; deterministic failures fail again on attempt 2, 2×40s sized inside the 90s test timeout); unrelated but same-run evidence: timeline-scroll-stability.spec.ts:125 "Received: 14" was deterministic mobile-only across 7 consecutive nightlies — the axis jump reused the pre-jump window size (mobile's wheel reveal grew it to 16) so the tail clamp started the window at day 14; fixed separately with a fixed-size jump window [run 6257 / fix/e2e-flaky-playwright]
[ISSUE-232] med resolved area:timeline timeline-scroll-stability.spec.ts:125 (mobile) failed deterministically — "Expected: 20, Received: 14" in EVERY attempt of 7 consecutive nightly full-test runs (6175→6257) while desktop stayed green: setVisibleWindowAroundDay sized the jump window with Math.max(visibleDayCount, 10), but on mobile the wheel gesture that reveals the axis crosses the bottom reveal sentinel and grows visibleDayCount to 16 before the click, so the tail clamp Math.max(0, days.length - windowSize) set the start index to 14 instead of 20 and the pinned-day assertion timed out (trace: day 8 items request still in flight at click time — the watch-lane had hydrated days 0–7 + 8 in flight); resolved by giving the jump a FIXED-size render window (10 days, target near its head — the pre-jump count carries no information the jump needs; the hydration pump fills the rest); A/B-proven: unpatched 3/3 red with the exact "14" signature, patched 24/24 green at --repeat-each=3 --retries=0 [run 6257 / fix/e2e-flaky-playwright]
[ISSUE-233] med resolved area:android Timeline filter chips overflowed on narrow screens: FilterRows laid all 5-6 Material3 FilterChips (Alle/Meine/[Familiegruppe]/Favoritter/Fotos/Videoer) into ONE non-scrollable fillMaxWidth Row, so Compose squeezed the trailing chips until their labels wrapped letter-under-letter — "Videoer" rendered as a vertical V/i/d/e/o/e/r stack; same latent overflow class in MapScopeRow (map scope chips + satellite chip squeezed by the trailing search/cast buttons); resolved to match the web timeline's two-row chip layout — FilterRows now renders scope chips (Alle/Meine/Familiegruppe) and filter chips (Favoritter + Fotos/Videoer) as two separate horizontally scrollable Rows (horizontalScroll per row in a Column wrapper, group-chip visibility + favorite toggle + media-type tap-again-clears semantics untouched), MapScopeRow scrolls its chip segment inside weight(1f, fill = false) keeping search/cast pinned trailing; all touched chips gain heightIn(min = 48.dp) + maxLines = 1 (Material3 default chip is 32dp tall, below the 48dp touch target; scrolling makes truncation unnecessary so labels keep intrinsic width); SearchScreen filter row was already a LazyRow and BackupSettings interval chips already horizontalScroll (touched only for the 48dp target) [fix/android-filter-chips]
[ISSUE-234] high resolved area:upload Android upload of a non-Latin-1 filename crashed the WHOLE batch twice over: (1) UploadSessionClient sent filename/client-relative-path RAW in HTTP headers, so OkHttp's header-value check threw IllegalArgumentException ("Unexpected char 0xf8 at 1 in filename value") on "Fødselsdagsinvitation med sportsmotiv3.png" before any bytes were sent; (2) that exception (and any per-item local IO fault) escaped BackupEngine's batch loop, so the session was never committed and the files already RECEIVED (404 items) were stranded unprocessed. Fixed client-side only, server verified already-safe: (a) fingerprint headers percent-encoded UTF-8 at the header boundary (encodeHeaderValue, encodeURIComponent semantics incl. space→%20), mirroring apps/web upload.ts — FileUpload fields stay RAW, encoding is wire-only; (b) batch loop catches per-item local faults (request-build/IO/IllegalArgumentException), marks the item FAILED (retryable next run), continues, and still commits with the received set; session-level faults (401, 403/404/409/410 session-gone, 5xx/network after retry budget) keep abort semantics; CancellationException never swallowed. Server commit-independence verified rather than changed: the finalizing commit (fileIds==null) processes ALL received/classified/commit_pending files (UploadCommitService.eligibleFiles) and commitFile per-file failures abort via transaction rollback (staging kept, idempotent replay) — receive is transactionless (ISSUE-210 shape intact), staging cleanup on the failure races unchanged, SEC-4 ownership checks untouched; a session that NEVER commits is the owner's discard/restore decision under spec 070 (stale), so no server change was forced. Tests: UploadSessionClientTest percent-encoding contract (Danish filename, ASCII client-file-key passthrough, DTO keeps raw name) + BackupEngineTest per-item isolation (local read failure isolates + commits; failed item retries in a fresh session; progress events fire for isolated items) [fix/android-upload-filename-peritem]
[ISSUE-235] med resolved area:timeline Android timeline stayed PENDING (spinner tiles) while this device uploaded: three stacked gaps — (1) the app-scoped SSE stream used a plain OkHttpClient() whose default 10 s read timeout tore the server's heartbeat-less event stream down every quiet 10 s (emitter only closes at 30 min; no keepalives), so during each 1–60 s backoff gap the thumbnails_ready/timeline_visibility_changed events fired by the worker's preview completion (WorkerResultService → publishThumbnailsReadyAfterCommit; emission itself verified healthy) never reached the client; (2) ThumbnailCell renders NO image request while previewStatus == pending, so the per-cell 6 s×10 retry loop never ran for pending tiles and a missed event pinned the spinner until manual refresh; (3) TimelineViewModel refreshed only on the active→terminal BackupWorkSnapshot transition, so new items appeared only after the whole run ended; resolved client-side (server emission left untouched — payload/back-compat intact for web): ServerEventStream.streamingClient() disables read+call timeouts (connect/write stay bounded; SseTransportTimeoutTest pins 0/0/15s/15s and guards the 10 s default), BackupEngine gains an additive outcome-only onBatchCommitted hook (fired once per successful batch commit inside the NonCancellable epilogue; fired for crash-replayed commits; never on abort/empty run) threaded through AndroidSyncController into a container-scoped MutableSharedFlow that the factory passes as TimelineViewModel.backupCommitEvents — each pulse sets windowRefreshRequested through the EXISTING scheduleTargetedRefresh worker (400 ms debounce, no second refresh path), so items appear mid-run without the terminal snapshot; ThumbnailCell arms ThumbnailRetryScheduler.startPendingProbes (60 s × 10, negative attempt numbers, separate budget from the 6 s failed-load lane, cancelled on status change/dispose) so an armed probe fetches the thumbnail even while status still reads pending [fix/android-timeline-thumb-refresh]
[ISSUE-229] high resolved area:api GET /api/media/{id}/original (spec 059 original-byte streaming: video-viewer playback fallback + viewer original download) answered 500 on EVERY request in production (BadSqlGrammarException "column oa.storage_key does not exist"): findRow selected oa.storage_key/oa.mime_type from original_assets, which has NEITHER column in the production schema (V001) — the key + mime type of an original live on stored_objects via original_assets.stored_object_id (original_assets carries only media_type photo|video); same drift as archived ISSUE-171 (public-share ZIP, 2026-10-18) whose stored_objects-join fix never reached the newer endpoint because the H2 media-original/media-token test schemas carried the SAME drifted columns (original_assets WITH storage_key/mime_type, no stored_objects table) so the 500 was invisible to tests from ship (2026-10-24) until a live-DB check during the selection-download feature work — resolved by joining stored_objects (LEFT JOIN so.id = oa.stored_object_id, select so.storage_key/so.mime_type; LEFT so a dangling original_assets row degrades to the existing 404/missing-file semantics instead of dropping the row) and aligning both H2 schemas + seed inserts to the production shape [PR#1476, feat/timeline-selection-download]
[ISSUE-237] high resolved area:timeline Rotating the screen while a photo was open in the full-bleed viewer wiped the item lists and showed the viewer empty state (R.string.viewer_empty, "Der er ikke noget at vise lige nu") instead of the photo: MainActivity has no configChanges, so rotation recreates the Activity; the six app-scoped shell ViewModels (search/albumDetail/personDetail/map/timeline/addToAlbum, hoisted in MemoriesApp.kt since ISSUE-121) SURVIVE recreation under the Activity's ViewModelStore, but each one's LaunchedEffect(session) { vm.reset() } re-launches in the fresh composition because authRepository.session is a StateFlow (app-scoped in AppContainer) that replays its current value — so every rotation ran all six resets, and TimelineViewModel.reset() emptied _uiState to TimelineUiState() just as the viewer destinations rendered items = timelineState.items, tripping MediaViewerScreen's empty-list guard (the ISSUE-121 belt-and-braces check; the reset re-run also contradicted reset()'s documented intent, which is to forget the PREVIOUS login, not the current one); fixed by routing all six sites through a ResetOnSessionChange(session) wrapper: a plain (deliberately NOT rememberSaveable — the anchor must NOT survive recreation) remembered last-session state seeded from the current session makes the reset fire only on a GENUINE session identity change (StoredSession is a data class, structural inequality is exact, login always mints a new token), so recreation re-seeds quietly while null→s1 login / s1→null logout / s1→s2 re-login still reset; process death also re-seeds (VMs are fresh anyway); the separate session-navigation LaunchedEffect is untouched (harmless on recreation: currentRoute is restored, not LOGIN), timeline refresh-on-re-entry keeps coming from ensureLoaded()'s own reentry throttle, and the viewer's empty-list guard stays as the ISSUE-121 safety net [PR#1538] [fix/android-viewer-rotation-empty]
Archived Issues
(2026-11-15) [ISSUE-236] high resolved area:android Android viewer video playback failed on Samsung devices — the viewer played the poster preview (previewUrl) or fell through to /original?purpose=view (original HEVC/HEIF, not decodable by ExoPlayer's default path) instead of the 720p H.264 video_rendition (spec 059); videoRenditionUrl was already mapped into TimelineItemUi and consumed by the cast path, but videoPlaybackUrl never preferred it, so local viewing errored ("Video konnte nicht geladen werden") while cast played — fixed by preferring the rendition in videoPlaybackUrl (rendition → preview poster → original; poster kept only defensively, original remains the old-server fallback) [fix/android-video-playback]
(2026-10-29) [ISSUE-222] med resolved area:people Legacy people candidates carried occurrence_refs {"mediaItemId":X} with no face id (discovery_people worker pre-faceAnnotationId shape, sporadic through 2026-10-26), so accepted occurrences degraded to occurrence_type='generic_evidence' keyed <mediaItemId>:<index> with no scope — the person-page day-items join (discovery_results.id = occurrence_key) never matched and ~68 live photos (e.g. Holger, occurrence 92af4e56) rendered with no face box and no "Vis område" zoom; resolved on both layers: the occurrence write path now binds media-only refs to the exact face annotation via payload-bbox identity (resolveExactFaceAnnotationId, unambiguous match only, ADR 0038 no-steal guards, generic_evidence fallback unchanged on any miss) and an hourly backfill (FallbackOccurrenceFaceRebindScheduler) repairs existing rows in place with preserved provenance + distinct fallback_occurrence_face_rebind action (candidates untouched, idempotent, ambiguous/decisively-assigned rows skipped) [PR pending] [fix/people-fallback-occurrence-scope]
(2026-10-29) [ISSUE-220] med resolved area:ci Container-registry blob uploads cut mid-stream (unexpected EOF in Forgejo UploadBlob → docker push "Client Closed Request"): converter worker's ~2.5 GB pip-layer upload failed on PR#1430 (7 CI attempts across two days, "Layer already exists" for every other layer, manifest PUT never reached, "Deploy converter worker" skipped). Two stacked causes: (1) Traefik entrypoints used the default respondingTimeouts.readTimeout=60s — every upload >60s died at exactly 60s (visible since 26.10); (2) after raising to 300s (2026-10-29, operator gateway config backup kept), the ~8 MB/s runner→registry path needed >300s for the 2.5 GB blob — its only symptom is Forgejo logging a 500 after exactly 300s, so very large layers may still need a higher value. Resolved for the converter image: job rerun succeeded post-raise (run #5901 all green incl. Deploy converter worker); orphaned partial uploads (21 GB) pruned from the Forgejo container's package-upload tmp dir. Watch for the same 500-after-exact-timeout signature on larger future images.
