fix(worker): scale slskd download backstop to album size (stop abandoning big albums)

The absolute download backstop was a flat _MAX_XFER_POLLS=1200 (~60min). A large
album (e.g. Drake "Scorpion", 521MB / 25 FLACs) from a slow-but-healthy serial
peer (~400KB/s, serving one file at a time) needs ~60-70min, so it was cut off
at 60min and — because an abandoned transfer is cancelled and re-enqueued from
scratch — re-downloaded from zero every attempt, never finishing and eventually
going needs_attention at the attempt cap.

Replace the flat cap with _max_polls(total_bytes): budget the backstop at a
conservative floor throughput (~64KB/s), clamped to [20min, 2h]. Scorpion now
gets the full 2h ceiling instead of 60min. The 90s stall timeout is unchanged,
so a truly dead/queued peer is still abandoned fast.

Follow-up (not in this change): resume across retries (skip files slskd already
completed) so an abandoned large download doesn't restart from zero.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

README.md

@@ -7,83 +7,54 @@ sources of differing quality: **Qobuz** (lossless/hi-res), **Soulseek** (P2P), a
 **ListenBrainz** and **Last.fm** (Spotify's recommendation/related-artists APIs were
 retired for new apps, so discovery is MetaBrainz-native and MBID-clean — no Spotify).
 
-## Status
+It runs on your own server as a small `docker compose` stack. This README covers getting
+it running; see **[docs/DEPLOYMENT.md](docs/DEPLOYMENT.md)** for the full feature tour,
+publishing pre-built images, and running behind a VPN.
 
-Built in three slices:
+## What it does
 
-1. **Acquisition engine** — ✅ **done**. Resolve a request against MusicBrainz,
-   match/rank across Qobuz/Soulseek/YouTube, download the best source into an
-   isolated staging dir, verify completeness, tag (mutagen), and atomically promote
-   the clean album into `Artist/Album (Year)/## Title.ext`.
-2. **Library manager** — ✅ **done**. Watched artists (live MB search + follow),
-   per-release monitor toggles + auto-monitor-future, a wanted list with retry +
-   quality-upgrade window, and a worker "monitor" that discovers new releases and
-   auto-grabs through the acquisition pipeline. Plus a **library scan** that ingests
-   existing on-disk albums as have + monitored + followed, and a per-artist
-   studio-only / show-all release-type filter.
-3. **Discovery** — ✅ **done**. Recommendation-driven finding of new music via
-   **ListenBrainz** similar-artists and **Last.fm** `artist.getSimilar` (both
-   MBID-native), behind a pluggable `SimilaritySource` interface — Last.fm registers
-   as a second source automatically once its API key is set, and cross-source scores
-   are summed per candidate. A background sweep (and a "Discover now" button)
-   aggregates artists similar to the ones you follow into a manual review feed of
-   suggested **artists** and **albums** at `/discover`; **Follow** and **Want**
-   reuse the slice-2 follow / wanted flows, **Dismiss** is permanent. Plus a
-   web-side seed-search box. Every suggestion and seed-search result opens a
-   read-only **artist preview** (`/discover/artist/[mbid]`) — discography,
-   per-album tracklists, and "in library" / "monitored" badges — where you can
-   Follow or Want before committing. No Spotify (its recommendation APIs are dead
-   for new apps).
+- **Acquisition** — resolve a request against MusicBrainz, match/rank across
+  Qobuz/Soulseek/YouTube, download the best source, verify completeness, tag, and organize
+  into `Artist/Album (Year)/## Title.ext`.
+- **Library manager** — follow artists (live MB search), per-release monitor toggles, a
+  wanted list with retry + quality-upgrade, a background monitor that auto-grabs new
+  releases, and a scan that ingests albums already on disk.
+- **Discovery** (`/discover`) — recommendation-driven suggestions from ListenBrainz +
+  Last.fm similar-artists, with Follow / Want / Dismiss and a read-only artist preview.
+- **Browse & watch** — Last.fm top artists/albums (`/lastfm`, with inline Follow/Want), a
+  cover-art **Library** grid (`/library`), and **The Floor** (home): a live queue of
+  in-flight acquisitions with three-phase progress and a Retry button.
 
-Alongside the three slices:
-
-- **Last.fm browse** (`/lastfm`) — your Last.fm top artists and albums (by period),
-  with a "hide items already in my library" toggle; each artist opens a modal of your
-  most-played songs and albums, and every album is one click to **Want**.
-- **Library** (`/library`) — a cover-art grid of everything on disk, filterable, each
-  album opening a tracklist modal.
-- **The Floor** (home) — a live queue of in-flight acquisitions with three-phase
-  progress (**Searching → Downloading → Finishing**), a live download percentage, a
-  per-job "pressing ticket" (source · format · tracks · elapsed), and a **Retry**
-  button on anything that needs attention.
-
-The whole UI runs a hand-rolled "Pressing Plant" design system (warm-paper theme, self-hosted
-Fraunces serif for prose vs. system mono for machine data, light/dark aware).
-
-**Operational notes:** the monitor and the discovery sweep are both **off by
-default** — set `monitor.enabled=true` (or Settings → Monitor, + optional `monitor.*`
-tuning) to activate auto-grab/upgrade, and `discover.enabled=true` (or
-Settings → Discovery, + optional `discover.*` tuning) to activate the scheduled
-discovery sweep. The `/discover` "Discover now" button and seed-search work with the
-sweep off. Credentials (Qobuz, Soulseek, Last.fm) are entered in Settings and persist
-across `docker compose up -d --build` rebuilds (don't use `down -v`; secrets are
-encrypted at rest with `LYRA_SECRET_KEY`). Tests run against a separate `lyra_test`
-database (a guard blocks the live DB). Deferred hardening + the full roadmap live in
-the implementation-plan "Notes" sections under
-[`docs/superpowers/plans/`](docs/superpowers/plans/).
+Full detail on each of these is in [docs/DEPLOYMENT.md → Features in depth](docs/DEPLOYMENT.md#features-in-depth).
 
 ## Architecture
 
-The `docker compose` stack is three services on a home server:
+The `docker compose` stack is a few services on one host:
 
-- **db** — Postgres, the shared source of truth.
+- **db** — Postgres, the shared source of truth (all durable state lives here).
 - **web** — the Next.js app: UI + API (search, request, queue, progress, settings).
-- **worker** — the Python worker: all source integration (streamrip, yt-dlp, slskd
-  client, mutagen; MusicBrainz + ListenBrainz + Last.fm for metadata and discovery),
-  running the six-stage acquisition pipeline plus the background monitor and discovery
-  sweeps.
+- **worker** — the Python worker: all source integration (streamrip, yt-dlp, slskd client,
+  mutagen; MusicBrainz + ListenBrainz + Last.fm for metadata and discovery), running the
+  acquisition pipeline plus the background monitor and discovery sweeps.
+- **db-backup** — a sidecar that `pg_dump`s the database on a schedule (see [Backup &
+  restore](#backup--restore)).
 
-Soulseek support talks to an **external slskd** daemon (an off-the-shelf headless
-Soulseek client) — the worker reaches it via the `slskd.url` + `slskd.api_key`
-credentials set in Settings, so run slskd separately rather than as part of this stack.
+Soulseek support talks to an **external slskd** daemon (an off-the-shelf headless Soulseek
+client) — the worker reaches it via the `slskd.url` + `slskd.api_key` credentials set in
+Settings, so run slskd separately rather than as part of this stack.
 
-A shared `/music` volume is the library, with a separate `/staging` volume for
-in-progress downloads (point `STAGING_DIR` at a fast local disk when `MUSIC_DIR` is a
-network share).
+A shared `/music` volume is the library, with a separate `/staging` volume for in-progress
+downloads (point `STAGING_DIR` at a fast local disk when `MUSIC_DIR` is a network share).
 
-See [`docs/superpowers/specs/`](docs/superpowers/specs/) for the full design.
+## Prerequisites
 