(2026-10-28) [ISSUE-217] low resolved area:ml owner_face_embedding_backfill jobs enqueued before the model-pair payload keys existed (pre-landmark deploys) lack "modelName"/"modelVersion" in their JSON, so the active-job anti-join's pair discriminator does not match them and a face covered by such a stale in-flight job can be re-enqueued for the same default pair once; benign double work only — worker-side embedding dedup (UNIQUE occurrence+model pair) collapses the duplicate result, no wrong data [feat/face-alignment-api]
(2026-10-27) [ISSUE-219] high resolved area:ml owner_face_embedding_backfill job payloads carried boundingBox (and potentially landmarks) double-encoded as JSON strings since the landmark-payload change (PR#1419), so every face failed with "string indices must be integers" and a full 4524-job backfill run completed with 0 embeddings written — no data corruption (embedding tables untouched, run is resumable); resolved on both sides: API packs a parsed JSONB object into the payload (jsonObjectFromDatabaseJson with {} fallback), worker _coerce_json tolerates string-encoded boxes/landmarks from stale queued jobs [PR#1420] [fix/backfill-box-doubleencode]
(2026-10-27) [ISSUE-218] low resolved area:ml owner_face_embedding_backfill results are stamped with the worker's ENV model pair, not the job payload's target pair, so faces enqueued under a target pair the worker does not emit (e.g. default arcface-r50/w600k-v1 enqueue while the worker env is w600k-v2-aligned) never reach target coverage and re-enqueue on every run of that pair — duplicate compute only (UNIQUE occurrence+model-pair dedup absorbs the rows, no wrong data); resolved: the backfill handler reads modelName/modelVersion from the job payload and resolves provenance from it — matching pairs compute normally, mismatched pairs short-circuit with a model_pair_mismatch issue (no embeddings computed, no relabeling), pair-less stale payloads keep the env fallback [PR#1418, PR#1419] [feat/face-backfill-payload-pair]
(2026-10-24) [ISSUE-007] low resolved area:explore flaky test 'loads People cards incrementally' — race condition in pagination mock data (nextCursor mismatch between endpoints). Passes ~50%. [audit 2026-10-24: resolved 2026-10-24: referenced test removed from explore-people.spec.ts, flake can no longer occur]
(2026-10-24) [ISSUE-009] med resolved area:ci test-retry plugin ASM Unsupported class file major version 65 — post-merge push CI fails while PR CI passes. Update test-retry to ASM 9.7+. [PR#682 run#1463, issue#685] [audit 2026-10-24: resolved 2026-10-24: test-retry pinned to 1.6.1 (ASM 9.7.1) since commit 7391ee4a]
(2026-10-24) [ISSUE-018] med resolved area:discovery V120 occurrence_refs reads mediaItemId but people-worker refs carry only faceAnnotationId → NULL → candidate queries silently empty → retrospective auto-accept skips new-person clusters (ENG-1). [audit:2026-09-04] [audit 2026-10-24: resolved 2026-10-24: superseded — matcher rework (ISSUE-043 fix) resolves faces from payload bbox; V120 residual tracked under ISSUE-043/053]
(2026-10-24) [ISSUE-020] med resolved area:workers No heartbeat on long jobs + 5-min reaper → lease expires → job requeued while first worker runs → duplicate discovery results, duplicate artifacts (ENG-3). [audit:2026-09-04] [audit 2026-10-24: resolved 2026-10-24: heartbeats landed (WorkerController:68, worker pings extend leased_until); reaper honours lease]
(2026-10-24) [ISSUE-022] med resolved area:db V108 passes legacy 'active' state through → violates V107 discovery_results_state_check → migration fails on affected installs (ENG-5). [audit:2026-09-04] [audit 2026-10-24: resolved 2026-10-24: not applicable to the only install — Flyway ran clean through V179 on the owner DB]
(2026-10-24) [ISSUE-024] med resolved area:workers ML worker advertised dormant face_detection/owner_face_recognition capabilities + face model env while Core credential excluded face job types — dead compose misconfig, removed by this change [PR#702]
(2026-10-24) [ISSUE-025] high resolved area:auth No password-reset path existed (no CLI, no admin endpoint; resetPasswordAsSuperAdmin dead code) — locked-out operator recovery required manual unaudited DB UPDATE; delivered: audited CLI reset (any user by email, session invalidation) [PR#704]
(2026-10-24) [ISSUE-026] med resolved area:ci detect-scopes.sh evaluates only push-head commit on PR-branch pushes — multi-commit PRs with docs-only head commits skip scoped tests (PR #704 API changes untested in CI); fix: detect on merge-base diff [PR#704 run#2378] [audit 2026-10-24: resolved 2026-10-24: detect-scopes diffs merge-base..HEAD since 71f266fc]
(2026-10-24) [ISSUE-029] med resolved area:people Suggested-people thumbnails all 404 — URL builders emitted face-crop URLs without checking the face_thumbnail derived asset exists (production: 0/16411 faces have thumbnail_derived_asset_id; legacy generation pipeline dead); URL now only emitted when asset resolves [fix/face-suggestion-thumbnail-404]
(2026-10-24) [ISSUE-031] high resolved area:people PR#718 recovery deferred: pre-existing fallback-keyed accepted occurrences ("mediaItemId:index" / generic_evidence) and occurrence-less person instances (349 on owner DB) are NOT repaired by the removal of the unsound heuristic rekey; repair endpoint DELIVERED in PR#719 (POST /api/admin/discovery/people/recover-faces, exact-ID insert-only, dry-run) — stays OPEN until the ops recovery run completes on the owner DB [PR#718, PR#719] [audit 2026-10-24: resolved 2026-10-24: ops recovery run completed on owner DB — occurrence-less active instances 349→4 (the 4 were created 2026-10-10/11, after the run), numeric fallback-keyed accepted occurrences 0; 9 accepted rows still carry uuid:index keys, data-side residual only]
(2026-10-24) [ISSUE-038] med resolved area:people Completion-time watermark could permanently starve embeddings created mid-job and terminal 'failed' jobs re-enqueued every 5 min without stamping; claim-time watermark (worker_jobs.started_at) rewinds the cursor past heartbeat inflation and terminal failures now stamp the same upsert [PR#724, PR#725]
(2026-10-24) [ISSUE-040] low resolved area:people workers/people test suite has 8 pre-existing failures in test_face_model.py (outdated mocks vs current flat-output FPN decoding in detect); also fails on clean main [PR#721 review] [audit 2026-10-24: resolved 2026-10-24: mocks rewritten for flat-2D SCRFD layout; test_face_model.py verified 25/25 green on main]
(2026-10-24) [ISSUE-043] high resolved area:people discovery_people worker candidates persist media-only occurrence_refs ([{"mediaItemId":X}]) with no faceAnnotationId, so retrospective "Apply" auto-approval matched 0 candidates (matcher bailed on missing face key + pool cross-joined discovery_results per media); matcher now resolves the exact face from the candidate payload bbox via ingest-identical IoU/center predicate, anchors restricted to named instances + accepted occurrences + matching dimensions, assigned faces never moved [PR pending] [audit 2026-10-24: resolved 2026-10-24: exact-face IoU/center matcher landed (ClassInstanceDiscoveryService:9600-9669); "[PR pending]" was stale]
(2026-10-24) [ISSUE-044] high resolved area:people remove/dissolve dialog on the instance detail page called archive instead of dissolve, archiving instances bare (no occurrence supersede, no release strategy) and leaving 5,748 faces stuck on archived instances (zombies); dialog now calls dissolve, Archive button hidden for people-domain rows, and POST /api/admin/discovery/people/release-archived-occurrences (dryRun, idempotent, audited) repairs existing zombies — stays open until the ops release run completes on the owner DB [PR#736] [audit 2026-10-24: ops release run verified complete — 0 accepted occurrences on archived instances]
(2026-10-24) [ISSUE-045] high resolved area:discovery pending class-instance candidates were never superseded when their target instance got archived or when newer owner_face_recognition jobs re-proposed the same faces — 58k stale pending rows pointed at one archived unnamed person and surfaced as nonsense "add to existing person" cards; fixed here via supersede-on-archive/merge, supersede-on-repropose, cleanup migration, read-path guard [diag 2026-09-16] [audit 2026-10-24: resolved 2026-10-24: supersede-on-archive/merge + V123 cleanup migration + mint-time guards landed]
(2026-10-24) [ISSUE-046] high resolved area:discovery persistWorkerCandidates embedding-match path bound worker face ids as VARCHAR against uuid columns (pending_face_embeddings.face_annotation_id) → PostgreSQL 500 "operator does not exist: uuid = character varying" on results POST → owner_face_recognition jobs failed 3x and cursor stamp blocked re-recognition (no suggestions surfaced); fixed via typed UUID binds + PG regression test (H2 coerces the cast so H2 tests stayed green) [diag 2026-09-17] [audit 2026-10-24: resolved 2026-10-24: typed UUID binds landed (:9626-9645) with PG-vs-H2 rationale + regression test]
(2026-10-24) [ISSUE-047] high resolved area:people Recognition candidates' faces were transitioned owner_tentative→owner_recognized on persist ("considered") but rejection never undid it — rejected clusters' faces left the recognition feed forever (owner 38a79de6: 6,553 stranded faces of job 1902cae5's rejected clusters; both owners' next runs deterministically submitted 0 candidates, DBSCAN replication proved 0 cores); rejection (decideCandidate/rejectCandidateByPolicy) now releases unassigned face_annotation refs back to owner_tentative + V125 repairs stranded faces and resets the cursor [diag 2026-09-18, PR pending] [audit 2026-10-24: resolved 2026-10-24: release-on-reject landed (PeopleDomainCandidateDecisionStrategy:300-312); V125 applied on owner DB (flyway 179); "[PR pending]" was stale]
(2026-10-24) [ISSUE-050] high resolved area:people PR#763 accidentally moved the people-worker job dispatch block OUT of the while RUNNING polling loop (4-space vs 8-space indent) — since that deploy the worker claimed a job, lazily loaded models, then looped forever on empty claim polls without ever dispatching: every claimed job (discovery_people, owner_face_recognition, owner_face_embedding_backfill) sat 'running' until lease expiry and was requeued by the recovery service in a never-processing loop (canary backfill job ebc57437: claimed 01:18:58, zero handler logs, lease expired 01:28:58, two orphaned attempts, completed only on attempt 3 under the fixed worker); fixed by restoring the pre-#763 indentation (#764) — caught by the staged canary before bulk-enqueueing 3,037 more jobs [PR#764]
(2026-10-24) [ISSUE-051] high resolved area:people FaceRecognitionDirtyMarkService.markDirty used an ANSI MERGE with an aliased target in the WHEN MATCHED SET clause (d.marked_at) — H2 (CI suite) accepts it, PostgreSQL rejects it ("column d of relation face_recognition_dirty does not exist"), so since PR#760/#761 every person mutation that marks dirty 500-ed on prod (remove-from-person, candidate accept, merge/rename/dissolve, occurrence confirm/move; also failed 13 owner_face_recognition jobs via the persistWorkerCandidates auto-accept→markDirty rollback) and face_recognition_dirty stayed empty → reconcile scheduler never re-armed recognition → 0 suggestions; also not portable in reverse: single-statement INSERT … ON CONFLICT DO UPDATE is rejected by the suite's H2; fixed with two portable statements (UPDATE, then INSERT when nothing changed) serialized by the caller's edit transaction (#768) — verified live: reconcile enqueued recognition for both owners, 71 fresh candidates within minutes, zero BadSqlGrammar [PR#760, PR#768]