-## Running it
+- **Docker** + **Docker Compose**.
+- A directory for the music library (set `MUSIC_DIR`).
+- Optional, entered in **Settings** after first run: a **Qobuz** account (lossless/hi-res),
+  a running **slskd** daemon (Soulseek), and a **Last.fm** API key (browse + discovery).
+  None are required to start — Lyra runs with whatever you configure.
+
+## Quick start
 
 ```sh
 cp .env.example .env          # then edit: set LYRA_SECRET_KEY (openssl rand -base64 32) and MUSIC_DIR
@@ -97,10 +68,26 @@ the placeholder. Once set, keep it stable: changing it makes already-stored cred
 undecryptable (you'd re-enter them in Settings). Optionally set `LYRA_PASSWORD` too (see
 [Security](#security)).
 
-The web entrypoint runs `prisma migrate deploy` on every rebuild, so schema changes
-reach the live DB automatically. Enter Qobuz / Soulseek / Last.fm credentials in
-**Settings** (they persist across rebuilds), then follow an artist or request an album
-to feed The Floor.
+The web entrypoint runs `prisma migrate deploy` on every rebuild, so schema changes reach
+the live DB automatically. Enter Qobuz / Soulseek / Last.fm credentials in **Settings** (they
+persist across rebuilds), then follow an artist or request an album to feed The Floor.
+
+`.env.example` documents every setting (`MUSIC_DIR`, `STAGING_DIR`, `SLSKD_DOWNLOADS_DIR`,
+backup knobs). Note in particular: point `SLSKD_DOWNLOADS_DIR` at slskd's own downloads
+directory (Lyra must run on the same host as slskd for Soulseek delivery to work), and point
+`STAGING_DIR` at a fast local disk if `MUSIC_DIR` is a network share.
+
+## Operational notes
+
+- The **monitor** and the **discovery sweep** are both **off by default** — set
+  `monitor.enabled=true` (or Settings → Monitor, + optional `monitor.*` tuning) to activate
+  auto-grab/upgrade, and `discover.enabled=true` (or Settings → Discovery, + optional
+  `discover.*` tuning) to activate the scheduled discovery sweep. The `/discover` "Discover
+  now" button and seed-search work with the sweep off.
+- **Credentials persist** across `docker compose up -d --build` rebuilds — don't use
+  `down -v` (it wipes the Postgres volume); secrets are encrypted at rest with
+  `LYRA_SECRET_KEY`.
+- **Tests** run against a separate `lyra_test` database (a guard blocks the live DB).
 
 ## Managing the library
 
@@ -110,6 +97,24 @@ it — re-Want it from the artist page if you change your mind). This is why the
 also bind-mounts `MUSIC_DIR` at `/music` (read-write) — deletion is guarded to only ever
 remove folders strictly inside the library root.
 
+## Security
+
+Lyra has **no per-user accounts**. Two layers protect it:
+
+- **Shared-password gate.** Set `LYRA_PASSWORD` in `.env` to require a single shared
+  password for the entire UI and API. A request without a valid session cookie is
+  redirected to `/login` (or gets `401` for `/api/*`). The cookie is httpOnly and carries
+  an HMAC of the password keyed by `LYRA_SECRET_KEY`, so it can't be forged. **If
+  `LYRA_PASSWORD` is unset, the app is completely open** — anyone who can reach the host
+  can browse the library, trigger downloads, and read/overwrite the stored credentials in
+  Settings. Set it for any deployment that isn't on a fully trusted, isolated network.
+
+- **Network.** The shared password is deliberately minimal (single password, bearer
+  cookie, plain HTTP on the LAN). For anything internet-facing, also put Lyra behind a
+  reverse proxy with TLS and/or a VPN or auth gateway — don't rely on the shared password
+  alone. See [docs/DEPLOYMENT.md](docs/DEPLOYMENT.md#running-behind-a-vpn-gluetun) for a
+  Gluetun VPN setup.
+
 ## Backup & restore
 
 **All** durable state — library metadata, monitored/wanted lists, discovery data,
@@ -140,105 +145,8 @@ docker compose start web worker
 is safe. To restore onto a brand-new host, bring up just `db` first
 (`docker compose up -d db`), restore, then `up -d --build` the rest.
 
-## Security
+## More
 
-Lyra has **no per-user accounts**. Two layers protect it:
-
-- **Shared-password gate.** Set `LYRA_PASSWORD` in `.env` to require a single shared
-  password for the entire UI and API. A request without a valid session cookie is
-  redirected to `/login` (or gets `401` for `/api/*`). The cookie is httpOnly and carries
-  an HMAC of the password keyed by `LYRA_SECRET_KEY`, so it can't be forged. **If
-  `LYRA_PASSWORD` is unset, the app is completely open** — anyone who can reach the host
-  can browse the library, trigger downloads, and read/overwrite the stored credentials in
-  Settings. Set it for any deployment that isn't on a fully trusted, isolated network.
-
-- **Network.** The shared password is deliberately minimal (single password, bearer
-  cookie, plain HTTP on the LAN). For anything internet-facing, also put Lyra behind a
-  reverse proxy with TLS and/or a VPN or auth gateway — don't rely on the shared password
-  alone.
-
-## Running behind a VPN (Gluetun)
-
-Lyra runs fine with its traffic routed through a [Gluetun](https://github.com/qdm12/gluetun)
-VPN container — useful for hiding the download traffic (Soulseek / Qobuz / YouTube). Nothing
-in Lyra is VPN-incompatible; the only thing that changes is container networking, because a
-service using `network_mode: "service:gluetun"` gives up its own network stack and shares
-Gluetun's. Two consequences:
-
-- Containers that share Gluetun's namespace reach each other over **`localhost`**, not by
-  Docker service name — so `DATABASE_URL` changes from `@db:5432` to **`@localhost:5432`**.
-- **Published ports move to the `gluetun` container** (the individual services can no longer
-  publish their own).
-
-The simplest, most consistent setup is to put the **whole stack** behind Gluetun. Sketch
-(merge into your existing Gluetun compose; adjust the VPN provider block to yours):
-
-```yaml
-services:
-  gluetun:
-    image: qmcgaw/gluetun
-    cap_add: [NET_ADMIN]
-    environment:
-      # ... your VPN provider / credentials ...
-      FIREWALL_INPUT_PORTS: "3000"                 # let your browser reach the web UI
-      FIREWALL_OUTBOUND_SUBNETS: "192.168.0.0/16"  # your LAN — so the worker can reach slskd
-    ports:
-      - "8770:3000"   # Lyra web  (published on gluetun, not on `web`)
-      # - "5432:5432" # only if you want DB access from the LAN
-
-  web:
-    build: ./web
-    network_mode: "service:gluetun"     # no `ports:` / `networks:` here
-    environment:
-      DATABASE_URL: postgresql://lyra:lyra@localhost:5432/lyra   # `db` -> `localhost`
-      LYRA_SECRET_KEY: ${LYRA_SECRET_KEY}
-      LYRA_PASSWORD: ${LYRA_PASSWORD:-}
-      HOSTNAME: 0.0.0.0
-      LYRA_LIBRARY_ROOT: /music
-    volumes:
-      - ${MUSIC_DIR:-./music}:/music
-    depends_on:
-      gluetun: { condition: service_healthy }
-      db: { condition: service_healthy }
-
-  worker:
-    build: ./worker
-    network_mode: "service:gluetun"
-    environment:
-      DATABASE_URL: postgresql://lyra:lyra@localhost:5432/lyra
-      LYRA_SECRET_KEY: ${LYRA_SECRET_KEY}
-      STAGING_ROOT: /staging
-      SLSKD_DOWNLOADS_ROOT: /slskd-downloads
-    volumes:
-      - ${MUSIC_DIR:-./music}:/music
-      - ${STAGING_DIR:-${MUSIC_DIR:-./music}/.staging}:/staging
-      - ${SLSKD_DOWNLOADS_DIR:-./slskd-downloads}:/slskd-downloads
-    depends_on:
-      gluetun: { condition: service_healthy }
-      db: { condition: service_healthy }
-
-  db:
-    image: postgres:17
-    network_mode: "service:gluetun"     # shares the namespace so `localhost:5432` resolves
-    environment:
-      POSTGRES_USER: ${POSTGRES_USER}
-      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD}
-      POSTGRES_DB: ${POSTGRES_DB}
-    volumes:
-      - postgres-data:/var/lib/postgresql/data
-```
-
-(Give `db-backup` the same `network_mode` and `PGHOST: localhost`.) The `HOSTNAME=0.0.0.0`
-and the built-in healthchecks keep working, since everything is in one namespace.
-
-Caveats worth knowing:
-
-- **Qobuz is geo-sensitive** — its catalog/availability varies by country, so pick a VPN
-  exit in a region where your Qobuz account works, or some downloads will fail to match.
-- The worker's `net.ipv6.conf.all.disable_ipv6` sysctl **can't apply behind Gluetun** (a
-  shared netns can't set sysctls) — harmless, because Gluetun blocks IPv6 anyway, which is
-  exactly what that sysctl worked around (Qobuz's CDN resolving to an unroutable IPv6).
-- MusicBrainz rate-limits per IP; a shared VPN exit is marginally more likely to hit the
-  1 req/s ceiling, but the built-in response cache makes that a non-issue in normal use.
-- Bind mounts (`/music`, `/staging`, `/slskd-downloads`), the Postgres volume, and backups
-  are filesystem, not network — **completely unaffected** by the VPN.
+- **[docs/DEPLOYMENT.md](docs/DEPLOYMENT.md)** — full feature tour, publishing pre-built
+  images to a registry, and running the whole stack behind a Gluetun VPN.
+- **[docs/superpowers/specs/](docs/superpowers/specs/)** — the full design specs.

docs/DEPLOYMENT.md

@@ -0,0 +1,177 @@
+# Lyra — deployment & advanced guide
+
+Topics beyond the [README](../README.md) quick-start: the full feature tour, publishing
+pre-built images to a registry, and running the whole stack behind a VPN. For a basic
+self-hosted setup you only need the README — start there.
+
+- [Features in depth](#features-in-depth)
+- [Publishing pre-built images](#publishing-pre-built-images)
+- [Running behind a VPN (Gluetun)](#running-behind-a-vpn-gluetun)
+
+---
+
+## Features in depth
+
+Lyra was built in three slices, all complete:
+
+1. **Acquisition engine.** Resolve a request against MusicBrainz, match/rank across
+   Qobuz/Soulseek/YouTube, download the best source into an isolated staging dir, verify
+   completeness, tag (mutagen), and atomically promote the clean album into
+   `Artist/Album (Year)/## Title.ext`.
+2. **Library manager.** Watched artists (live MB search + follow), per-release monitor
+   toggles + auto-monitor-future, a wanted list with retry + quality-upgrade window, and a
+   worker "monitor" that discovers new releases and auto-grabs through the acquisition
+   pipeline. Plus a **library scan** that ingests existing on-disk albums as
+   have + monitored + followed, and a per-artist studio-only / show-all release-type filter.
+3. **Discovery.** Recommendation-driven finding of new music via **ListenBrainz**
+   similar-artists and **Last.fm** `artist.getSimilar` (both MBID-native), behind a
+   pluggable `SimilaritySource` interface — Last.fm registers as a second source
+   automatically once its API key is set, and cross-source scores are summed per candidate.
+   A background sweep (and a "Discover now" button) aggregates artists similar to the ones
+   you follow into a manual review feed of suggested **artists** and **albums** at
+   `/discover`; **Follow** and **Want** reuse the slice-2 follow / wanted flows, **Dismiss**
+   is permanent. Plus a web-side seed-search box. Every suggestion and seed-search result
+   opens a read-only **artist preview** (`/discover/artist/[mbid]`) — discography, per-album
+   tracklists, and "in library" / "monitored" badges — where you can Follow or Want before
+   committing. No Spotify (its recommendation APIs are dead for new apps).
+
+Alongside the three slices:
+
+- **Last.fm browse** (`/lastfm`) — your Last.fm top artists and albums (by period), with a