(2026-10-24) [ISSUE-054] low resolved area:discovery Tentative candidate listing is unpaginated (GET returns all pending candidates for the instance); V131 target index supports a future keyset query; add limit/cursor when real libraries hit large pending sets [spec:041 review] [audit 2026-10-24: resolved 2026-10-24: limit/offset/totalRemaining landed (DiscoveryController:440-451)]
(2026-10-24) [ISSUE-055] high resolved area:people "New person instead" on a cluster_person_match suggestion ignored the entered displayName and silently accepted into the SUGGESTED person: decideCandidate fell through to the preLinkedTarget fallback (existing_instance_id on the candidate) because the frontend posts accepted+displayName without existingInstanceId and nothing read the createdAsNew marker; fixed via explicit createNewInstance Boolean on the decision request (legacy compat: details.createdAsNew=true) that skips the pre-linked fallback so the new-instance resolver applies displayName and creates a fresh instance carrying the accepted occurrences; guard: signal + blank name → 400, candidate stays pending; suggested person and stash-shadow path unchanged [fix/new-person-instead-persistence]
(2026-10-24) [ISSUE-056] high resolved area:people Profile people-settings form called the generic class-instance-auto-accept-policy endpoint with people-specific field names (confidenceThreshold/sharedPersonAutoApprovalEnabled/sharedPersonConfidenceThreshold) instead of minConfidence/sharedInstance* → enabling auto-approve failed with "classInstanceAutoAccept.minConfidence is required when enabled" and local/shared threshold edits were silently dropped (backend COALESCE kept old values) and reverted to 0.85/0.9 defaults on reload; fixed web-side contract alignment + normalization of both response shapes (generic endpoint vs retrospective-run policy); e2e mocks that encoded the wrong contract corrected [PR#800]
(2026-10-24) [ISSUE-057] high resolved area:people Media-viewer face dialog typed-name assignment ("Find or create person") returned 409 "Face is already assigned to a named person" whenever the face merely had a PENDING "This photo looks like X" suggestion: currentAssignmentForFace fell back to pending candidates with existing_instance_id, so the accepted-assignment guard fired on unaccepted suggestions; fixed by making currentAssignmentForFace accepted-only, creating the owner-named person and superseding the stale named-target suggestions (state='superseded', leaves the carousel) when the typed name differs, keeping the unnamed-cluster name-the-candidate merge path, excluding rejected faces from facesForMedia, and mapping a not-assignable face to 400 (was 409/500-ish); web now shows ProblemDetail text via describeApiError (+ +json response parsing) and resyncs faces/suggestions on 409 [fix/face-naming-409]
(2026-10-24) [ISSUE-070] med resolved area:ci memories-deploy runner-2 workspace broken for JS actions: android job on main merge (run 2900, job 36728) died before executing any step — git clone https://data.forgejo.org/actions/upload-artifact failed with "could not create leading directories of /home/chris/git/memories-deploy-runner-2/workspace/db/" (exit 128), i.e. host filesystem state prevents action checkout; PR runs on the sibling runner were green; fix is host-side runner workspace cleanup/inspection, not workflow changes [PR#877, run 2900] [audit 2026-10-24: resolved 2026-10-24: host workspace healthy again — db/ hash dirs create fine, runs green since]
(2026-10-24) [ISSUE-072] med resolved area:web-tests upload-status.spec.ts "activity badge stays hidden for failed-only lifetime work" races Svelte hydration: the click lands on the SSR-rendered "Open recent activity" button before handlers attach (~80% failure at retries=0, both on main and feature branches; CI retries masked it as flaky); fix by awaiting hydration/summary request before the click as sibling tests do [spec:049, PR#925 attribution] [audit 2026-10-24: resolved 2026-10-24: hydration guard landed (upload-status.spec.ts:550-554)]
(2026-10-24) [ISSUE-081] high resolved area:people Constrained forward clustering (ADR 0100) built its density graph over anchors + all unassigned faces as one full N x N float64 similarity matrix plus two full-size temporaries (~20.7 GiB simultaneous allocations at n=26k, dim=512) and segfaulted OpenBLAS mid-GEMM on the 16 GB deploy host — every owner_face_recognition job died hard after fetch (kernel: segfault in libscipy_openblas64, worker RestartCount 13), failed after 3 attempts, and submitted nothing; resolved: blockwise float32 graph build with ~192 MiB tile, vectorized label-broadcast cannot-link, in-place mask handling (measured n=26213: 12 s, 2.4 GiB peak, adjacency byte-identical to brute-force reference); legacy _cosine_distance_matrix got the same treatment; regression pins for equivalence + household-scale memory bound [PR#975]
(2026-10-24) [ISSUE-082] high resolved area:people Owner-face-embeddings feed dropped confirmed anchors: auto-accepted candidates move discovery_results.state to 'accepted' (afterAcceptance), but the accepted-occurrence branch filtered owner_tentative/owner_recognized only, so all anchor embeddings vanished from the worker feed — anchor_labels resolved empty, the constrained pass found zero identities, and 100% of step-2 candidates degraded to new_person_cluster (no existing_person_match / merge suggestions ever emitted; carousel showed only new people); resolved: accept dr.state='accepted' in the accepted-occurrence branch (still gated on accepted occurrence + stored occurrence embedding); regression test pins anchor+tentative in one feed response; verified live: next run emitted 24 existing_person_match + 250 cluster_person_match, 0 new_person_cluster [PR#976]
(2026-10-24) [ISSUE-084] high resolved area:people Owner-face-embeddings feed excluded post-reset faces: soft "Personen-Clustering zurücksetzen" removes occurrences but deliberately keeps faces at dr.state='owner_recognized' ("the face stays in the recognition pool") and re-arms their embeddings as pending_face_embeddings rows, yet the feed's pending branch filtered dr.state='owner_tentative' only — 17,970 orphaned faces of the affected owner were invisible to every recognition run, the worker clustered an 8,681-face residual pool with ZERO density-graph triangles at maxDistance 0.35 (complete-linkage cliques ≥3 impossible for ANY method: anchored-hybrid, constrained-dbscan and legacy DBSCAN all emitted 0 candidates; with the full pool the new anchored-hybrid method emits 5,832 new_person_cluster, verified by in-container repro); resolved: pending branch admits dr.state IN ('owner_tentative','owner_recognized'), accepted-preference dedup untouched; 3 regression tests incl. failing-on-main proof [PR#987]
(2026-10-24) [ISSUE-085] low resolved area:people Clustering-method switch was invisible in logs: worker logged no line naming the dispatched method (only a WARNING for unknown names), so an anchored-hybrid switch looked like a no-op during the PR#987 zero-candidate incident, and the operator job label even claimed "Face recognition (DBSCAN clustering)" (hardcoded); resolved: one INFO line per run "Clustering run: job=… method=… pool=N anchors=M tentative_assigned=K" before dispatch + method-agnostic operator label "Face recognition (clustering)" [PR#988, PR#989]
(2026-10-24) [ISSUE-077] high resolved area:discovery UserDiscoveryService reprocess-all was broken at scale: enqueueAllItems batch SELECT had no cursor/exclusion so it re-selected the same first batch forever (infinite loop inside one @Transactional method), per-row enqueueDiscoveryDomainJob calls all joined the request transaction → 21k-media reprocess produced a zombie PostgreSQL "idle in transaction" backend holding invisible inserts after the 120s proxy 524 (workers saw zero jobs, UI frozen at "0 of N processed"); incremental DiscoveryReprocessScheduler cannot backfill because it only enqueues media created after its watermark (created_at > reprocess_last_enqueued_at), so pre-existing libraries had no working full-rebuild path; fix: cursor pagination (mi.id > :lastId), REQUIRES_NEW chunk commits (~250), async executor + per-owner/domain guard, enqueueState/enqueuedCount in domain status; caveat: re-running after a mid-run crash may re-enqueue some media (duplicate jobs are benign — same payload, idempotent worker results) [fix/reprocess-all-batching] [audit 2026-10-24: resolved 2026-10-24: cursor pagination (mi.id > :lastId) + REQUIRES_NEW chunks + enqueueState/enqueuedCount landed]
(2026-10-24) [ISSUE-087] high resolved area:people Multi-pass recognition pipeline floods persons with unverifiable assignments: step-1 nearest-anchor emission auto-confirms wrong faces when Core strong band + auto-accept policy pass (34 faces dist ≤ 0.347, Holger chain), anchored-hybrid Phase-B cliques never consume their faces (persons not auto-created) so every run re-clusters the same pool into overlapping ~25-face cliques (William 28d20e79: 894 silent tentative occurrences from 134 clusters in 3 runs via the embedding-match medium band, no cards); implemented: semisupervised-v1 single-pass A/B/C pipeline landed (ADR 0101, spec 050) — no step-1 emit, no weak/legacy passes, no Core rematch, tentative always with visible backing card, ambiguity always review-only; mitigation active once method is switched to semisupervised-v1 (deploy API+people-worker together first, then switch + recluster) [spec:050] [audit 2026-10-24: resolved 2026-10-24: semisupervised pipeline live on this instance (ADR 0101; v3/v4 runs 2026-10-16/19); step-1 flood mechanics gone]
(2026-10-24) [ISSUE-090] med resolved area:people Stale people suggestions whose face refs are ALL occupied now resolve on "Stash away" without moving/removing ownership or creating an empty shadow; audit preserves client details with authoritative owner_action/stashed/all_faces_already_assigned fields; own-allocation and removed-occurrence cards no longer 409 (insert-only zero-write stash resolves as rejected/outcome all_faces_already_assigned, PR#1005); remaining 409s are genuine races only, surfaced as ProblemDetail code candidate_changed; automated paths unchanged [PR#1002, PR#1005]
(2026-10-24) [ISSUE-096] high resolved area:media:dimensions HEIC imports leave media_items.width/height NULL (metadata-extractor HEIF gap + swallowed extraction failures marked metadata_status 'complete') so day-items emit dimensions:null, which disables the face-zoom gate on affected photos; worker now resolves HEIC dims via ImageMagick identify/render fallback, incomplete results no longer fake-complete (parked 'degraded' + extractionIncomplete marker), and reset-faces requeues extract_metadata; existing null-dims items still need the missing-dimensions backfill run [PR pending] [audit 2026-10-24: resolved 2026-10-24: ImageMagick dims + extractionIncomplete marker landed; owner DB shows 0 HEIC null-dims (residual null-dims are 1743 videos + 11 photos, non-HEIC, video backfill pending)]
(2026-10-24) [ISSUE-097] high resolved area:people Stash away ("hide person") and "New person instead" on an ambiguous semisupervised cluster_person_match card 400'd ("ambiguous cluster_person_match requires an explicit existingInstanceId target") because the decideCandidate guard checked requestedTarget==null alone, before the preLinkedStashed shadow branch (spec 039) and the createNewInstanceRequested fresh-instance resolver, neither of which can default to the suggested identity; scoped the guard (defaultsToSuggestedIdentity) to decisions that would actually fall through to the pre-linked suggested identity [PR pending]
(2026-10-24) [ISSUE-101] med resolved area:people every full recognition-run result submission (≈3.8k candidates) exceeded the worker's hard 120 s httpx read timeout and died on httpx.ReadTimeout; the recovery service auto-retried (~4 min later), doubling owner-visible run time and pairing with the pre-V154 persistence cost (88 s × 86 supersede chunks) into 2–8 min result-application windows that held the tenant_clustering_lock (accept/reject clicks blocked); fixed in PR#1040 — submission timeout 600 s and persistence now runs the GIN-backed jsonb containment supersede + rejected-pair pushdown (measured on prod: chunks 88 s → ~0.5 ms, B-card check 220 ms → sub-ms), first post-deploy run pending verification [PR#1040]
(2026-10-05) [ISSUE-102] high resolved area:web the PR#1040 gate rework renamed the suggestions carousel's ONLY content branch from {:else if hasContent()} to {:else if suggestionsStaleNow && hasContent()} — outside an in-flight run no branch matched and the section rendered empty (owner-reported minutes after deploy, no suggestions at all); the in-branch background-refresh banner is itself keyed on suggestionsStaleNow, so the branch reverted to the plain content predicate (chain: unloaded → reset-run spinner → content); no E2E test covers carousel gating, which is why it shipped [PR#1042]
(2026-10-06) [ISSUE-103] med resolved area:web people-list "N tentative" badges disappeared and the instance-detail Tentative filter pill showed 0 while activating the filter rendered cards — both surfaces read autoAssignedCount (tentative OCCURRENCES, by design 0 for cards-only pending rows: 6 of 24 prod people with pending cards had zero occurrences; join loss ruled out: missing_dr=0, superseded_dr=0) while the filter wall renders pending candidate CARDS; fixed cards-first via tentativeBadgeCount() (tentativeCount > 0 wins, occurrence fallback so no existing badge disappears) across list/stashed/explore/merge-dialog/person-multi-select badges and the pill; occurrences stay in header-tentative-count + amber grid rings; reverts the ISSUE-088 pill semantics [PR#1048]
(2026-10-06) [ISSUE-104] high resolved area:people Accept on an ambiguous semisupervised cluster_person_match card on the person page 400'd ("ambiguous cluster_person_match requires an explicit owner-chosen target; use the candidate decision endpoint") because Accept posted the refs-only bulk tentative confirm, whose body can never carry a target (Guard A fail-closed by design); fixed by exposing ambiguous on ClassInstanceTentativeCandidateView and routing such cards — individually and inside mixed batches — through decideClassInstanceCandidate with the current instance as explicit existingInstanceId; Guard A/B untouched [PR#1047]
(2026-10-24) [ISSUE-193] med resolved area:map Manual location overrides (metadata_overrides field 'location') were ignored by map viewport/bounds queries; fixed in this PR [PR#1102] [audit 2026-10-24: resolved 2026-10-24: LOCATION_OVERRIDE_JOIN in viewport+bounds queries (MapQueryService:20-25,353,391)] (renumbered from ISSUE-114: id collision)
(2026-10-24) [ISSUE-123] low resolved area:api OpenAPI /api/search does not document the mediaType and approval query parameters that SearchController/SearchFilters actually accept, so a client generated from the spec cannot discover two narrowing filters the running server honours (the Android screen does not use them in Phase 1, so nothing is broken today) [spec:001] [audit 2026-10-24: resolved 2026-10-24: spec documents mediaType+approval (core-api.openapi.yaml:439-453)]
(2026-10-24) [ISSUE-126] low resolved area:android A person's media wall loads the preview rendition for every tile: PersonMediaItemView exposes only previewUrl (no thumbnailUrl, no dimensions), so tiles decode a far larger asset than the ~120dp cell needs and the viewer opens without an intrinsic aspect ratio — network and memory cost only, nothing is broken; fix upstream by adding thumbnailUrl/dimensions to PersonMediaItemView [spec:055, spec:001] [audit 2026-10-24: resolved 2026-10-24: PersonMediaItemView retired; PersonDetailView.media carries thumbnail/dimensions]
(2026-10-24) [ISSUE-127] med resolved area:android Person detail answers PersonDetailView | MergePeopleResponse, and while a merge is in flight the second shape (no media) is what arrives — PersonDetailDto.media is required by design, so decoding fails and the screen shows its generic error state instead of redirecting to the merged person; merge UI is out of Phase 2 scope, so a user who opens a person mid-merge sees "Die Personen konnten nicht geladen werden." with a retry that succeeds once the merge settles [spec:055, spec:001] [audit 2026-10-24: resolved 2026-10-24: union shape gone; person detail decodes from class-instances endpoints only]
(2026-10-24) [ISSUE-129] low resolved area:android A tentative row whose personAssignmentId is null renders as an undecidable suggestion card (confirm/reject disabled plus a note) instead of a working decision, because POST /confirm is keyed on assignment ids and there is no id to send; the card is not silently accepted or dropped, but it cannot be cleared from the strip until the server fills the id [spec:001] [audit 2026-10-24: resolved 2026-10-24: personAssignmentId retired; candidateId always present, cards always decisionable]
(2026-10-24) [ISSUE-176] med resolved area:upload Unknown upload-session ids answered HTTP 500 instead of 404 on GET /api/upload-sessions/{id}, POST …/files (receive) and POST …/commit: UploadSessionService.requireOwnership (and the session-status loads in UploadReceiveService/UploadCommitService) did a JdbcClient .single()/singleRow() whose EmptyResultDataAccessException ApiExceptionHandler does not map (same class of bug as ISSUE-168, but scoped here); fixed by normalizing only the upload-session lookups to NoSuchElementException → 404 ProblemDetail (no global EmptyResultDataAccessException mapping, SEC-4 403-not-owner semantics untouched), pinned by MockMvc endpoint tests [fix/api-session-404]
(2026-10-24) [ISSUE-184] high resolved area:android A cancellation (WorkManager REPLACE/constraint loss) landing between the upload session's successful 2xx commit and the local markUploaded loop left committed items PENDING — observed in production (session ece22011: 17 committed server-side, only 9 marked, 8 re-uploaded into a fresh session on the next run); fixed by wrapping the post-commit durability epilogue (markUploaded + batch markCommitted) in withContext(NonCancellable) so a cancellation immediately after commit cannot interrupt the marks, while everything before the commit (scan, hashing, uploads, the commit call) stays fully cancellable; pinned by an engine test that parks inside markUploaded while the run is cancelled and asserts all items end UPLOADED [fix/backup-progress]
(2026-10-24) [ISSUE-186] low resolved area:android The SICHERUNG LastRunCard stats row (R.string.backup_stats, rendered by BackupScreen.LastRunCard with uploaded/failed/skipped counts) printed the SAME value three times in every locale — de/en/da all used %1$d for all three placeholders while the call site correctly passes three arguments, so "hochgeladen 18 · fehlgeschlagen 0 · übersprungen 0" showed "18" three times (argument 2 and 3 were silently ignored); fixed by making the 2nd/3rd placeholders %2$d/%3$d in values/strings.xml, values-en/ and values-da/, pinned by BackupStatsFormatTest which formats the resource with three distinct values and asserts the numbers come out in order [fix/android-stats-format]
(2026-10-24) [ISSUE-187] high resolved area:android The silent-re-upload bug class: an UPLOADED device file whose server media item was trashed or permanently deleted (by any family member, from any surface) stayed skipped in the local Room diff forever — the reconciler never looked at server state, so when the file's fingerprint changed (re-save, edit, re-download) it was re-pended and silently uploaded AGAIN, resurrecting deleted content in the library (INV-2 violation; also re-uploaded after every local edit of an intentionally-deleted item); fixed per spec 065 PR-B + spec 044 S3 amendment: the engine now runs a server delta pull (GET /api/sync/changes, change_seq cursor + per-owner epoch, SyncPullClient + SyncPullApplier) BEFORE the local diff — the pending=empty early-exit no longer bypasses the pull — and binds each uploaded row to its server item (content-hash+byteSize match on cached hashes, ChangeBinder, plus serverMediaItemId from the upload path when a server echoes it), persists the binding in Room (serverMediaItemId/serverChangeSeq/serverLifecycleState shadow columns, DB migration 1→2, single-row sync_state cursor table); the reconciler's new shadowed rule (UPLOADED + lifecycle trashed/permanently_deleted → shadowed, checked BEFORE the fingerprint diff) never re-pends those rows and INV-1 is preserved (shadow columns are Room-only; no code path touches MediaStore); a fully-received page is applied atomically with the cursor advance (Room transaction), offline/failed pulls apply NOTHING and never block the run, epoch mismatch/invalid cursor auto-relists from 0; the backup screen shows the distinct non-error library-removed bucket with the spec copy ("Bewusst nicht in Memories" / "Bild bleibt auf deinem Gerät", de/en/da) — four distinct states, never a merged "not synced" label. Residual: historical rows without a cached contentHash stay unbound until their next upload-path hash (no server-id backfill job, per spec scope); committedMediaItemId binding is future-proofing only — today's server never echoes it on the upload response (ids are allocated at commit, no files-list endpoint) [spec:065, PR-B]
(2026-10-24) [ISSUE-191] high resolved area:android Upgrading an existing install to 0.10.2 broke the Backup tab at first DB open: IllegalStateException: Migration didn't properly handle client_intent… — Room validates the post-migration schema against the @Entity definitions and MIGRATION_2_3 created client_intent with an index client_intent (status, createdAtEpochMillis) that the entity never declares (@Entity has no @Index; Room deliberately omits entity-declared indices from validation, so migration-created extra indices always fail), plus DEFAULT NULL on three nullable columns, which left explicit column defaults the entity lacks (tolerated by Room's createdFrom rule only because the migrated table is read-back, not entity-created — latent, not fatal); every device upgrading 0.9.x/0.10.x (DB version 1 since 0.9.0) hit 1→3 via MIGRATION_1_2+MIGRATION_2_3; fixed by mirroring the entity's generated DDL exactly in the migration SQL (index dropped, bare nullable adds in 1→2, DEFAULT NULL removed in 2→3) and pinning it with MemoriesDatabaseMigrationTest, which seeds a faithful v1/v2 file DB, opens through the production MemoriesDatabase.build (real migration chain + Room validation; red before the fix reproducing the production exception, green after) [fix/room-client-intent-migration]
(2026-10-23) [ISSUE-137] med resolved area:android The people suggestion strip has no name-entry surface, so an unnamed new_person_cluster candidate (no existingInstanceId, no payload displayName) renders reject-only — the server 400s an accept without displayName and the app cannot ask for one; the web carousel's create-person mode (type a name, then accept) has no Android equivalent yet, so such clusters can only be declined or left pending — resolved: confirming a create-mode card with no name at all now opens an AlertDialog name prompt (blank stays disabled) and sends accepted + trimmed displayName + createNewInstance=true; cards whose payload already suggested a name still confirm straight away, one-decision-at-a-time and the 401/error paths unchanged (SuggestionsViewModel.requestConfirmWithName/confirmWithName, VM tests cover typed/blank/cancelled/direct paths) [spec:055, feat/parity-android-usability]