+  "hide items already in my library" toggle and inline **Follow** / **Want** buttons on each
+  row; each artist also opens a modal of your most-played songs and albums.
+- **Library** (`/library`) — a cover-art grid of everything on disk, filterable, each album
+  opening a tracklist modal.
+- **The Floor** (home) — a live queue of in-flight acquisitions with three-phase progress
+  (**Searching → Downloading → Finishing**), a live download percentage, a per-job "pressing
+  ticket" (source · format · tracks · elapsed), and a **Retry** button on anything that needs
+  attention.
+
+The whole UI runs a hand-rolled "Pressing Plant" design system (warm-paper theme, self-hosted
+Fraunces serif for prose vs. system mono for machine data, light/dark aware).
+
+Deferred hardening and the full roadmap live in the implementation-plan "Notes" sections under
+[`docs/superpowers/plans/`](superpowers/plans/); the full design specs are under
+[`docs/superpowers/specs/`](superpowers/specs/).
+
+---
+
+## Publishing pre-built images
+
+> **Maintainer release process** — this is how the project maintainer publishes images to a
+> private registry. If you're self-hosting your own copy, ignore this and just build from
+> source with `docker compose up -d --build` (see the README). You'd only need this if you
+> run your own registry and want servers to **pull** rather than build.
+
+The `web` and `worker` images can be published to a container registry and consumed by a
+compose file that pulls instead of builds. The maintainer publishes to a Gitea container
+registry consumed by a compose file in the
+[`vm-download`](https://git.jger.nl/Jonathan/vm-download) ops repo (`docker/lyra/`).
+
+Publish a new release from this repo (amd64):
+
+```sh
+docker login <your-registry>       # once — registry user + a token with package:write
+./scripts/build-push.sh            # builds + pushes lyra-web / lyra-worker (:latest and :<sha>)
+```
+
+`scripts/build-push.sh` defaults the registry namespace to `git.jger.nl/jonathan`; override it
+with `LYRA_REGISTRY=<your-registry>/<namespace>`.
+
+Then on the server:
+
+```sh
+cd /opt/git/vm-download && git pull
+cd docker/lyra && docker compose pull && docker compose up -d
+```
+
+`db` (Postgres) and `db-backup` use the stock `postgres:17` image, so only the two custom
+images are built here.
+
+---
+
+## Running behind a VPN (Gluetun)
+
+Lyra runs fine with its traffic routed through a [Gluetun](https://github.com/qdm12/gluetun)
+VPN container — useful for hiding the download traffic (Soulseek / Qobuz / YouTube). Nothing
+in Lyra is VPN-incompatible; the only thing that changes is container networking, because a
+service using `network_mode: "service:gluetun"` gives up its own network stack and shares
+Gluetun's. Two consequences:
+
+- Containers that share Gluetun's namespace reach each other over **`localhost`**, not by
+  Docker service name — so `DATABASE_URL` changes from `@db:5432` to **`@localhost:5432`**.
+- **Published ports move to the `gluetun` container** (the individual services can no longer
+  publish their own).
+
+The simplest, most consistent setup is to put the **whole stack** behind Gluetun. Sketch
+(merge into your existing Gluetun compose; adjust the VPN provider block to yours):
+
+```yaml
+services:
+  gluetun:
+    image: qmcgaw/gluetun
+    cap_add: [NET_ADMIN]
+    environment:
+      # ... your VPN provider / credentials ...
+      FIREWALL_INPUT_PORTS: "3000"                 # let your browser reach the web UI
+      FIREWALL_OUTBOUND_SUBNETS: "192.168.0.0/16"  # your LAN — so the worker can reach slskd
+    ports:
+      - "8770:3000"   # Lyra web  (published on gluetun, not on `web`)
+      # - "5432:5432" # only if you want DB access from the LAN
+
+  web:
+    build: ./web
+    network_mode: "service:gluetun"     # no `ports:` / `networks:` here
+    environment:
+      DATABASE_URL: postgresql://lyra:lyra@localhost:5432/lyra   # `db` -> `localhost`
+      LYRA_SECRET_KEY: ${LYRA_SECRET_KEY}
+      LYRA_PASSWORD: ${LYRA_PASSWORD:-}
+      HOSTNAME: 0.0.0.0
+      LYRA_LIBRARY_ROOT: /music
+    volumes:
+      - ${MUSIC_DIR:-./music}:/music
+    depends_on:
+      gluetun: { condition: service_healthy }
+      db: { condition: service_healthy }
+
+  worker:
+    build: ./worker
+    network_mode: "service:gluetun"
+    environment:
+      DATABASE_URL: postgresql://lyra:lyra@localhost:5432/lyra
+      LYRA_SECRET_KEY: ${LYRA_SECRET_KEY}
+      STAGING_ROOT: /staging
+      SLSKD_DOWNLOADS_ROOT: /slskd-downloads
+    volumes:
+      - ${MUSIC_DIR:-./music}:/music
+      - ${STAGING_DIR:-${MUSIC_DIR:-./music}/.staging}:/staging
+      - ${SLSKD_DOWNLOADS_DIR:-./slskd-downloads}:/slskd-downloads
+    depends_on:
+      gluetun: { condition: service_healthy }
+      db: { condition: service_healthy }
+
+  db:
+    image: postgres:17
+    network_mode: "service:gluetun"     # shares the namespace so `localhost:5432` resolves
+    environment:
+      POSTGRES_USER: ${POSTGRES_USER}
+      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD}
+      POSTGRES_DB: ${POSTGRES_DB}
+    volumes:
+      - postgres-data:/var/lib/postgresql/data
+```
+
+(Give `db-backup` the same `network_mode` and `PGHOST: localhost`.) The `HOSTNAME=0.0.0.0`
+and the built-in healthchecks keep working, since everything is in one namespace.
+
+Caveats worth knowing:
+
+- **Qobuz is geo-sensitive** — its catalog/availability varies by country, so pick a VPN
+  exit in a region where your Qobuz account works, or some downloads will fail to match.
+- The worker's `net.ipv6.conf.all.disable_ipv6` sysctl **can't apply behind Gluetun** (a
+  shared netns can't set sysctls) — harmless, because Gluetun blocks IPv6 anyway, which is
+  exactly what that sysctl worked around (Qobuz's CDN resolving to an unroutable IPv6).
+- MusicBrainz rate-limits per IP; a shared VPN exit is marginally more likely to hit the
+  1 req/s ceiling, but the built-in response cache makes that a non-issue in normal use.
+- Bind mounts (`/music`, `/staging`, `/slskd-downloads`), the Postgres volume, and backups
+  are filesystem, not network — **completely unaffected** by the VPN.

docs/superpowers/plans/2026-07-15-post-deploy-ux-fixes.md

@@ -0,0 +1,96 @@
+# Post-deploy UX fixes (3 items) — handoff plan
+
+> **Context:** Lyra is DEPLOYED + LIVE on the media VM (ssh alias `download` = jonathan@10.10.0.103,
+> hostname `servarr`, x86_64). Pre-built images from `git.jger.nl/jonathan/lyra-{web,worker}`, whole
+> stack behind its own gluetun VPN, deploy repo `Jonathan/vm-download` → `docker/lyra/`. Library was
+> wiped to a blank slate; user rebuilds via Last.fm. Two test presses succeeded end-to-end (John Mayer
+> Continuum q3 hi-res, Room for Squares q2) — pipeline is solid. These 3 are UX polish the user hit
+> in real use. See MEMORY.md → project-state.md "VM DEPLOYMENT" for full deploy details.
+
+**Workflow for EACH fix:** implement in `worker/` → run worker tests (`cd worker && source .venv/bin/activate
+&& python -m pytest`) → **on the DEV box** `./scripts/build-push.sh` (needs `docker login git.jger.nl`
+already done this session; a fresh session may need re-login — a **real TTY**, not the `!`-prefix) →
+on VM `ssh download 'cd /opt/git/vm-download/docker/lyra && docker compose pull && docker compose up -d worker'`
+→ verify with a test press (queue via SQL, see below) and watch `Job.downloadProgress`/on-disk result.
+Commit per fix; branch `feat/post-deploy-ux`, ff-merge to main, push.
+
+**Queue a test press on the VM** (mimics the Wanted "Search now"): pick a `MonitoredRelease` id from
+`SELECT id,"artistName",album FROM "MonitoredRelease" WHERE monitored AND NOT ignored AND state<>'fulfilled'`,
+then insert a Request+Job (use a `/tmp/*.sql` file + `docker cp` + `psql -f` to avoid ssh quoting hell):
+`INSERT INTO "Request"(id,artist,album,"monitoredReleaseId",status,"createdAt") SELECT gen_random_uuid()::text,"artistName",album,id,'pending',now() FROM "MonitoredRelease" WHERE id='<mrid>' RETURNING id` → capture into a Job insert `(id,"requestId",state,"currentStage",attempts,"downloadProgress","createdAt","updatedAt") VALUES (gen_random_uuid()::text,rid,'requested','intake',0,0,now(),now())`.
+
+---
+
+## Fix 1 — Smooth Qobuz download progress (byte-level)
+
+**Problem:** the Floor progress bar sits at 0% for the whole Qobuz download then jumps to 100%.
+Qobuz/streamrip's per-track byte callback is NOT hooked, so `Job.downloadProgress` comes only from the
+file-count poller (`pipeline._poll_download_progress`) — and streamrip downloads tracks concurrently, so
+0 files are "complete" until they all land at once. yt-dlp + slskd already report bytes via `on_progress`
+(smooth); only Qobuz doesn't.
+
+**Files:** `worker/lyra_worker/adapters/_streamrip.py` (the Qobuz download), `worker/lyra_worker/pipeline.py`
+(threads `on_progress` → `_make_on_progress` → `Job.downloadProgress`; poller is the fallback when the
+per-attempt `reports` Event is set).
+
+**Approach:** hook streamrip's progress. streamrip drives progress via a callback/console; investigate its
+API (it uses a `rip.utils`/`Downloadable` byte-progress hook or a `tqdm` callback). Aggregate downloaded
+bytes / total bytes summed across the album's concurrent tracks → a single fraction → call the pipeline's
+`on_progress(frac)` (which sets `reports` and writes throttled `Job.downloadProgress`). Keep the file-count
+poller as the fallback (adapters that don't report). Match the throttle already in `_make_on_progress`
+(≥1% move or every 0.5s).
+
+**Test:** hard to unit-test streamrip internals — add a unit test that a fake byte-callback aggregates to a
+monotonic 0→1 fraction, plus **live-verify** with a test press: `Job.downloadProgress` climbs mid-download
+instead of 0→100. (Known caveat if streamrip exposes no usable byte hook: fall back to a smoother
+file-count cadence, or accept the jump — document the finding.)
+
+## Fix 2 — Parallelize the source searches (Soulseek adds ~2 min)
+
+**Problem:** every acquisition spends up to ~2 min in the Soulseek search. `SlskdClient.search_album` polls
+`_SEARCH_POLLS=40 × _POLL_SECONDS=3s = 120s` (early-exits on `isComplete`, but slskd searches genuinely run
+long). The pipeline searches adapters SEQUENTIALLY (`for adapter in adapters: adapter.search(target)`), so
+Soulseek's 2 min is added on top of Qobuz's ~instant search.
+
+**Files:** `worker/lyra_worker/pipeline.py` (the match phase, `for adapter in adapters` loop);
+optionally `worker/lyra_worker/adapters/_slskd.py` (`_SEARCH_POLLS`).
+
+**Approach:** run the per-adapter searches **concurrently** (e.g. `ThreadPoolExecutor`, one thread per
+adapter), collect all candidates when the last returns → total match wall-clock = slowest single search,
+not the sum. So Qobuz returns instantly and only Soulseek's ~2 min gates. Keep the existing per-adapter
+try/except (a failing source contributes nothing). **Thread-safety caveat:** streamrip uses a module-level
+persistent asyncio loop (`_streamrip._run`) — verify concurrent Qobuz search from a worker thread is safe
+(it runs coroutines on that shared loop; may need a lock or to keep Qobuz on the calling thread while
+Soulseek/YouTube go to threads). Optionally also drop `_SEARCH_POLLS` to ~20 (60s) with the early-exit.
+
+**Test:** a pipeline test asserting all adapters' candidates are collected (order-independent) when searched
+concurrently; existing match/ranking tests stay green. Live-verify a press: match phase finishes in ~the
+Soulseek time, not Soulseek+Qobuz.
+
+## Fix 3 — Canonicalize bonus-track filenames on import
+
+**Problem:** tracks BEYOND the MusicBrainz tracklist keep streamrip's raw name. Real example: Room for
+Squares imported 12 clean `## Title.flac` + one `13. John Mayer - St. Patrick's Day (Album Version).flac`
+(MB lists 12; Qobuz had a 13th bonus). The main tracklist gets `## Title` but extras don't.
+
+**Files:** `worker/lyra_worker/library.py` (`import_album` — how it names/moves audio into
+`Artist/Album (Year)/## Title.ext`), and/or the tagger `worker/lyra_worker/_mutagen.py`.
+
+**Approach:** find where files are renamed to `## Title.ext`. For a track with no MB position/title (an
+extra), derive a clean name from its own embedded tags (track number + title) or by stripping the
+`^\d+\.?\s*(Artist\s*-\s*)?` prefix from the streamrip filename → `## Title.ext`. Ensure the position keeps
+counting past the MB list (13, 14, …) and the on-disk `trackNames` capture stays consistent.
+
+**Test:** unit-test the filename-canonicalizer on inputs like `13. John Mayer - St. Patrick's Day (Album
+Version).flac` → `13 St. Patrick's Day (Album Version).flac`; live-verify by re-pressing an album with a
+bonus track (or `docker compose exec worker` a rename dry-run). NOTE: the already-imported Room for Squares
+track 13 can be fixed by re-pressing (force upgrade) or a one-off rename — mention to the user.
+
+---
+
+## Self-review / notes
+- All three are worker-only → only the `worker` image rebuilds/redeploys (`docker compose up -d worker`).
+- Don't disrupt the live stack: the VM is in use; deploy the worker, verify one press, done.
+- If context/time is tight, do them in priority order: **2 (search speed)** and **3 (filenames)** are
+  self-contained and low-risk; **1 (streamrip progress)** is the deepest (streamrip internals) — timebox it
+  and fall back to documenting if streamrip has no clean byte hook.

docs/superpowers/specs/2026-07-15-press-controls-design.md

@@ -0,0 +1,179 @@
+# Press controls + inter-job delay — design
+
+Date: 2026-07-15
+
+## Problem
+
+The monitor can enqueue a large backlog of acquisition jobs at once. The worker drains
+them serially with **no pacing** (only a 2s idle poll when the queue is empty) and offers
+**no runtime control** over the queue — the only per-job action on the Floor is Retry (on
+`needs_attention`). The user wants:
+
+1. A **configurable delay between downloads** to spread sustained Qobuz load over time.
+2. **Per-item controls** on the Floor ("the Press"): pause and remove a queued item.
+3. **Whole-press controls**: pause/resume, clear the queue, and a force-stop of the
+   currently-downloading album.
+
+## Constraints (from the existing code)
+
+- The worker is a **single, strictly-serial process** (`worker/lyra_worker/main.py:256-298`).
+  One job runs to completion in a blocking `run_pipeline` call before the loop does anything
+  else.
+- **Config is read fresh every loop iteration** (`main.py:262`, `get_config`), so any Config
+  change takes effect within ~2s with no restart.
+- An **in-flight download cannot be cleanly interrupted** — `run_pipeline` and the adapter
+  download are synchronous/blocking with only an `on_progress` callback, no cancel token. The
+  only interruption is process death, after which `reclaim_stuck_jobs` (`claim.py:8-24`)
+  requeues the in-flight job at startup and `clear_staging_root` (`main.py:233`) wipes the
+  partial download.
+- `claim_next` (`claim.py:27-57`) claims the oldest `state='requested'` job.
+- Deleting a `Request` cascades to its `Job` and `Candidate` rows (`schema.prisma:51,66`); the
+  linked `MonitoredRelease` is `SetNull`, not cascaded.
+- `enqueue_due` (`monitor.py:115-147`) re-enqueues a monitored release once it has no live
+  (non-terminal) job and its `lastSearchedAt` is older than `retryIntervalHours` (default 6).
+  So a removed/cleared monitored item reappears later on the normal cadence unless `ignored`.
+- The web container runs `prisma migrate deploy` on startup (`web/docker-entrypoint.sh`), so a
+  new migration auto-applies on deploy.
+
+## Decisions (confirmed with the user)
+
+- **Active item:** add a global **force-kill** ("Stop now") that aborts the in-flight download
+  via worker process-restart. Per-item controls act only on queued items.
+- **Remove / Clear semantics:** **cancel-for-now**. Manual one-off requests are gone for good;
+  monitored/wanted releases stay wanted and may reappear on the monitor's normal cadence.
+  Permanent skip remains the existing per-release **Ignore**.
+- **Pause scope:** **downloads only**. Pausing halts claiming/downloading; the monitor keeps
+  finding and enqueueing releases (they pile up in the queue) and Resume drains them.
+
+## Data model
+
+- **New column** `Job.paused Boolean @default(false)` — a Prisma migration
+  (`add_job_paused`). Auto-applies on deploy.
+- **Config keys** (key/value `Config` rows, no migration):
+  - `floor.paused` — `"true"`/`"false"`; global press pause.
+  - `floor.jobDelaySeconds` — integer string, default `"0"`; delay between downloads.
+  - `floor.killRequested` — `"true"`/`"false"`; one-shot force-kill signal, cleared by the
+    worker on consumption.
+
+## Worker changes
+
+`worker/lyra_worker/main.py` and `worker/lyra_worker/claim.py`.
+
+1. **Inter-job delay.** After `finish_job` (`main.py:297`), read `floor.jobDelaySeconds` from
+   the already-fresh `config` and sleep that many seconds before the next loop iteration.
+   Sleep in small (~2s) chunks so the loop stays responsive; the force-kill watcher runs on a
+   separate thread and works regardless. Parse with the existing `_int`-style guard (invalid →
+   0). A value of 0 preserves today's back-to-back behavior.
+
+2. **Global pause.** Immediately before `claim_next` (`main.py:286`), if `floor.paused` is
+   truthy (`_TRUE` idiom), skip claiming: `time.sleep(IDLE_SLEEP); continue`. Because the loop
+   is serial, any in-flight job has already finished before this check — pause stops the
+   *next* download. Monitor sweep / scan / discovery (above the claim in the loop body)
+   continue to run, so the queue keeps filling while paused.
+
+3. **Per-item pause.** `claim_next` `WHERE` gains `AND NOT "paused"`. A paused queued job is
+   skipped by the claim but remains `state='requested'`, so it is not "lost" and resumes
+   claiming as soon as `paused` is cleared. `reclaim_stuck_jobs` is unaffected (it only touches
+   in-flight states).
+
+4. **Force-kill watcher.** A daemon thread started in `run_forever` (modeled on
+   `_liveness_heartbeat`, `main.py:216-226`) polls `floor.killRequested` on its own DB
+   connection every ~1.5s. When set, it clears the flag (commit) and calls `os._exit(0)`. The
+   container restart policy brings the worker back; startup `clear_staging_root` +
+   `reclaim_stuck_jobs` wipe the partial download and requeue the aborted job. The "Stop now"
+   action also sets `floor.paused=true`, so the worker returns **idle** rather than instantly
+   re-downloading the aborted album; the user resumes when ready.
+
+## Infrastructure (ops repo)
+
+The prod worker's Docker restart policy is currently **`no`** (`docker inspect` →
+`RestartPolicy=no`), so process death — from force-kill *or any crash* — leaves it down. The
+force-kill design requires `restart: unless-stopped` on the `worker` service in the ops repo
+(`/opt/git/vm-download/docker/lyra/compose.yaml`). The in-repo `docker-compose.yml` already has
+this (`docker-compose.yml:53`); this change brings prod in line and is good hygiene
+independent of this feature. This is a one-line change in a separate repo and must be applied
+before the force-kill feature is usable in prod.
+
+## API (web) — follows existing `retry` / `library DELETE` patterns
+
+- **`GET /api/floor`** → `{ paused: boolean, jobDelaySeconds: number }` (reads the Config
+  rows). Used by the Floor to render the paused banner and control states.
+- **`POST /api/floor`** with `{ action }`:
+  - `"pause"` / `"resume"` → upsert `floor.paused`.
+  - `"clear"` → delete every `Request` whose `Job.state = 'requested'` (queued, not yet
+    claimed). Cascade removes the Jobs and Candidates. The in-flight and terminal jobs are
+    untouched. Returns the number cleared.
+  - `"stop"` → upsert `floor.paused='true'` and `floor.killRequested='true'`.
+- **`POST /api/requests/[id]/pause`** with `{ paused: boolean }` → set `Job.paused`. Guard:
+  the job must be `state='requested'` (else `400`, matching the existing retry route), since
+  pausing an in-flight job has no effect.
+- **`DELETE /api/requests/[id]`** → delete the `Request` (cascade Job + Candidates). Guard:
+  only when `Job.state = 'requested'`, so the active download is never touched (that is the
+  job of `Stop now`). Mirrors `DELETE /api/library/[id]`.
+
+## Web UI
+
+`web/src/app/queue.tsx` and `web/src/app/_ui/job-row.tsx`; styling via existing `.btn`
+variants in `design.css`.
+
+- **Press header** (above the queue list):
+  - `Pause the press` / `Resume` toggle (from `GET /api/floor`).
+  - `Clear queue` — confirm dialog showing the queued count; calls `POST /api/floor
+    {action:"clear"}`.
+  - `Stop now ⚠` — rendered only when a job is actively downloading (derivable from the
+    existing `/api/requests` data: any job in `matching`/`matched`/`downloading`/`tagging`).
+    Confirm dialog ("aborts the current download and pauses the press"); calls `POST /api/floor
+    {action:"stop"}`.
+  - A "Press paused — N queued" banner/chip when `paused`.
+- **Per-item controls** (queued rows only — `job.state === 'requested'`), placed in the row
+  `note` slot as `.btn sm ghost`:
+  - `Pause` / `Resume` → `POST /api/requests/[id]/pause`.
+  - `Remove` → `DELETE /api/requests/[id]`.
+  - A per-item paused row shows a subtle "paused" chip and the `Resume` button.
+- Active-download and `needs_attention` rows keep their current rendering (Retry unchanged).
+- Polling cadence stays 2s; `GET /api/floor` is fetched alongside the existing three calls in
+  `refresh()`.
+
+## Settings
+
+Add a numeric field **"Delay between downloads (seconds)"** (default 0, min 0) to the settings
+form (`web/src/app/settings/settings-form.tsx`), persisted to `floor.jobDelaySeconds` via the
+existing settings config-write path (the same mechanism used for the other numeric/monitor
+settings). Include short helptext explaining it paces sustained Qobuz load.
+
+## Testing
+
+- **Worker** (pytest, `worker/tests/`):
+  - `claim_next` skips `paused=true` queued jobs and still claims the next unpaused one.
+  - Global pause: the loop skips claiming when `floor.paused` is set (unit-test the claim-gate
+    helper).
+  - Delay: the delay value is read and applied (unit-test the parse/gate helper; keep the
+    sleep injectable/mockable).
+  - Kill watcher: when `floor.killRequested` is set, the watcher clears the flag and triggers
+    exit (mock `os._exit`, assert flag cleared + exit called).
+- **Web** (route tests, mirroring `requests/[id]/retry/route.test.ts` and
+  `library/bulk/route.test.ts`):