(2026-10-23) [ISSUE-120] low resolved area:android Timeline scope chip offers only Alle/Meine: scope=group requires a sharingGroupId (SHARED_MEDIA_JOIN_GROUP pins c.sharing_group_id = :sharingGroupId, so a null id matches nothing), and the app has no group-picker surface yet — the only source is GET /api/sharing-groups. Web parity labels the third option "Familiegruppe"; ship it once a picker (or a single-group default) exists — resolved: TimelineViewModel probes GET /api/sharing-groups once per session; with exactly one status='active' group the "Familiegruppe" chip appears and pins sharingGroupId on both the days and day-items requests, with zero/multiple/failed probes the chip stays hidden and a group tap is a no-op (a failed probe is silent, never an error banner); a group-scoped request also survives filters (favorite/mediaType ride the same requests) [spec:055, feat/parity-android-usability]
(2026-10-18) [ISSUE-130] med resolved area:android The Android app never requests READ_EXTERNAL_STORAGE at runtime, so on devices running Android 9 or older (API <= 28) the backup MediaStore scanner finds nothing to back up until the user manually grants the permission in system settings; on API 33+ behavior is unchanged (no photo-picker permission needed), but a first-launch backup on the target Android 7.1.1 device will silently show an empty album list — resolved: the SICHERUNG flow now derives the media-read permission set from the SDK via StorageCompat.mediaReadPermissions (legacy READ_EXTERNAL_STORAGE on 25–32, READ_MEDIA* from 33), requests it at runtime, reloads albums/status after the grant, and retry hooks re-check the grant; failed runs additionally surface their exception (type + message) in the last-run card so a silent "Fehlgeschlagen" can no longer hide the cause [PR#1125, fix/android-711-media-permission]
(2026-10-24) [ISSUE-133] high resolved area:people Self-sustaining re-clustering loop: every owner_face_recognition run re-marked its owner dirty while persisting its own results (auto-accept / ADR 0102 tentative allocation inside the accept unit fired the same assignment-set-changed hook that user edits use, PeopleDomainCandidateDecisionStrategy.afterAssignmentSetChanged), and since a full-library run lasts 10-14 min > the 10-min RE_ARM_INTERVAL debounce, each 60s FaceRecognitionReconcileScheduler tick saw a fresh marker with no recent job and re-enqueued immediately — 24/7 full-library re-clustering of ~22k faces with zero uploads (prod owner 38a79de6: 19-28 runs/day for a week, 45,930 completed lifetime jobs); fixed by origin-aware dirty markers (V158 origin column: 'user' drains and re-arms as before, 'worker' is provenance-only and ignored by the drain; a later user edit still re-arms via the ISSUE-048 semantics; worker markers never overwrite a pending user marker) [fix/people-reconcile-loop] [audit 2026-10-24: resolved 2026-10-24: V158 origin column + drain honours origin=user (FaceRecognitionReconcileScheduler:92); flyway 179 applied]
(2026-10-24) [ISSUE-135] low resolved area:web tests/discovery-instance-tentative.spec.ts:2086 flakes on the branch baseline (3/10 failures without any change): the scheduled loadTentativeCandidates() background refresh clears tentativeError, removing the tentative-accept-retry button mid-click; independent of the toast-overlay conversion — do NOT "fix" it by waiting for toast expiry in the test, since the wait guarantees the refresh fires and the button disappears (click the retry immediately instead) [a0d73202] [audit 2026-10-24: resolved 2026-10-24: click-retry-without-wait fix landed (discovery-instance-tentative.spec.ts:2183-2192)]
(2026-10-24) [ISSUE-143] med resolved area:search face:unassigned browse showed only undated media due to COALESCE capture ordering crowding out dated items within LIMIT 100 (fixed in this PR: dated-first ordering in the search legs) [PR#1176] [audit 2026-10-24: resolved 2026-10-24: dated-first ordering in all search legs (SearchService:626,732,796)]
(2026-10-18) [ISSUE-171] high resolved area:sharing spec-065 public share ZIP download returned a 22-byte empty archive in production (BadSqlGrammarException "column oa.storage_key does not exist" in PublicShareService.downloadRows, thrown AFTER writeZip had committed the 200+headers → client got a truncated archive + "cannot render error page" log): downloadRows selected oa.storage_key/oa.mime_type from original_assets, which has neither column — the key/mime live on stored_objects (original_assets.stored_object_id; only media_type photo|video) while derived_assets does carry storage_key/mime_type, so rendition paths and the guest media endpoint kept working; second latent defect: rows were resolved inside the open ZipOutputStream, so any downloadRows failure flushed a valid empty-archive EOCD onto a committed 200 instead of surfacing a 500 — fixed by joining stored_objects (LEFT JOIN so.id = oa.stored_object_id, COALESCE(so.mime_type, oa.media_type) AS original_mime) and resolving downloadRows before the first byte is written (skip-missing-files semantics kept); the H2 public-share-test-schema.sql carried the same drift (original_assets WITH storage_key/mime_type, no stored_objects table) so tests could never catch it — schema aligned with production, plus a central-directory ZIP read-back test and a rows-resolved-before-any-byte test [fix/065-zip-original-key]
(2026-10-13) [ISSUE-146] med resolved area:web Hardcoded mixed EN/DE/DA UI strings in apps/web await Paraglide extraction (spec 058 waves 2-3): the API ProblemDetail surface is now localized (messages/api-messages en/de/da) and the per-user locale preference is stored/served, but every Svelte surface (setup/admin, storage banners, timeline, viewer, people, albums, search, toasts…) still ships literal strings, so non-English users see a mixed-language UI until the frontend extraction waves land — resolved by spec 058 waves 1-5 (2163 keys/locale en/de/da, hardcoded-string guard with committed baseline, EXTRACTION.md exclusions documented) [spec:058]
(2026-10-12) [ISSUE-140] high resolved area:discovery PR#1144 interpolated REPRESENTATIVE_FACE_THUMBNAIL_SUBQUERY into suggestSimilarInstances' embedding-path candidate-rows statement where the correlated subquery ran once per candidate embedding ROW (~4,675 executions for ~350 persons → 39.3 s per media-viewer "Assign person" dropdown open, measured live; pre-#1144: 75 ms) — not an index problem: EXPLAIN showed the existing idx_class_instance_occurrences_instance_state* indexes already serve the subquery (batched form: ~110 ms for the whole universe) and the pinned-rep ORDER BY CASE is not indexable; resolved by batching thumbnail resolution into ONE post-loop query over distinct scored instance ids (chunks of 1000, same shape as personRanking personRows), display-only — all six #1144 thumbnail regression tests pass unchanged, endpoint ~200 ms [PR#1146]
(2026-10-12) [ISSUE-112] low resolved area:discovery suggestSimilarInstances (the instance-suggestions endpoint feeding the media-viewer "Assign person" dropdown, NOT the three picker rankings) selected its preview thumbnail from an arbitrary accepted-or-tentative crop (three inline subqueries with LIMIT 1, no ORDER BY, no named-instance guard) — the same bug class fixed for personRanking/mergeTargetRankings/mergeSuggestion in 9e11330d, deliberately deferred then to keep that PR thumbnail-only; all three subqueries (embedding path, ILIKE name-search fallback, blank-query named-instance listing) now interpolate REPRESENTATIVE_FACE_THUMBNAIL_SUBQUERY (pinned rep first, accepted beats tentative, named instances never serve a tentative crop, unnamed keep tentative fallback); six regression tests, red on unmodified main / green with fix (326/326) [PR#1144]
(2026-10-12) [ISSUE-138] high resolved area:api Deployed memories-api failed to start after PR#1137 merge (run 4284, Deploy API): TimelineQueryService gained a second public constructor (3-arg with Clock) without @Autowired, making Spring's constructor resolution ambiguous — it fell back to a no-arg default that does not exist (NoSuchMethodException at context startup). Unit tests construct the service directly and the smoke job only compiles, so nothing caught it pre-deploy. Fixed by @Autowired on the 3-arg constructor + ApplicationContextRunner wiring regression test (TimelineQueryServiceWiringTest, red/green verified); deploy re-verified green in run 4294 [PR#1140]
(2026-10-12) [ISSUE-121] high resolved area:timeline Android timeline→viewer hands the photo to a SECOND, empty TimelineViewModel: NavHost wraps every destination in NavBackStackEntry.LocalOwnersProvider, which (verified in navigation-compose 2.10.2 bytecode) provides LocalViewModelStoreOwner = that destination's NavBackStackEntry, so the viewModel(factory) in composable(Routes.PHOTOS) and the one in composable(Routes.MEDIA_DETAIL) resolved to different ViewModelStores — the comment in MemoriesApp.kt claiming they were shared was wrong, and MediaViewerScreen's if (items.isEmpty()) LaunchedEffect { onBack() } guard then immediately popped the viewer back to the timeline (flicker, viewer never opened); resolved by hoisting TimelineViewModel to app scope the same way SearchViewModel/MapViewModel are (reset on session change, ensureLoaded() re-entry on the screen), plus a belt-and-braces friendly empty state in MediaViewerScreen instead of the instant pop-back; the other viewer entry points (search/album/map/person) were unaffected — their ViewModels were already hoisted [fix/viewer-vm-scope]
(2026-10-11) [ISSUE-118] high resolved area:api Spec-054 merge (PR#1111) crash-looped memories-api in the deploy: three beans gained a second public constructor (StorageAvailabilityService, AdminBootstrapService, StoragePathsAdminController), which disables Spring's implicit constructor autowiring — context startup failed until @Autowired was added on the intended constructors; unit/PG tests construct services manually, so only the real container boot surfaced it; resolved via two hotfixes [PR#1113, PR#1115]
(2026-10-11) [ISSUE-114] med resolved area:discovery Person-history rejected batches for zero-occurrence candidates (owner-rejected weak single-face/cluster suggestions rejected before any Spec-048 tentative occurrence rows were written — prod person 32cd0e03: 8 recent rejected candidates, occ_rows 0,0,1,1,0,1,1,0) serialized with occurrenceCount=0 and no items, so the Historie wall rendered text-only cards with NO thumbnails even though the candidate still carries its occurrence_refs; fixed by synthesizing one history item per face_annotation ref (capped at 36 = the frontend wall cap) with occurrenceKey=ref:<type>:<id>, media/dimensions resolved via the same owner-scoped discovery_results→media_items join the tentative-candidate listing uses; candidates WITH superseded rows keep the real rows as the single source of truth (no ref duplication) [feat/history-rejected-thumbnails]
(2026-10-10) [ISSUE-017] high resolved area:metadata MetadataController passes ownerId=null into createOverride → requireOwner always denied → every metadata-override call 403'd (fail-closed but the feature was dead, SEC-8); fixed by resolving the owner via LocalAuthService.currentSession(authentication).ownerId() on create and the new DELETE revert endpoint, with server-side per-field validation and the FR-041/SC-019 revert flow (activeMetadataOverrides + sourceValue on media detail) [feat/metadata-edit-api, PR#1095]
(2026-10-08) [ISSUE-108] med resolved area:discovery Suggestion-candidate listing: the stash-target anti-join (NOT EXISTS + CAST(occurrence_refs AS VARCHAR) LIKE '%'||occurrence_key||'%') was evaluated for all ~3.5k pending rows (top-N sort gives the LIMIT no early exit; Seq Scan on class_instances inside SubPlan 1 × 3066 loops) so the PAGE query alone cost ~2.2-2.3 s live after PR#1075; fixed by the two-step post-filter: page SQL drops the anti-join, fetches a limit + 24 window, then applies the identical SQL predicate as one batch over just the fetched ids (max 2 top-up rounds, trim to limit; limit == null = one full fetch + same batch filter) — live EXPLAIN ANALYZE 2166 ms → ~30-35 ms, differ test over all listable rows 0 deviations; countCandidates intentionally keeps the full anti-join (a total must see every pool row, set-rewrite measured 740 ms) [PR#1076]