+  - `POST /api/floor` for each action; `clear` deletes only `requested` jobs; `stop` sets both
+    flags.
+  - `POST /api/requests/[id]/pause` toggles `Job.paused`; rejects non-`requested`.
+  - `DELETE /api/requests/[id]` removes a queued Request; refuses an in-flight/terminal job.
+  - `GET /api/floor` returns the Config-derived state.
+
+## Edge cases
+
+- **Remove/Clear of monitored items** reappear later by design (cancel-for-now). No extra
+  handling; the per-release Ignore is the permanent path.
+- **Kill with nothing downloading** still restarts the worker (brief blip). Minimized by only
+  showing `Stop now` while a download is active.
+- **Delay during pause:** the delay sleeps only *between* jobs; pause is checked before each
+  claim, so a paused press never starts a delayed job.
+- **Race:** clicking per-item Pause/Remove on a job that was just claimed — the guards
+  (`state='requested'`) reject it; the UI only renders those buttons on queued rows, and the
+  2s refresh reconciles.
+
+## Out of scope
+
+- Resuming a partially-downloaded album (no download-resume support; force-killed jobs
+  re-download from scratch).
+- Reordering the queue / per-item priority.
+- Any change to the per-release Ignore or the wanted list UI.

docs/superpowers/specs/2026-07-15-qobuz-pacing-design.md

@@ -0,0 +1,83 @@
+# Qobuz pacing ("budget gate") — design
+
+Date: 2026-07-15
+
+## Problem
+
+Bulk hi-res downloading got the Qobuz account blocked (`403 USER_BLOCKED`). We need to keep
+using Qobuz for hi-res but at a volume and cadence that looks like a human listener, so the
+account isn't flagged again — while a large wanted-list still drains promptly via other sources.
+
+## Approach: a per-job Qobuz eligibility gate
+
+One gate decides, **per album**, whether Qobuz is an eligible source *right now*. If the gate is
+closed, Qobuz candidates are dropped before ranking, so the worker falls through to the best
+non-Qobuz source (Soulseek/YouTube) and keeps moving — no idling. Monitored albums grabbed that
+way are upgraded to Qobuz hi-res later, automatically, via the existing quality-upgrade window,
+once Qobuz has budget again.
+
+**Qobuz is eligible for a job only if ALL hold:**
+1. **Active hours** — current hour ∈ [activeStart, activeEnd) (a daytime window; start < end).
+2. **Under today's cap** — Qobuz downloads counted today < today's cap.
+3. **Spacing elapsed** — now ≥ the randomized "next allowed" time set after the last Qobuz
+   download.
+
+After each successful Qobuz download the worker bumps today's counter and sets the next-allowed
+time. The counter resets at the day boundary (tracked by a stored date).
+
+**Today's cap = warm-up ramp of the steady cap**, by account age (days since `warmupStartDate`):
+
+| Account age (days) | Cap |
+|---|---|
+| 0–2 | min(5, steady) |
+| 3–6 | min(10, steady) |
+| 7–13 | min(20, steady) |
+| 14+ | steady |
+
+If `warmupStartDate` is unset, no ramp (steady cap applies).
+
+**Spacing / jitter (combined):** the gap to the next Qobuz download is randomized around
+`activeWindowSeconds / todayCap` (so the day's budget spreads evenly across active hours), ±30%,
+with a hard floor (≥ 60 s). This *is* the jitter — never a fixed metronome — and it prevents
+front-loading the whole daily cap into the first hour.
+
+## Configuration (Config keys; editable in Settings → Qobuz → Pacing)
+
+Settings:
+- `qobuz.pacing.enabled` (default `true`)
+- `qobuz.pacing.activeStartHour` (default `8`)
+- `qobuz.pacing.activeEndHour` (default `23`)
+- `qobuz.pacing.dailyCap` (steady cap, default `40`)
+- `qobuz.pacing.warmupStartDate` (ISO date; set to the new account's start day)
+
+Worker-managed state (not user-edited):
+- `qobuz.pacing.countDay` (date the counter is for)
+- `qobuz.pacing.countToday` (int)
+- `qobuz.pacing.nextAllowedAt` (epoch seconds)
+
+Current hour / day come from the **DB clock** (`now()`), avoiding container-timezone ambiguity.
+Config is read fresh each worker loop, so changes take effect with no restart.
+
+## Worker integration
+
+- New module `worker/lyra_worker/qobuz_gate.py`: a `QobuzPacing` config dataclass
+  (`from_config`), a pure `today_cap(...)`, a pure `is_allowed(...)` decision, and thin
+  DB helpers `gate_open(conn, now)` (reads state + decides) and `record_download(conn, pacing,
+  now)` (bumps counter + sets next-allowed).
+- `pipeline.py`: after searching sources, if `gate_open` is false AND at least one non-Qobuz
+  candidate exists, drop Qobuz candidates before ranking (fall-through). If Qobuz is the *only*
+  source available, keep it (a Qobuz-only setup isn't gated — holding would strand the album).
+- After a successful download where `winner.source == "qobuz"`, call `record_download`.
+
+## Testing
+
+Pure-logic unit tests for `is_allowed` (each gate condition), `today_cap` (ramp boundaries), and
+the spacing computation; pipeline tests that a gated Qobuz candidate is dropped in favour of a
+non-Qobuz source, that an open gate keeps Qobuz, and that a successful Qobuz download records the
+counter/next-allowed.
+
+## Out of scope
+
+- Holding/parking an album for later Qobuz when Qobuz is the only source (fails/last-source as
+  today). Multi-source setups fall through, which is the common case.
+- Timezone-aware active-hours windows that wrap past midnight (start < end only).

scripts/build-push.sh

@@ -0,0 +1,29 @@
+#!/usr/bin/env bash
+# Build the Lyra web + worker images and push them to the Gitea container registry, so the VM
+# can pull pre-built images instead of building from source.
+#
+# Usage (from the Lyra repo root):
+#   docker login git.jger.nl        # once — Gitea username + a token with package:write
+#   ./scripts/build-push.sh         # builds + pushes :latest and :<short-sha>
+#
+# Override the registry namespace with LYRA_REGISTRY (default git.jger.nl/jonathan).
+set -euo pipefail
+
+REGISTRY="${LYRA_REGISTRY:-git.jger.nl/jonathan}"
+here="$(cd "$(dirname "$0")/.." && pwd)"
+sha="$(git -C "$here" rev-parse --short HEAD)"
+
+for svc in web worker; do
+  img="$REGISTRY/lyra-$svc"
+  echo "==> building $img ($sha)"
+  docker build -t "$img:latest" -t "$img:$sha" "$here/$svc"
+done
+
+for svc in web worker; do
+  img="$REGISTRY/lyra-$svc"
+  echo "==> pushing $img:latest and $img:$sha"
+  docker push "$img:latest"
+  docker push "$img:$sha"
+done
+
+echo "Done. Pulled on the VM with: docker compose pull && docker compose up -d"

web/prisma/migrations/20260715000000_add_job_paused/migration.sql

@@ -0,0 +1,2 @@
+-- AlterTable
+ALTER TABLE "Job" ADD COLUMN     "paused" BOOLEAN NOT NULL DEFAULT false;

web/prisma/migrations/20260720200000_add_job_download_eta/migration.sql

@@ -0,0 +1,2 @@
+-- Estimated seconds remaining for the active download (measured throughput). Nullable.
+ALTER TABLE "Job" ADD COLUMN "downloadEtaSeconds" INTEGER;

web/prisma/schema.prisma

@@ -58,6 +58,10 @@ model Job {
   createdAt        DateTime    @default(now())
   updatedAt        DateTime    @updatedAt
   downloadProgress Float       @default(0)
+  // Estimated seconds until the current download finishes, from measured throughput (slskd byte
+  // transfer rate). Null when unknown (no speed sample yet, or a source that doesn't report bytes).
+  downloadEtaSeconds Int?
+  paused           Boolean     @default(false)
   candidates       Candidate[]
 }
 

web/src/app/_ui/artist-picker.tsx

@@ -0,0 +1,49 @@
+"use client";
+
+import { Modal } from "./modal";
+import type { ArtistHit } from "@/lib/artist-resolve";
+
+/** Disambiguation picker: several MusicBrainz artists share a name (e.g. two bands both called
+ *  "Looking Glass"), so let the user choose the right one by its MB disambiguation text instead
+ *  of silently trusting Last.fm's — sometimes wrong — MBID. Candidates arrive in MB relevance
+ *  order, so the first is flagged "Suggested". */
+export function ArtistPicker({
+  open,
+  name,
+  candidates,
+  onPick,
+  onClose,
+}: {
+  open: boolean;
+  name: string;
+  candidates: ArtistHit[];
+  onPick: (mbid: string, name: string) => void;
+  onClose: () => void;
+}) {
+  return (
+    <Modal open={open} onClose={onClose} title="Which artist?">
+      <p className="muted-note">
+        More than one artist is named “{name}”. Pick the right one.
+      </p>
+      <ul className="list">
+        {candidates.map((c, i) => (
+          <li key={c.mbid} className="list-row">
+            <div className="main">
+              <div className="rtitle">
+                <button className="linkish" onClick={() => onPick(c.mbid, c.name)}>
+                  {c.name}
+                </button>
+              </div>
+              <div className="rmeta">
+                <span>{c.disambiguation || "no description on MusicBrainz"}</span>
+              </div>
+            </div>
+            <div className="actions">
+              {i === 0 ? <span className="chip working">Suggested</span> : null}
+            </div>
+          </li>
+        ))}
+      </ul>
+    </Modal>
+  );
+}

web/src/app/_ui/job-row.tsx

@@ -14,6 +14,7 @@ export function JobRow({
   album,
   kind,
   label,
+  marker,
   meta,
   note,
   bar,
@@ -22,6 +23,7 @@ export function JobRow({
   album: string;
   kind: Kind;
   label: string;
+  marker?: string;
   meta?: ReactNode;
   note?: ReactNode;
   bar?: ReactNode;
@@ -31,6 +33,7 @@ export function JobRow({
       <div className="stripe" />
       <div>
         <h3 className="title">
+          {marker ? <span className="nextup">{marker}</span> : null}
           {album} <span className="artist">· {artist}</span>
         </h3>
         {meta ? <div className="meta">{meta}</div> : null}

web/src/app/_ui/kebab-menu.tsx

@@ -0,0 +1,47 @@
+"use client";
+
+import { useEffect, useRef, useState, type ReactNode } from "react";
+
+/** A vertical "⋮" trigger that opens a small popup menu. Closes on outside click, Escape, or any
+ *  click inside the popup (so menu items don't each need to manage it). Children are the items —
+ *  use <button className="menu-item">…</button>. */
+export function KebabMenu({ label = "Actions", children }: { label?: string; children: ReactNode }) {
+  const [open, setOpen] = useState(false);
+  const ref = useRef<HTMLDivElement>(null);
+
+  useEffect(() => {
+    if (!open) return;
+    function onDoc(e: MouseEvent) {
+      if (ref.current && !ref.current.contains(e.target as Node)) setOpen(false);
+    }
+    function onKey(e: KeyboardEvent) {
+      if (e.key === "Escape") setOpen(false);
+    }
+    document.addEventListener("mousedown", onDoc);
+    document.addEventListener("keydown", onKey);
+    return () => {
+      document.removeEventListener("mousedown", onDoc);
+      document.removeEventListener("keydown", onKey);
+    };
+  }, [open]);
+
+  return (
+    <div className="menu" ref={ref}>
+      <button
+        type="button"
+        className="menu-trigger"
+        aria-haspopup="menu"
+        aria-expanded={open}
+        aria-label={label}
+        onClick={() => setOpen((o) => !o)}
+      >
+        ⋮
+      </button>
+      {open ? (
+        <div className="menu-pop" role="menu" onClick={() => setOpen(false)}>
+          {children}
+        </div>
+      ) : null}
+    </div>
+  );
+}

web/src/app/_ui/status.test.ts

@@ -1,5 +1,20 @@
 import { describe, it, expect } from "vitest";
-import { describeJob, timeAgo } from "./status";
+import { describeJob, etaLabel, timeAgo } from "./status";
+
+describe("etaLabel", () => {
+  it("formats seconds, minutes, and hours", () => {
+    expect(etaLabel(45)).toBe("~45s left");
+    expect(etaLabel(240)).toBe("~4m left");
+    expect(etaLabel(3900)).toBe("~1h 5m left");
+  });
+
+  it("returns null for missing or non-positive input", () => {
+    expect(etaLabel(null)).toBeNull();
+    expect(etaLabel(undefined)).toBeNull();
+    expect(etaLabel(0)).toBeNull();
+    expect(etaLabel(-5)).toBeNull();
+  });
+});
 
 describe("describeJob", () => {
   it("maps pipeline states to literal labels + severity kind", () => {

web/src/app/_ui/status.ts

@@ -27,6 +27,16 @@ export function timeAgo(value: string | Date | null | undefined, now: number): s
   return `${Math.floor(h / 24)}d`;
 }
 
+/** Compact download ETA like "~4m left", "~45s left", "~1h 5m left". Null/≤0 → no label. */
+export function etaLabel(seconds: number | null | undefined): string | null {
+  if (seconds == null || seconds <= 0) return null;
+  if (seconds < 60) return `~${seconds}s left`;
+  const m = Math.round(seconds / 60);
+  if (m < 60) return `~${m}m left`;
+  const h = Math.floor(m / 60);
+  return `~${h}h ${m % 60}m left`;
+}
+
 export function describeJob(state: string, stage: string): { label: string; kind: Kind } {
   void stage; // stage no longer drives the label; kept for call-site compatibility
   return (

web/src/app/api/discover/reset/route.test.ts

@@ -0,0 +1,17 @@
+import { describe, it, expect } from "vitest";
+import { prisma } from "@/lib/db";
+import { POST } from "./route";
+
+describe("POST /api/discover/reset", () => {
+  it("sets the reset + requested flags so the worker rebuilds on the next sweep", async () => {
+    const res = await POST();
+    expect(res.status).toBe(202);
+    expect((await res.json()).reset).toBe(true);
+    const cfg = Object.fromEntries(
+      (await prisma.config.findMany({ where: { key: { in: ["discover.resetRequested", "discover.requested"] } } }))
+        .map((c) => [c.key, c.value]),
+    );
+    expect(cfg["discover.resetRequested"]).toBe("true");
+    expect(cfg["discover.requested"]).toBe("true");
+  });
+});

web/src/app/api/discover/reset/route.ts

@@ -0,0 +1,17 @@
+import { prisma } from "@/lib/db";
+
+// POST /api/discover/reset — request a full rebuild: the worker clears the pending
+// suggestions, contributions, and seed throttles at the start of the next sweep, then
+// regenerates everything with the current logic. Kicks a sweep immediately.
+export async function POST() {
+  await Promise.all(
+    ["discover.resetRequested", "discover.requested"].map((key) =>
+      prisma.config.upsert({
+        where: { key },
+        create: { key, value: "true", secret: false },
+        update: { value: "true" },
+      }),
+    ),
+  );
+  return Response.json({ reset: true }, { status: 202 });
+}

web/src/app/api/floor/route.test.ts

@@ -0,0 +1,63 @@
+import { describe, it, expect } from "vitest";
+import { GET, POST } from "./route";
+import { prisma } from "@/lib/db";
+
+function post(action: string) {
+  return new Request("http://localhost/api/floor", {
+    method: "POST",
+    headers: { "content-type": "application/json" },
+    body: JSON.stringify({ action }),
+  });
+}
+
+async function flag(key: string) {
+  return (await prisma.config.findUnique({ where: { key } }))?.value ?? null;
+}
+
+async function queuedRequest(album: string, state = "requested") {
+  return prisma.request.create({
+    data: { artist: "A", album, status: "pending", job: { create: { state: state as never } } },
+    include: { job: true },
+  });
+}
+
+describe("floor API", () => {
+  it("GET reports paused=false by default", async () => {
+    const res = await GET();
+    expect(await res.json()).toEqual({ paused: false });
+  });
+
+  it("pause and resume toggle floor.paused", async () => {
+    await POST(post("pause"));
+    expect(await flag("floor.paused")).toBe("true");
+    const paused = await (await GET()).json();
+    expect(paused).toEqual({ paused: true });
+
+    await POST(post("resume"));
+    expect(await flag("floor.paused")).toBe("false");
+  });
+
+  it("stop sets both paused and killRequested", async () => {
+    const res = await POST(post("stop"));
+    expect(res.status).toBe(200);
+    expect(await flag("floor.paused")).toBe("true");
+    expect(await flag("floor.killRequested")).toBe("true");
+  });
+
+  it("clear deletes only queued (requested) requests", async () => {
+    const queued = await queuedRequest("Queued");
+    const active = await queuedRequest("Downloading", "downloading");
+
+    const res = await POST(post("clear"));
+    const body = await res.json();
+    expect(body.cleared).toBe(1);
+
+    expect(await prisma.request.findUnique({ where: { id: queued.id } })).toBeNull();
+    expect(await prisma.request.findUnique({ where: { id: active.id } })).not.toBeNull();
+  });
+
+  it("rejects an unknown action", async () => {
+    const res = await POST(post("nope"));
+    expect(res.status).toBe(400);
+  });
+});

web/src/app/api/floor/route.ts

@@ -0,0 +1,49 @@
+import { prisma } from "@/lib/db";
+
+async function getFlag(key: string): Promise<boolean> {
+  const row = await prisma.config.findUnique({ where: { key } });
+  return row?.value === "true";
+}
+
+async function setFlag(key: string, value: boolean): Promise<void> {
+  const v = value ? "true" : "false";
+  await prisma.config.upsert({
+    where: { key },
+    create: { key, value: v, secret: false },
+    update: { value: v },
+  });
+}
+
+export async function GET() {
+  return Response.json({ paused: await getFlag("floor.paused") });
+}
+
+export async function POST(request: Request) {
+  let body: { action?: string };
+  try {
+    body = (await request.json()) as { action?: string };
+  } catch {
+    return Response.json({ error: "invalid JSON" }, { status: 400 });
+  }
+
+  switch (body.action) {
+    case "pause":
+      await setFlag("floor.paused", true);
+      return Response.json({ ok: true, paused: true });
+    case "resume":
+      await setFlag("floor.paused", false);
+      return Response.json({ ok: true, paused: false });
+    case "stop":
+      await setFlag("floor.paused", true);
+      await setFlag("floor.killRequested", true);
+      return Response.json({ ok: true, paused: true, stopped: true });
+    case "clear": {
+      const { count } = await prisma.request.deleteMany({
+        where: { job: { is: { state: "requested" } } },
+      });
+      return Response.json({ ok: true, cleared: count });
+    }
+    default:
+      return Response.json({ error: "unknown action" }, { status: 400 });
+  }
+}

web/src/app/api/monitor/config/route.test.ts

@@ -34,3 +34,23 @@ describe("monitor config API", () => {
     expect(await prisma.config.count()).toBe(before);
   });
 });
+
+describe("monitor config — floor.jobDelaySeconds", () => {
+  it("defaults jobDelaySeconds to 0 and round-trips a new value", async () => {
+    const before = await (await GET()).json();
+    expect(before["floor.jobDelaySeconds"]).toBe("0");
+
+    await PATCH(
+      new Request("http://localhost/api/monitor/config", {
+        method: "PATCH",
+        headers: { "content-type": "application/json" },
+        body: JSON.stringify({ "floor.jobDelaySeconds": "45" }),
+      }),
+    );
+
+    const stored = await prisma.config.findUnique({ where: { key: "floor.jobDelaySeconds" } });
+    expect(stored?.value).toBe("45");
+    const after = await (await GET()).json();
+    expect(after["floor.jobDelaySeconds"]).toBe("45");
+  });
+});

web/src/app/api/monitor/config/route.ts

@@ -6,6 +6,7 @@ const DEFAULTS: Record<string, string> = {
   "monitor.retryIntervalHours": "6",
   "monitor.qualityCutoff": "2",
   "monitor.upgradeWindowDays": "14",
+  "floor.jobDelaySeconds": "0",
 };
 
 export async function GET() {

web/src/app/api/qobuz/pacing/route.test.ts

@@ -0,0 +1,37 @@
+import { describe, it, expect } from "vitest";
+import { GET, PATCH } from "./route";
+import { prisma } from "@/lib/db";
+
+function patch(body: Record<string, string>) {
+  return new Request("http://localhost/api/qobuz/pacing", {
+    method: "PATCH",
+    headers: { "content-type": "application/json" },
+    body: JSON.stringify(body),
+  });
+}
+
+describe("qobuz pacing config API", () => {
+  it("returns defaults when nothing is stored", async () => {
+    const cfg = await (await GET()).json();
+    expect(cfg["qobuz.pacing.enabled"]).toBe("true");
+    expect(cfg["qobuz.pacing.activeStartHour"]).toBe("8");
+    expect(cfg["qobuz.pacing.activeEndHour"]).toBe("23");
+    expect(cfg["qobuz.pacing.dailyCap"]).toBe("40");
+    expect(cfg["qobuz.pacing.warmupStartDate"]).toBe("");
+  });
+
+  it("round-trips allow-listed settings and persists them", async () => {
+    await PATCH(patch({ "qobuz.pacing.dailyCap": "25", "qobuz.pacing.activeStartHour": "9" }));
+    const stored = await prisma.config.findUnique({ where: { key: "qobuz.pacing.dailyCap" } });
+    expect(stored?.value).toBe("25");
+    const cfg = await (await GET()).json();
+    expect(cfg["qobuz.pacing.dailyCap"]).toBe("25");
+    expect(cfg["qobuz.pacing.activeStartHour"]).toBe("9");
+  });
+
+  it("ignores keys not in the allow-list (e.g. worker-managed counters)", async () => {
+    await PATCH(patch({ "qobuz.pacing.countToday": "999", "qobuz.pacing.dailyCap": "30" }));
+    expect(await prisma.config.findUnique({ where: { key: "qobuz.pacing.countToday" } })).toBeNull();
+    expect((await prisma.config.findUnique({ where: { key: "qobuz.pacing.dailyCap" } }))?.value).toBe("30");
+  });
+});

web/src/app/api/qobuz/pacing/route.ts

@@ -0,0 +1,39 @@
+import { prisma } from "@/lib/db";
+
+// Qobuz pacing ("budget gate") settings. The worker reads these Config keys live; state keys
+// (countDay/countToday/nextAllowedAt) are worker-managed and intentionally NOT editable here.
+const DEFAULTS: Record<string, string> = {
+  "qobuz.pacing.enabled": "true",
+  "qobuz.pacing.activeStartHour": "8",
+  "qobuz.pacing.activeEndHour": "23",
+  "qobuz.pacing.dailyCap": "40",
+  "qobuz.pacing.warmupStartDate": "",
+};
+
+export async function GET() {
+  const rows = await prisma.config.findMany({ where: { key: { in: Object.keys(DEFAULTS) } } });
+  const stored = Object.fromEntries(rows.map((r) => [r.key, r.value]));
+  return Response.json(
+    Object.fromEntries(Object.entries(DEFAULTS).map(([k, d]) => [k, stored[k] ?? d])),
+  );
+}
+
+export async function PATCH(request: Request) {
+  let body: unknown;
+  try {
+    body = await request.json();
+  } catch {
+    return Response.json({ error: "invalid JSON" }, { status: 400 });
+  }
+  const entries = Object.entries((body ?? {}) as Record<string, unknown>).filter(([k]) =>
+    Object.prototype.hasOwnProperty.call(DEFAULTS, k),
+  );
+  for (const [key, value] of entries) {
+    await prisma.config.upsert({
+      where: { key },
+      create: { key, value: String(value), secret: false },
+      update: { value: String(value) },
+    });
+  }
+  return Response.json({ updated: entries.length });
+}

web/src/app/api/requests/[id]/pause/route.test.ts

@@ -0,0 +1,49 @@
+import { describe, it, expect } from "vitest";
+import { POST } from "./route";
+import { prisma } from "@/lib/db";
+import type { Prisma } from "@prisma/client";
+
+function req(id: string, paused: boolean) {