(2026-10-07) [ISSUE-107] low resolved area:web zoomed face tiles (suggestion carousel cards + dialog crops, people-page "Vis område") rendered blurry — they CSS-zoomed a whole-photo 512px thumbnail (scopeCropFit), leaving only ~80-120 source px for small faces; fixed by the spec 052 pre-composed scope-crop assets: web prefers crop=scope assets with whole-photo fallback (PR#1058), worker generates face_scope_thumbnail (PR#1059), and the backfill is complete in production (6,415/6,415 jobs completed, every active people face has an asset) [spec:052]
(2026-10-05) [ISSUE-100] med resolved area:people semisupervised-v2 (ADR 0104) shipped with component-cap fail-closed: a single connected component above the reference HAC member cap (512 deployed, 768 in dev with the experiment harness) aborted the recognition run with ComponentTooLargeError instead of clustering — dense libraries that grow one giant threshold-graph component saw repeated job failures; no silent degradation by design [PR#1035]; fixed by the D4 scalable-exact partition (exact global-argmin over cluster sum vectors via reducibility-justified lazy queue, reference full-matrix partition kept as test oracle, tie-break parity fixture-enforced, per-component cap now SCALABLE_MAX_HAC_NODES=30000 with fail-closed above; production component of 13,649 nodes now partitions in minutes) [PR#1036]
(2026-10-04) [ISSUE-098] med resolved area:media:metadata URL-encoded ingested filenames made MediaJobHandlers.timestampFromFilename match the YYYY-MM-DD patterns against the encoded name, reading '%40'+DD as a 4-digit year (4001..4027) with no year guard anywhere on the write path → start-page timeline sections with year >4000; fixed by percent-decoding the filename (literal '+' preserved, malformed escapes fall back to raw) before filename date parsing plus a plausibility guard (1800..current year+1, worker zone) on the final captureTimestamp in extractMetadata and on the incoming captureTimestamp in WorkerResultService.acceptMetadataCandidate (implausible value treated as absent, so COALESCE preserves the stored timestamp/timezone/confidence); 21 corrupted media_items rows repaired directly in postgres (capture_timestamp reset from the decoded filename), year>4000 count = 0; deployed as 0.65.1534 with green smoke [PR#1028]
(2026-10-03) [ISSUE-092] high resolved area:ci PGDG dropped the bullseye-pgdg apt suite after Debian 11 LTS → every postgis/postgis:16-* base image 404s on its Release file → postgres image build (and with it the whole deploy pipeline) failed; fixed: deploy/postgres/Dockerfile now bases on official postgres:16 + postgresql-16-postgis-3 + postgresql-16-pgvector (PG major 16 and /var/lib/postgresql/data unchanged) [PR#1000]
(2026-10-03) [ISSUE-089] high resolved area:api PR#996 broke API boot: PeopleClusteringResetService gained a second public constructor (legacy 4-arg delegating + new 5-arg with TenantClusteringLockService) with NO @Autowired on either → Spring cannot pick a constructor at context startup ("No default constructor found"), memories-api crash-looped in production (~15 min, caught by Deploy API smoke; unit/PG tests construct the service manually so only the real context boot hits it); resolved: @Autowired on the 5-arg primary constructor (mirrors ClassInstanceDiscoveryService pattern), production healthy again (Deploy + Smoke green, 0 restarts); follow-up tracked separately: an ApplicationContext boot smoke test in CI so constructor-ambiguity regressions fail in PRs, not deploys [PR#996, PR#997]
(2026-10-02) [ISSUE-088] med resolved area:web DiscoveryInstanceDetail confirmation pills never filtered the person wall (instanceFilters() omitted the confirmation state, so "Confirmed" showed all occurrences incl. tentative — badge 15 vs 810 wall items) and the Tentative badge counted pending candidate CARDS (21) instead of tentative occurrences (894), making header/pills/wall three inconsistent numbers; resolved: confirmation=accepted|tentative param on the class-instance days/dayItems endpoints (absent = legacy both-states), Confirmed pill filters + refetches the wall (one-way prop instead of bind: so the refetch comparison works), Tentative badge shows the tentative occurrence count (autoAssignedCount); bucket counts stay consistent with item listings on both sort bases [PR#995] — SEMANTICS REVERSED 2026-10-06 by PR#1048: pill/badges back to pending-candidate cards (tentativeBadgeCount, cards-first with occurrence fallback); occurrences remain only in header-tentative-count chip and amber grid rings
(2026-09-29) [ISSUE-076] high resolved area:ci All image builds/deployments on main failed since memories-deploy-runner-2 activation: build-deploy.yml/daily-full-tests.yml hard-coded MEMORIES_DEPLOY_ENV_FILE=/home/chris/git/memories/deploy/.env (runner-1 checkout), but deploy-runner-2 has an ephemeral workspace without that file → every Build * image job died with "couldn't find env file" (runs 3269/3279/3295/3297, affected all agents' merges); resolved: env file moved to stable host path /home/chris/git/memories-deploy-runner/.ci-secrets/deploy.env (shared by both runners, 750/640 perms), workflows updated (#958), backlog merges shipped via compose full-rebuild trigger (#959); verified run 3325 green: all 8 images built, all services deployed, smokes passed [PR#958, PR#959]
(2026-09-28) [ISSUE-073] med resolved area:workers workers/people could not decode HEIC (pillow without pillow-heif → PIL.UnidentifiedImageError) so discovery_people failed for all image/heic items — HEIC photos (1,686 active) had never had face detection/results (zero people discovery_results for heic since 2025-05); surfaced by the EXIF-migration resets (743 failed jobs, terminal, no retry loop); resolved: pillow-heif registered at face_model import time, detection + embedding-backfill covered, orientation-6-HEIC equivalence tested (109 tests green), CI smoke job installs the dep, all 1,107 failed HEIC items re-queued via reset-faces and detection verified live (HEIC results appearing within 30 min) [PR#945, PR#942 migration]
(2026-09-27) [ISSUE-071] high resolved area:web-tests admin-operator.spec.ts never mocked /api/admin/upload-events added to the admin setup page load() Promise.all by PR#917 (509f3513), so the request returned "404 page not found", the promise rejected, tabs never rendered, and all 5 US5 admin-operator tests (desktop+mobile) failed; fixed with a describe-level upload-events fixture, 10/10 green — the same commit also gated the discovery-instance-tentative history page 2 so its transient toHaveCount(1) cannot race the auto-load-more observer [PR#917, fix/admin-operator-e2e]
(2026-09-25) [ISSUE-013] high resolved area:upload Upload session get/receive/commit never verified session ownership — anyone with the session UUID read session state, staged bytes into it, and committed media into the victim's library (SEC-4). Fixed: UploadSessionService.requireOwnership (created_by_user_id for user principals; api_credential_id + media:upload scope for credential principals, no cross-access, Super Admin not exempt) runs as the first statement of receive()/commit() before any staging write or markCommitStarted, and UploadFileController guards get() before returning session state; unknown session ids still surface as 404 before 403 [docs/adr/0097, PR pending]
(2026-09-23) [ISSUE-066] high resolved area:ci deploy-api job (introduced by #826's needs-chain change) lacked always() in its if, so when deploy-postgres was skipped (postgres unchanged) Forgejo skipped deploy-api BEFORE evaluating the if — even though the expression accepted 'skipped': runs 2742 (#831) and 2753 (#835) built API images successfully but silently never deployed them, leaving production memories-api stale while web deployed (user-visible: 404 on the new face-unassign endpoint); fixed in #837 by mirroring deploy-web's always() && pattern; the stuck #831/#835 API changes went live via the api=true touch PR that archived this issue [PR#837, runs 2742/2753]
(2026-09-23) [ISSUE-062] med resolved area:ci Concurrent CI jobs can exhaust runner memory: run 2707 (scoped API tests for PR #816) hung from 17:32Z with zero log output after java.lang.OutOfMemoryError thrown from the UncaughtExceptionHandler in thread "testcontainers-ryuk" while run 2708 (web tests, PR #817) ran concurrently on the same host — same failure class as the v1.21.3 build OOM whose sequential-deps fix covered only build jobs; the hung run had to be cancelled manually and CI re-triggered via a real merge-base update because Forgejo has no rerun API for scoped jobs; mitigation candidates: host-wide concurrency group serializing scoped-test jobs, lower Gradle test JVM heap on runners, or ryuk disable with explicit container budgets [runs 2707/2708/2710, PR#816] — resolved (recurrences 2740/2749): job-level concurrency: is not enforceable on this Forgejo (concurrency groups are stored per RUN only), so every memory-heavy test step in both workflows now sources a host-file flock shared by both memories runners on the host (scripts/ci/acquire-heavy-test-lock.sh, CI_HEAVY_TEST_LOCK under /home/chris/git/memories-deploy-runner/.ci-locks) and the Gradle test JVMs are bounded (2g heap / 1g direct, one fork per module) [fix/ci-oom-serialization]
(2026-09-23) [ISSUE-063] med resolved area:ci CI has no test job for workers/people — build-deploy.yml contains pytest/unittest jobs for workers/ml and workers/media but zero references to workers/people (verified: no workers/people or people pytest invocation in the workflow), so people-worker test regressions reach main unnoticed and the scoped-detection path only builds/deploys the image without running its tests; add a people-worker test job mirroring the ML worker smoke-test pattern (pip install numpy pillow httpx pytest && PYTHONPATH=workers/people/src python -m pytest -q workers/people/tests) gated on the worker_people change scope [quality-consensus-worker recon]
(2026-09-23) [ISSUE-064] med resolved area:workers 8 pre-existing failures in workers/people/tests/test_face_model.py (baseline on pristine main 973e8c41: 8 failed / 77 passed): commit 97bf4cc9 changed _detect to flat-2D SCRFD output decoding while the test mocks still feed batched CHW tensors, so detector-decode tests fail on mocked-session output shape; production detection unaffected (real ONNX sessions return flat output) but the suite is red and masks future regressions until mocks are updated to the flat-2D format; discovered while landing the weak-match consensus change [commit 97bf4cc9, PR quality-consensus]
(2026-09-22) [ISSUE-061] med resolved area:web Thumbnail retry loop self-retriggered: thumbnailRetry.svelte.ts ran $effect(() => thumbnailRetryRequest.subscribe(() => retryNow())), the subscriber executed synchronously during effect setup, and retryNow() → beginRetryProbe() read the attempts $state — so every attempts += 1 in scheduleRetry re-ran the effect, reset attempts to 0 and probed again; permanently failing thumbnails never reached maxAttempts and hammered the server with cache-busted requests (isolated repro: 5,950 requests/30s for 6 tiles; caused albums.spec.ts:1308 CI timeouts). Fixed by untracking the subscribe/retryNow calls; new Playwright guard counts probes for a permanently-404 tile (red 2,665 probes/15s before, bounded after) [fix/thumbnail-retry-loop]