+  return new Request(`http://localhost/api/requests/${id}/pause`, {
+    method: "POST",
+    headers: { "content-type": "application/json" },
+    body: JSON.stringify({ paused }),
+  });
+}
+
+async function seed(jobData: Prisma.JobCreateWithoutRequestInput) {
+  return prisma.request.create({
+    data: { artist: "A", album: "B", status: "pending", job: { create: jobData } },
+    include: { job: true },
+  });
+}
+
+describe("per-item pause API", () => {
+  it("pauses a queued job", async () => {
+    const created = await seed({ state: "requested" });
+    const res = await POST(req(created.id, true), { params: Promise.resolve({ id: created.id }) });
+    expect(res.status).toBe(200);
+    const job = await prisma.job.findUnique({ where: { id: created.job!.id } });
+    expect(job?.paused).toBe(true);
+  });
+
+  it("resumes a paused job", async () => {
+    const created = await seed({ state: "requested", paused: true });
+    await POST(req(created.id, false), { params: Promise.resolve({ id: created.id }) });
+    const job = await prisma.job.findUnique({ where: { id: created.job!.id } });
+    expect(job?.paused).toBe(false);
+  });
+
+  it("rejects pausing an in-flight job", async () => {
+    const created = await seed({ state: "downloading", paused: true });
+    const res = await POST(req(created.id, false), { params: Promise.resolve({ id: created.id }) });
+    expect(res.status).toBe(400);
+    const job = await prisma.job.findUnique({ where: { id: created.job!.id } });
+    expect(job?.paused).toBe(true);
+  });
+
+  it("404s for a missing request", async () => {
+    const res = await POST(req("nope", true), { params: Promise.resolve({ id: "nope" }) });
+    expect(res.status).toBe(404);
+  });
+});

web/src/app/api/requests/[id]/pause/route.ts

@@ -0,0 +1,26 @@
+import { prisma } from "@/lib/db";
+
+export async function POST(request: Request, { params }: { params: Promise<{ id: string }> }) {
+  const { id } = await params;
+
+  let body: { paused?: unknown };
+  try {
+    body = (await request.json()) as { paused?: unknown };
+  } catch {
+    return Response.json({ error: "invalid JSON" }, { status: 400 });
+  }
+  if (typeof body.paused !== "boolean") {
+    return Response.json({ error: "paused must be a boolean" }, { status: 400 });
+  }
+
+  const req = await prisma.request.findUnique({ where: { id }, include: { job: true } });
+  if (!req || !req.job) {
+    return Response.json({ error: "not found" }, { status: 404 });
+  }
+  if (req.job.state !== "requested") {
+    return Response.json({ error: "only queued jobs can be paused" }, { status: 400 });
+  }
+
+  await prisma.job.update({ where: { id: req.job.id }, data: { paused: body.paused } });
+  return Response.json({ ok: true, paused: body.paused });
+}

web/src/app/api/requests/[id]/route.test.ts

@@ -0,0 +1,37 @@
+import { describe, it, expect } from "vitest";
+import { DELETE } from "./route";
+import { prisma } from "@/lib/db";
+import type { Prisma } from "@prisma/client";
+
+function del(id: string) {
+  return new Request(`http://localhost/api/requests/${id}`, { method: "DELETE" });
+}
+
+async function seed(jobData: Prisma.JobCreateWithoutRequestInput) {
+  return prisma.request.create({
+    data: { artist: "A", album: "B", status: "pending", job: { create: jobData } },
+    include: { job: true },
+  });
+}
+
+describe("remove request API", () => {
+  it("deletes a queued request and cascades its job", async () => {
+    const created = await seed({ state: "requested" });
+    const res = await DELETE(del(created.id), { params: Promise.resolve({ id: created.id }) });
+    expect(res.status).toBe(200);
+    expect(await prisma.request.findUnique({ where: { id: created.id } })).toBeNull();
+    expect(await prisma.job.findUnique({ where: { id: created.job!.id } })).toBeNull();
+  });
+
+  it("refuses to remove an in-flight job", async () => {
+    const created = await seed({ state: "downloading" });
+    const res = await DELETE(del(created.id), { params: Promise.resolve({ id: created.id }) });
+    expect(res.status).toBe(400);
+    expect(await prisma.request.findUnique({ where: { id: created.id } })).not.toBeNull();
+  });
+
+  it("404s for a missing request", async () => {
+    const res = await DELETE(del("nope"), { params: Promise.resolve({ id: "nope" }) });
+    expect(res.status).toBe(404);
+  });
+});

web/src/app/api/requests/[id]/route.ts

@@ -0,0 +1,16 @@
+import { prisma } from "@/lib/db";
+
+export async function DELETE(_request: Request, { params }: { params: Promise<{ id: string }> }) {
+  const { id } = await params;
+
+  const req = await prisma.request.findUnique({ where: { id }, include: { job: true } });
+  if (!req) {
+    return Response.json({ error: "not found" }, { status: 404 });
+  }
+  if (req.job && req.job.state !== "requested") {
+    return Response.json({ error: "only queued items can be removed" }, { status: 400 });
+  }
+
+  await prisma.request.delete({ where: { id } }); // cascades Job + Candidates
+  return Response.json({ ok: true });
+}

web/src/app/api/requests/route.ts

@@ -64,6 +64,8 @@ export async function GET() {
               claimedAt: r.job.claimedAt,
               updatedAt: r.job.updatedAt,
               downloadProgress: r.job.downloadProgress,
+              downloadEtaSeconds: r.job.downloadEtaSeconds,
+              paused: r.job.paused,
               candidateCount: cands.length,
               chosen: chosen ? { source: chosen.source, format: chosen.format, trackCount: chosen.trackCount } : null,
             };

web/src/app/design.css

@@ -245,6 +245,9 @@ nav.contents .sep { flex: 1; }
 .btn.ghost { border-color: var(--rule-2); color: var(--graphite); }
 .btn.ghost:hover { background: transparent; border-color: var(--ink); color: var(--ink); }
 
+.press-controls { display: flex; gap: 0.5rem; align-items: center; margin: 0 0 1rem; flex-wrap: wrap; }
+.item-controls { display: inline-flex; gap: 0.5rem; align-items: center; }
+
 .toggle {
   font-family: var(--mono); font-size: 0.64rem; letter-spacing: 0.1em; text-transform: uppercase;
   color: var(--graphite); display: inline-flex; align-items: center; gap: 7px; cursor: pointer;
@@ -287,16 +290,19 @@ nav.contents .sep { flex: 1; }
 .edit-meta { display: flex; flex-direction: column; gap: 8px; width: 100%; }
 .rmeta.why { font-style: italic; color: var(--graphite); margin-top: 2px; }
 
-/* Discover unified row: artist block | inline album thumbnails | actions */
-.disco-row { align-items: flex-start; flex-wrap: wrap; gap: 14px; }
-.disco-row .main { flex: 1 1 180px; }
-.disco-albums { display: flex; gap: 12px; flex-wrap: wrap; }
-.disco-album { display: flex; flex-direction: column; gap: 4px; width: 92px; }
-.disco-thumb { display: flex; flex-direction: column; gap: 4px; background: none; border: 0; padding: 0; cursor: pointer; text-align: left; }
-.disco-thumb .thumb { width: 92px; height: 92px; }
-.disco-album .da-title { font-size: 12px; line-height: 1.2; color: var(--ink); overflow: hidden; display: -webkit-box; -webkit-line-clamp: 2; -webkit-box-orient: vertical; }
-.disco-album .badge { font-family: var(--mono); font-size: 10px; text-transform: uppercase; letter-spacing: 0.04em; color: var(--graphite); }
-.disco-album .btn.sm { padding: 2px 8px; }
+/* Discover unified row: artist block | compact album list (stacked) | actions.
+   Each album is a small horizontal item (cover + title/badge on one line) so two stack
+   vertically without making the row taller than the artist block. */
+.disco-row { align-items: center; flex-wrap: wrap; gap: 14px 20px; }
+.disco-row .main { flex: 1 1 200px; }
+.disco-albums { display: flex; flex-direction: column; gap: 6px; flex: 0 1 300px; min-width: 220px; }
+.disco-album { display: flex; align-items: center; gap: 8px; }
+.disco-thumb { display: flex; align-items: center; gap: 8px; flex: 1; min-width: 0; background: none; border: 0; padding: 0; cursor: pointer; text-align: left; }
+.disco-thumb .da-cover { width: 30px; height: 30px; flex-shrink: 0; }
+.da-text { display: flex; flex-direction: column; min-width: 0; line-height: 1.15; }
+.disco-album .da-title { font-size: 12px; color: var(--ink); white-space: nowrap; overflow: hidden; text-overflow: ellipsis; }
+.disco-album .badge { font-family: var(--mono); font-size: 9px; text-transform: uppercase; letter-spacing: 0.04em; color: var(--graphite); }
+.disco-album .btn.sm.want { padding: 1px 8px; flex-shrink: 0; }
 .sel-tools { display: inline-flex; align-items: center; gap: 10px; }
 .bulk-bar { display: flex; align-items: center; gap: 12px; flex-wrap: wrap; margin: 4px 0 12px; padding: 8px 12px; border: 1px solid var(--rule); background: var(--paper-2); }
 .album-card { position: relative; }
@@ -380,6 +386,39 @@ nav.contents .sep { flex: 1; }
 .tab:hover { color: var(--ink); }
 .tab.active { color: var(--ink); border-bottom-color: var(--accent); }
 .tab .n { color: var(--rule-2); margin-left: 6px; }
+.tab.has-errors .n { color: var(--alert); }
+
+/* "Next up" badge on the first queued row (the item the worker claims next) */
+.nextup {
+  display: inline-block; vertical-align: middle; margin-right: 9px; transform: translateY(-2px);
+  font-family: var(--mono); font-size: 0.56rem; letter-spacing: 0.13em; text-transform: uppercase;
+  color: var(--accent-ink); background: var(--accent); padding: 2px 7px; border-radius: 4px;
+}
+
+/* ── Kebab menu (⋮ popup for press actions) ───────────── */
+.dept.has-actions { align-items: center; }
+.dept.has-actions .fill { transform: none; }
+.menu { position: relative; display: inline-flex; }
+.menu-trigger {
+  font-size: 1.15rem; line-height: 1; color: var(--graphite); background: transparent;
+  border: 1px solid var(--rule); border-radius: 6px; width: 32px; height: 30px; cursor: pointer;
+  display: inline-flex; align-items: center; justify-content: center;
+}
+.menu-trigger:hover { color: var(--ink); border-color: var(--rule-2); }
+.menu-pop {
+  position: absolute; top: calc(100% + 6px); right: 0; z-index: 30; min-width: 190px;
+  background: var(--paper-2); border: 1px solid var(--rule-2); border-radius: 9px;
+  box-shadow: 0 10px 30px rgba(0, 0, 0, 0.22); padding: 5px;
+  display: flex; flex-direction: column; gap: 1px;
+}
+.menu-item {
+  font-family: var(--mono); font-size: 0.68rem; letter-spacing: 0.1em; text-transform: uppercase;
+  text-align: left; background: transparent; border: 0; color: var(--ink);
+  padding: 10px 11px; border-radius: 6px; cursor: pointer; white-space: nowrap;
+}
+.menu-item:hover { background: var(--rule); }
+.menu-item:disabled { color: var(--rule-2); cursor: default; background: transparent; }
+.menu-item.danger { color: var(--alert); }
 
 /* ── Discography row (cover thumb + click-to-modal) ───── */
 .disco-open {

web/src/app/discover/discover-client.tsx

@@ -36,6 +36,12 @@ function similarTo(seeds: string[]): string | null {
   return `Similar to ${shown}${extra > 0 ? ` +${extra} more` : ""}`;
 }
 
+/** Full-length albums only — hide singles/EPs and secondary-type releases (live/comp) that
+ *  pre-date the albums-only derivation and still linger in the DB. */
+function isFullAlbum(a: Album): boolean {
+  return (a.primaryType ?? "Album") === "Album" && (a.secondaryTypes?.length ?? 0) === 0;
+}
+
 /** Human badge for why an album is surfaced. */