(2026-09-19) [ISSUE-049] high resolved area:people People-domain accepted face occurrences carried class_occurrence_embeddings in two incompatible model spaces — legacy buffalo_l/arcface-v1 (old worker-ml stack, removed PR#702) vs current arcface-r50/w600k-v1 — and 115 of 116 active unnamed clusters had ONLY legacy rows while all 21 named persons had ONLY current rows; all similarity/matching sites group by EmbeddingSpaceKey and pair only within a shared space, so legacy clusters could never be merge-suggested or auto-matched against named persons (the "no merge suggestions" report); fixed via embedding-only owner_face_embedding_backfill job (PR#763) re-embedding 4,416 affected occurrences from stored bboxes without re-detection — staged canary verified end-to-end, then 3,038/3,038 media jobs completed, coverage gap 0 (arcface rows 7,244 → 11,660); residual note: ~1.4k owner_tentative pending_face_embeddings rows remain legacy-only (accepted-occurrences-only backfill by design — those faces re-enter matching only via re-clustering) [spec:039 recon, PR#763]
(2026-09-19) [ISSUE-048] high resolved area:people Person edits never re-armed face recognition — the recognition loop only armed on new embedding evidence (face_recognition_cursor watermark, V121), so previously considered owner_tentative faces were never re-matched against the updated assignment set and stale "new person" clusters stayed separate after naming/merging similar persons; fixed via owner dirty-markers on person edits + FaceRecognitionReconcileScheduler (60s drain) in PR#760, plus the missed candidate-accept decision paths (decideCandidate/acceptCandidateForExistingInstance/autoAcceptExistingInstanceCandidate — the actual review-UI flow) in PR#761; rejections intentionally do not re-arm to avoid re-proposing released faces [spec:038, PR#760, PR#761]
(2026-06-20) [ISSUE-001] med resolved area:ci scoped-tests failed on non-main branches; fetch --depth 2 too shallow for multi-commit pushes [PR#409]
(2026-06-20) [ISSUE-003] med resolved area:discovery 409 'Could not parse discovery result JSON' persisted after PR#301; diagnostic logging added in PR#303. Fixed in PR#411 — jsonFromObject validates JSON strings before passthrough, all discovery service jsonList/jsonMap degrade gracefully instead of throwing IllegalStateException [PR#411]
(2026-06-20) [ISSUE-005] med resolved area:tests Playwright timeline destructive face reset test uses "Check people again" menuitem text removed in commit 6e2d8d97; rewritten to use RerunDiscoveryDialog flow via /api/media/{id}/discovery/reset [PR#416]
(2026-06-14) [ISSUE-004] high resolved area:plants YOLO-M bbox float rounding → bounds outside [0,1] → 400 on result submit → fail-retry flood → pool exhaustion [PR#348]
(2026-06-11) [ISSUE-002] high resolved area:discovery PGobject JSONB parse fail → 409 on /api/media/{id}/discovery [PR#301]
(2026-06-28) [ISSUE-006] high resolved area:discovery minBboxAreaPx hardcoded null for plants_image -- 6400 never sent to worker [PR#491]
(2026-10-24) [ISSUE-115] med resolved area:android Android media-viewer video playback was blocked server-side: no user-facing /api/media/{id}/original existed and the video_preview job reported video_preview_not_implemented, so ExoPlayer requests 404'd. Server half now shipped (spec 059): range-capable GET /api/media/{id}/original, ffmpeg-generated 720p H.264 video_rendition + posters, original fallback streaming [spec:055, spec:059] [audit 2026-10-24: resolved 2026-10-24: GET /api/media/{id}/original + video_rendition + Android playback shipped (spec 059)]
(2026-10-24) [ISSUE-035] med resolved area:workers O(n²) pure-Python face clustering (_density_clusters, _best_existing_person_match) looped scalar cosine distances over ~16k×512-dim vectors (19h runtime); vectorized with NumPy pairwise distance matrices, golden tests prove identical cluster output [fix/face-pipeline-embedding-flush]
(2026-10-24) [ISSUE-058] med resolved area:discovery Instance/person/face thumbnails whose derived asset was still generating rendered as broken 404 images (no onerror, no retry); web now keeps the <img> mounted but hidden behind a spinner placeholder, retries with a cache-busted probe (max 10 × 6s), short-circuits on SSE (processing_summary_changed/discovery_results_changed) while thumbnails are pending, and pages silently refetch JSON for pending imagery; genuinely-null URLs keep "No preview" [fix/instance-thumbnail-spinner]
(2026-09-23) [ISSUE-065] med resolved area:ci Deploy API failed (run 2729, PR#826) because V136's one-time HNSW index build (20k×512-d embeddings) took ~4.5 min and outlasted the deploy compose up --wait 300s healthcheck — the memories-api container had already been recreated with the new image, completed migration + boot healthy on its own ~4 min later (flyway v136 success, idx_occurrence_embeddings_hnsw present), so production was unaffected beyond the delayed rollout and a red deploy run; lesson: migrations that build large/HNSW indexes need a longer wait-timeout on postgres-coupled deploys or CONCURRENT index builds [PR#826, run 2729]
(2026-09-25) [ISSUE-067] med resolved area:web Search page person dropdown appeared disabled/unselectable on the deployed instance: /api/discovery/domains/people/class-instances returns ClassInstanceView (classInstanceId, occurrenceCount, representativeBbox) while the frontend PersonRecordView expected personRecordId, so every keyed-each key was undefined → Svelte each_key_duplicate runtime error aborted the update → select stayed disabled with no options; getPeople() now normalizes both shapes, only named people with ids are offered, and selects carry explicit [&>option] dark-theme styling [feat/search-tokens-web]
(2026-09-25) [ISSUE-068] med resolved area:web Media-viewer face edit banner "Bekræft" was a silent no-op for tentative faces: clicking an unnamed face set editingDomain='faces' without loading an edit config (the people domain key is 'people', so getDomainEditConfig('faces') always failed), the DiscoveryEditModeOverlay therefore never mounted and the banner called triggerConfirm() on a null ref — and even with an overlay mounted, handleEditConfirm only applied staged region edits, so no path accepted the pending recognition candidate; confirm now accepts this photo's pending people candidates that reference faces (via the same decideClassInstanceCandidate endpoint the suggestion carousel uses, matched through overlay payload.faceAnnotationIds/occurrenceRefs), reports a brief inline message when there is genuinely nothing to confirm (message/error surfaces were also hidden during edit mode), and both the banner and the sidebar Persons header expose icon-only confirm (green check) / cancel (red undo) actions [fix/face-edit-confirm]
(2026-10-24) [ISSUE-075] med resolved area:ci Version-bump-only commits (gradle.properties) fell through detect-scopes.sh's catch-all branch (pattern gradle/* matches only the gradle/ dir, not gradle.properties), setting mark_all_services + fallback smoke — every version-bump PR branch ran the entire JVM/web/CLI/ML suite (30+ min) for a 1-line bookkeeping change; gradle.properties is now inert in scope detection (image versions come from git tags via compute-version.sh, gradle.properties only feeds the local Gradle jar version) [PR]
(2026-09-28) [ISSUE-074] high resolved area:timeline Deep timeline axis jumps rendered placeholder walls then the scrollbar visibly jumped repeatedly (once per network round trip) until the view settled at the library top — schedulePreviousDaysRevealCheck's inverted geometry (sentinel above the viewport satisfies top < innerHeight+900 with no bottom >= 0 test) misfired on every hydration batch and revealPreviousDays prepended 8 day sections with no mutation-time scroll anchor (removed by PR#895), a deterministic feedback loop driving visibleDayStartIndex to 0; fixed via corrected reveal predicate (sentinelWithinRevealMargin in timelineReveal.ts), mutation-time anchor capture/restore in revealPreviousDays composing with the hydrateDays merge-time anchor, observer/reveal suppression while jumpingDayKey is set, overflow-anchor:none on the timeline route (single JS scroll owner), persist guard while jumping/hydrating, and viewport-intersecting anchor preference; regression test proven failing on main (window drift 20→0), passing patched [PR#939]
(2026-10-24) [ISSUE-095] med resolved area:people A person's representative thumbnail could come from a tentative (auto-assigned, unconfirmed) face occurrence: the people representative enrichment ranked accepted and tentative faces together (tentative could win), the auto-populated representative_occurrence_id pin ignored occurrence state, the "Set thumbnail" picker (face-thumbnails) listed tentative occurrences, and the enrichment-absent fallback passed a tentative occurrence's bbox through, letting the frontend crop the plain thumbnail into a tentative face region; representative paths now require accepted state — named persons with no accepted face fall back to the plain media thumbnail (no face crop, no bbox), unnamed candidate persons keep accepted-first tentative crops as fallback, and the face picker lists approved occurrences only [PR pending]
(2026-10-24) [ISSUE-142] high resolved area:ci CI computed bogus 0.0.x versions on main (runs 4412/4424 → images tagged memories-api:0.0.3, web UI showed v0.0.3): detect-scopes.sh ran before compute-version.sh in the changes job and its git fetch --depth 1 origin "$BEFORE_SHA" wrote .git/shallow, truncating history at BEFORE_SHA so no tag was reachable and all compute-version.sh describe fallbacks failed into the 0.0.{commit_count} bootstrap branch (count from shallow root); intermittent because the fetch does not always succeed and BEFORE_SHA is all-zeros on force-push/branch-creation; fixed by reordering the changes job so compute-version.sh runs immediately after checkout (tags intact) and by failing loudly in compute-version.sh when .git/shallow exists AND no tag resolves (tagless non-shallow repos keep the bootstrap fallback) [PR#1172] [audit 2026-10-24: resolved 2026-10-24: compute-version runs before detect-scopes (build-deploy.yml:77-91) + fail-loud shallow check (compute-version.sh:34)]
(2026-10-24) [ISSUE-154] low resolved area:tests Playwright upload-status.spec.ts "US1 upload and status smoke" flaked/failed terminally on desktop+mobile (runs#4682/4688/4699). Verified root cause (not i18n — EN copy uploadReadyCount/uploadSummaryLine intact and PR#1202 touched no uploads files): the file inputs' on:change handlers attach during SvelteKit hydration, but page.goto resolves on the load event, which can precede hydration. A setInputFiles/change fired too early is silently dropped → files stays empty → the "Upload selected files" button (disabled while files.length === 0) never enables → :269 click times out at 90s (terminal); :52/:131 fail an upstream assert then pass on retry. Deterministic mechanism, load-triggered: loses on a cold preview server under a queued CI runner, wins on a warm box. Fixed by gating each file-input test on a client-boot signal before interacting — poll AppShell's onMount processing-summary fetch (gotoUploadsHydrated helper for :52/:269; the tests' own summary counter for :131) [PR#1207] [audit 2026-10-24: resolved 2026-10-24: gotoUploadsHydrated boot-signal gate landed (upload-status.spec.ts:57-81)]
(2026-10-24) [ISSUE-157] low resolved area:videos extract_metadata for videos loads the whole original via byte[] downloadOriginal (CoreApiWorkerClient contract) before ffprobe runs — memory spike proportional to file size (a 4 GB clip costs ~4 GB heap); worker lane documented the tradeoff in MediaJobHandlers javadoc; proper fix is a streaming download overload of the worker client [spec:059] [audit 2026-10-24: resolved 2026-10-24: downloadOriginalToFile streaming landed (MediaJobHandlers:164-172)]
(2026-10-24) [ISSUE-160] high resolved area:cast Every token-authenticated media fetch 500s after the bytes are served: MediaTokenClaim is not Serializable, but the JDBC-backed session store serializes the SecurityContext after each request, so ?st= casts fail with NotSerializableException (Chromecast shows cast icon, never the image — first on-hardware cast test 2026-10-14, 20 occurrences in api logs). Mitigated by fix/media-token-serializable (claim implements Serializable + JDK-serialization regression test) [spec:060] [audit 2026-10-24: resolved 2026-10-24: MediaTokenClaim implements Serializable + JDK-serialization regression test landed]
(2026-10-24) [ISSUE-161] med resolved area:people Assign-person dialog showed person A's face on person B's suggestion row: the suggestion list was keyed positionally ({#each} without key), so rows kept the previous list's DOM/img and the stale bitmap stayed visible until the new crop decoded (ThumbnailImage only hid during 'pending'/'exhausted', not 'loading'); additionally refreshFaceDialogState applied a refreshed list without bumping faceSuggestionRequestId so a slow in-flight debounced fetch could clobber it; fixed by keying rows on instanceId, an opt-in hideWhileLoading on ThumbnailImage used only by dialog rows, invalidating in-flight fetches in refreshFaceDialogState, and requesting the 256px face rendition (?size=256) for the 48px row [fix/face-suggest-perf-thumbs]
(2026-10-24) [ISSUE-164] med resolved area:people semisupervised-v3 result ENVELOPE stamped clusteringContractVersion=1 (hardcoded SEMISUPERVISED_CONTRACT_VERSION in main.py single-pass stamp) while its candidates correctly carried 2 via _stamped — Core rejected the first live G4 run's submission with 400 "expected 2, got 1" (job 8ae71e61, 2026-10-16); the v3 worker suite pinned candidate stamps and job-settings validation but never the envelope stamp; fixed by single-source contract_version_for_method() shared by _stamped and the envelope, plus envelope-stamp regression tests [ADR 0113 V5 / fix/v3-contract-version-stamps]
(2026-10-24) [ISSUE-165] med resolved area:people owner enqueue exports cio.occurrence_key as faceAnnotationId and 15 of 388 tentative rows carry non-canonical ":0" keys (generic_evidence occurrence convention) — Core's releasedTentativeFaces validator requires canonical UUIDs and rejected the v3 G4 submission with 400 (job 604def57, 2026-10-16); fixed worker-side by a canonical-UUID filter on releasedTentativeFaces (conservative: such rows stay milestone'd instead of being released under an invalid id); root data quirk (occurrence_key with index suffix in state=tentative) remains open for a data-side cleanup [ADR 0102 / fix/release-tentative-canonical-uuid]
(2026-10-24) [ISSUE-169] med resolved area:people semisupervised-v4 margin-demoted B cards (V6: margin < 0.05 → reviewOnly + ambiguous + conflictingPersonIds) omitted the compatiblePersonIds array that Core's semisupervised envelope validator requires on every ambiguous candidate — the first live v4 run (job fc05c137, 2026-10-19, ~31 min of HAC work) was rejected at submission with 400 "requires a compatiblePersonIds array"; fixed worker-side: the compatible set of a margin-demoted card is its own suggested target (v4 has no wholeGroup concept; the runner-up lives in conflictingPersonIds), pinned by extended demotion fixtures [ADR 0116 V6 / fix/v4-margin-demotion-envelope-shape]
(2026-10-24) [ISSUE-170] low resolved area:ci post-merge CI run #5201 failed once at Build converter worker image — docker registry pull retries ended in "Client Closed Request" (transient registry/network flake, no code change involved; merge was web-only spec-065 work); rerun API returns 404 (not supported), so the run was left red; the scoped detection had already skipped Deploy converter worker and all other image builds/deploys were green [PR#1300] [audit 2026-10-24: resolved 2026-10-24: one-off transient registry flake, no recurrence, nothing fixable in repo]
(2026-10-24) [ISSUE-172] med resolved area:videos ffprobe coded width/height were stored as media_items dimensions without applying the display-matrix rotation, so portrait videos (coded landscape + ±90°) got landscape timeline tiles and their autorotated posters were cover-cropped top/bottom; fixed for new ingests by rotation-aware VideoTools.probe and repaired on this instance via the derived-poster-based swap SQL (9 rows 1920×1080→1080×1920, idempotent re-run = 0) — other self-hosters with pre-fix portrait videos must run the same SQL [PR#1304] [audit 2026-10-24: resolved 2026-10-24: rotation-aware VideoTools.probe landed; repair SQL already run on this instance (9 rows swapped)]
(2026-10-24) [ISSUE-192] high resolved area:android Device rows holding DUPLICATE content (same SHA-256, e.g. one image in several albums/folders) never bound via the delta pull's hash rule because ChangeBinder demanded a UNIQUE row per hash, so a server-side trash/purge verdict never reached any copy: the rows stayed plain UPLOADED and the reconciler kept counting them skipped ("N übersprungen") while the library-removed card ("Bewusst nicht in Memories", spec 065 S8/INV-2) never appeared and deletion stayed invisible on the device — resolved: hash binding now fans out to ALL unbound UPLOADED rows matching (contentHash, byteSize) — content identity is the server's identity rule (spec 001 FR-068e/FR-068b1 dedupe), so copies of one hash are one server identity, not an ambiguity — and where several changes share a hash (spec 065 C5 purge+reupload twins) the binder elects per hash before binding: lifecycle active > trashed > permanently_deleted, tiebreak higher changeSeq; already-bound rows keep their direct rule-1 binding (never rebound to another item) [fix/sync-binder-duplicates]
(2026-10-24) [ISSUE-197] high resolved area:android Consumed-history-not-rebound after the binder-bug builds: 0.10.3 (the first build with sync pull, pre-#1359) pulled the change feed with the unique-match ChangeBinder that under-bound duplicate content — the pull still CONSUMED the entire feed (cursor → 28594) while the consumed purge/trash tombstones produced zero shadow writes, and consume-once cursor semantics mean 0.10.4 (with the fixed #1359 fan-out binder) then has nothing new to pull: the historical purge entries are lost client-side, rows stay plain UPLOADED, the reconciler keeps counting "14/18 übersprungen" and "Bewusst nicht in Memories" never appears; resolved one-time client-side by Room migration 3→4 (DB v4): UPDATE sync_state SET lastCursor = 0 keeps the epoch so the next pull is a PLAIN full re-list (no destructive epoch-reset path) whose replayed tombstones re-bind correctly through the fixed fan-out binder — safe because tombstone retention is PERMANENT and a replay from cursor 0 is always possible (spec 065 S2); fresh installs are unaffected (absent sync_state row = cursor 0 via entity/repository defaults) [fix/sync-cursor-relist]
[ISSUE-199] high mitigated area:android First-sync/epoch-reset resurrection (spec 065 INV-2 violation, F4): on a fresh install (or after epoch/cursor reset) the delta pull's hash rule had nothing to bind (it matches only already-UPLOADED rows), so reconcile pended EVERY scanned file and the run uploaded all of them — server dedupe absorbed active/trashed content harmlessly, but PURGED content had been GC'd and findExistingOwnerMedia excludes permanently_deleted, so the re-upload created fresh media_items rows: deleted photos resurrected; same shape for a long-offline device whose rows are PENDING with cached hashes. Mitigated by first-sync pre-binding: a COMPLETED full listing (cursor-0 start, epoch reset, or invalid-cursor re-list — never an incremental delta, which cannot prove library-wide absence) now binds scanned files to the elected per-hash change (same ChangeBinder election as PR#1359) BEFORE reconcile pends them; matched rows are written UPLOADED+bound in one write (matched-active → skipped, matched-trashed/purged → shadowed, never uploaded), no-match files upload normally with their hash cached in-pass (each pending file hashed at most once per run, replacing the upload I/O, never adding to it). NOTE: the re-upload path for accepted reinstate-on-purged (ISSUE-190) remains a separate, open item. [fix/sync-firstsync-binding]
(2026-10-25) [ISSUE-190] med resolved area:sync Reinstate-on-purged acceptance (spec 065 C6/PR-D) is an RE-AUTHORIZATION, not a resurrection: the media_items row stays permanently_deleted and the client is expected to re-upload its INV-1-untouched original, but NO server or client path today performs that re-upload automatically — the Android shadow rule (INV-2) keeps blocking the file, so after an accepted reinstate the item returns to the library only if the user re-uploads it manually (e.g. re-add the album); a follow-up should either flip the shadow state on accepted reinstate (delta pull carries no "reinstate-authorized" signal) or define an explicit re-upload intent — resolved by the user-approved per-file override design: the "Nicht in Memories" timeline's purged entries now offer „Trotzdem neu sichern" (one-step confirm; the deliberate INV-2 exception, spec 065 S6 RE-AMENDED + spec 044 S3 amendment), which FORGETS the local shadow state rows via MediaStateRepository.forgetMediaStates (whole-row delete — a merely unshadowed UPLOADED row would be skipped by Reconciler rule 5; deletion makes rule 2 pend the file fresh) for ALL group copies (dupes share one entry), so the next backup scan re-uploads the file as a NEW server item — strictly local (no server call, no restore intent), the purged twin stays permanently_deleted in the Entfernungs-Verlauf and server dedupe never merges the fresh upload with it (findExistingOwnerMedia excludes permanently_deleted); trashed entries keep „Wiederherstellen" (server restore), local entries keep „Wieder sichern" [spec:065, S6] [feat/exclusions-purged-override]
[ISSUE-216] med open area:web Scoped-PR web Playwright suite fails/flakes reproducibly across unrelated PRs (observed on #1415 and #1417 with only favicon/app.html changes): backup-export "On" toggle assertion + discovery-domain Rename/Merge dialog tests hit 30 s timeouts (4 failed / 9 flaky / 527 passed on one rerun); not PR-caused (a rerun went green on #1415), suspected runner-parallelism or mock-server slowness — needs triage of the scoped test job [audit 2026-10-30: mitigations landed — PR#1444 (trace preservation on failure, nightly flake-detection retries=0/repeat-each=10, --shm-size=1g on all Playwright containers) + PR#1445 (hydration gates across the affected specs); the ⋮-menu timeout class is honestly attributed to runner starvation pending trace evidence from the next red run (traces now survive CI); stays open until attributed] [audit 2026-10-30: PR#1453 split discovery-review.spec.ts into four focused specs + discovery-review-shared.ts (describe name preserved) — split verified the rename/merge/dissolve suites green at retries=0 locally, 74/74; ⋮-menu timeout class still attributed to runner starvation pending trace attribution; shm + trace evidence in place since PR#1444; stays open]
Memories Wiki
Using Memories
- Accounts, the Library Hub, and Data Export
- Albums and Automatic Suggestions
- Explore and Discovery Domains
- The Media Viewer
- People and Faces
- Places and Map
- Plants (In Development)
- Datenschutzerklärung (Privacy Policy)
- Fast, Progressive Image Loading
- Search
- Sharing and Households
- Timeline and Photos
- Uploading and Import
Self-Hosting
- Backup and Restore
- CLI Password Reset
- Deployment Architecture
- Known Issues and Operator Diagnostics
- Map Tiles and Place Data Privacy
- Storage Layout
- Upgrades and Versioning
- Workers
Reference