 function badgeFor(reason: string | null): string | null {
   if (reason === "newest") return "Newest";
@@ -63,6 +69,7 @@ export function DiscoverClient() {
   const [query, setQuery] = useState("");
   const [search, setSearch] = useState<SearchResult | null>(null);
   const [busy, setBusy] = useState(false);
+  const [confirmReset, setConfirmReset] = useState(false);
   const [followedSimilar, setFollowedSimilar] = useState<Set<string>>(new Set());
   const [albumModal, setAlbumModal] = useState<{ a: Album; artistName: string; artistMbid: string } | null>(null);
 
@@ -100,6 +107,18 @@ export function DiscoverClient() {
     setRunning(true);
   }
 
+  async function resetDiscover() {
+    setConfirmReset(false);
+    const res = await fetch("/api/discover/reset", { method: "POST" }).catch(() => null);
+    if (res && (res.ok || res.status === 202)) {
+      setRows([]); // clears immediately; the rebuild repopulates
+      setRunning(true);
+      toast("Rebuilding suggestions from scratch…");
+    } else {
+      toast("Couldn't reset");
+    }
+  }
+
   async function runSearch(e: React.FormEvent) {
     e.preventDefault();
     if (!query.trim()) return;
@@ -163,13 +182,26 @@ export function DiscoverClient() {
         <button className="btn accent" onClick={discoverNow} disabled={running}>
           {running ? "Discovering…" : "Discover now"}
         </button>
-        <span className="rmeta">
-          {running
-            ? "Sweeping your followed artists for new suggestions…"
-            : lastRunAt
-              ? `Last discovered ${agoLabel(lastRunAt)}${result ? ` · ${result}` : ""}`
-              : "Not yet run — press Discover now to sweep your followed artists"}
-        </span>
+        {confirmReset ? (
+          <>
+            <span className="rmeta">Clear all suggestions and rebuild from scratch?</span>
+            <button className="btn sm accent" onClick={resetDiscover}>Reset</button>
+            <button className="btn sm ghost" onClick={() => setConfirmReset(false)}>Cancel</button>
+          </>
+        ) : (
+          <>
+            <button className="btn sm ghost" onClick={() => setConfirmReset(true)} disabled={running}>
+              Reset
+            </button>
+            <span className="rmeta">
+              {running
+                ? "Sweeping your followed artists for new suggestions…"
+                : lastRunAt
+                  ? `Last discovered ${agoLabel(lastRunAt)}${result ? ` · ${result}` : ""}`
+                  : "Not yet run — press Discover now to sweep your followed artists"}
+            </span>
+          </>
+        )}
       </div>
 
       <SectionHeader title="Find similar" />
@@ -214,7 +246,9 @@ export function DiscoverClient() {
         <p className="muted-note">No suggestions yet — run Discover, or follow a few artists to seed it.</p>
       ) : (
         <ul className="list">
-          {rows.map((row) => (
+          {rows.map((row) => {
+            const albums = row.albums.filter(isFullAlbum);
+            return (
             <li key={row.artistMbid} className="list-row disco-row">
               <div className="main">
                 <div className="rtitle">
@@ -230,20 +264,22 @@ export function DiscoverClient() {
                 {similarTo(row.seeds) ? <div className="rmeta why">{similarTo(row.seeds)}</div> : null}
               </div>
 
-              {row.albums.length > 0 ? (
+              {albums.length > 0 ? (
                 <div className="disco-albums">
-                  {row.albums.map((a) => (
+                  {albums.map((a) => (
                     <div key={a.id} className="disco-album">
                       <button
                         className="disco-thumb"
                         onClick={() => setAlbumModal({ a, artistName: row.artistName, artistMbid: row.artistMbid })}
                         aria-label={`Open ${a.album ?? ""}`}
                       >
-                        <CoverArt rgMbid={a.rgMbid} alt={a.album ?? ""} className="thumb" />
-                        <span className="da-title">{a.album}</span>
-                        {badgeFor(a.albumReason) ? <span className="badge">{badgeFor(a.albumReason)}</span> : null}
+                        <CoverArt rgMbid={a.rgMbid} alt={a.album ?? ""} className="da-cover" />
+                        <span className="da-text">
+                          <span className="da-title">{a.album}</span>
+                          {badgeFor(a.albumReason) ? <span className="badge">{badgeFor(a.albumReason)}</span> : null}
+                        </span>
                       </button>
-                      <button className="btn sm" onClick={() => act(a.id, "want")}>
+                      <button className="btn sm ghost want" onClick={() => act(a.id, "want")}>
                         Want
                       </button>
                     </div>
@@ -262,7 +298,8 @@ export function DiscoverClient() {
                 ) : null}
               </div>
             </li>
-          ))}
+            );
+          })}
         </ul>
       )}
 

web/src/app/settings/settings-form.tsx

@@ -31,6 +31,7 @@ const MONITOR_FIELDS: [string, string][] = [
   ["monitor.retryIntervalHours", "Retry interval (hours)"],
   ["monitor.qualityCutoff", "Quality cutoff"],
   ["monitor.upgradeWindowDays", "Upgrade window (days)"],
+  ["floor.jobDelaySeconds", "Delay between downloads (seconds)"],
 ];
 
 export function SettingsForm() {
@@ -53,6 +54,8 @@ export function SettingsForm() {
   const [discoverSaved, setDiscoverSaved] = useState(false);
   const [monitorCfg, setMonitorCfg] = useState<Record<string, string>>({});
   const [monitorSaved, setMonitorSaved] = useState(false);
+  const [pacingCfg, setPacingCfg] = useState<Record<string, string>>({});
+  const [pacingSaved, setPacingSaved] = useState(false);
 
   useEffect(() => {
     fetch("/api/scan")
@@ -105,6 +108,12 @@ export function SettingsForm() {
       .then((c: Record<string, string>) => setMonitorCfg(c));
   }, []);
 
+  useEffect(() => {
+    fetch("/api/qobuz/pacing")
+      .then((r) => r.json())
+      .then((c: Record<string, string>) => setPacingCfg(c));
+  }, []);
+
   async function saveCredentials(e: React.FormEvent) {
     e.preventDefault();
     await fetch("/api/config", {
@@ -159,6 +168,21 @@ export function SettingsForm() {
     setTimeout(() => setMonitorSaved(false), 1500);
   }
 
+  function setPacing(key: string, value: string) {
+    setPacingCfg((c) => ({ ...c, [key]: value }));
+  }
+
+  async function savePacing(e: React.FormEvent) {
+    e.preventDefault();
+    await fetch("/api/qobuz/pacing", {
+      method: "PATCH",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify(pacingCfg),
+    });
+    setPacingSaved(true);
+    setTimeout(() => setPacingSaved(false), 1500);
+  }
+
   const secretHint = (set: boolean) => (set ? <span className="set"> · set</span> : null);
 
   async function clearField(field: "qobuzToken" | "slskdApiKey" | "lastfmApiKey") {
@@ -192,6 +216,7 @@ export function SettingsForm() {
       </div>
 
       {tab === "Qobuz" ? (
+        <>
         <form className="settings-section" onSubmit={saveCredentials}>
           <div className="subhead">
             Auth token
@@ -234,6 +259,55 @@ export function SettingsForm() {
             {saved ? <span className="saved-note">Saved</span> : null}
           </div>
         </form>
+        <form className="settings-section" onSubmit={savePacing}>
+          <div className="subhead">
+            Pacing
+            <HelpButton title="Qobuz pacing">
+              Throttles Qobuz to a human-looking volume and cadence so the account isn&apos;t flagged
+              for bulk downloading. Within active hours it downloads up to the daily cap, spaced out
+              with randomized gaps; when throttled, albums fall through to Soulseek/YouTube and
+              upgrade to Qobuz hi-res later. The daily cap ramps up over the first two weeks from the
+              new-account start date.
+            </HelpButton>
+          </div>
+          <label className="field-inline">
+            <input type="checkbox" aria-label="pace qobuz downloads"
+              checked={pacingCfg["qobuz.pacing.enabled"] !== "false"}
+              onChange={(e) => setPacing("qobuz.pacing.enabled", e.target.checked ? "true" : "false")} />
+            <span>Pace Qobuz downloads</span>
+          </label>
+          <div className="field-grid">
+            <label className="field">
+              <span>Active hours — start (0–24)</span>
+              <input type="number" min={0} max={24} aria-label="qobuz active start hour"
+                value={pacingCfg["qobuz.pacing.activeStartHour"] ?? ""}
+                onChange={(e) => setPacing("qobuz.pacing.activeStartHour", e.target.value)} />
+            </label>
+            <label className="field">
+              <span>Active hours — end (0–24)</span>
+              <input type="number" min={0} max={24} aria-label="qobuz active end hour"
+                value={pacingCfg["qobuz.pacing.activeEndHour"] ?? ""}
+                onChange={(e) => setPacing("qobuz.pacing.activeEndHour", e.target.value)} />
+            </label>
+            <label className="field">
+              <span>Daily cap (steady)</span>
+              <input type="number" min={0} aria-label="qobuz daily cap"
+                value={pacingCfg["qobuz.pacing.dailyCap"] ?? ""}
+                onChange={(e) => setPacing("qobuz.pacing.dailyCap", e.target.value)} />
+            </label>
+            <label className="field">
+              <span>New-account start (warm-up)</span>
+              <input type="date" aria-label="qobuz warmup start date"
+                value={pacingCfg["qobuz.pacing.warmupStartDate"] ?? ""}
+                onChange={(e) => setPacing("qobuz.pacing.warmupStartDate", e.target.value)} />
+            </label>
+          </div>
+          <div className="save-row">
+            <button type="submit" className="btn accent">Save</button>
+            {pacingSaved ? <span className="saved-note">Saved</span> : null}
+          </div>
+        </form>
+        </>
       ) : null}
 
       {tab === "Soulseek" ? (

web/src/lib/artist-resolve.test.ts

@@ -0,0 +1,46 @@
+import { describe, it, expect } from "vitest";
+import { resolveArtistChoice, type ArtistHit } from "./artist-resolve";
+
+const US: ArtistHit = { mbid: "us", name: "Looking Glass", disambiguation: "70's US pop group" };
+const AU: ArtistHit = { mbid: "au", name: "Looking Glass", disambiguation: "Australian stoner rock" };
+
+describe("resolveArtistChoice", () => {
+  it("is ambiguous when 2+ candidates share the exact name", () => {
+    expect(resolveArtistChoice("Looking Glass", [US, AU], "au")).toEqual({
+      kind: "ambiguous",
+      candidates: [US, AU],
+    });
+  });
+
+  it("uses the single exact-name match even over a disagreeing fallback mbid", () => {
+    const hit: ArtistHit = { mbid: "r", name: "Radiohead", disambiguation: "" };
+    expect(resolveArtistChoice("Radiohead", [hit], "wrong-lastfm-mbid")).toEqual({
+      kind: "single",
+      mbid: "r",
+    });
+  });
+
+  it("matches exact name case- and whitespace-insensitively", () => {
+    expect(resolveArtistChoice("  looking GLASS ", [US, AU])).toEqual({
+      kind: "ambiguous",
+      candidates: [US, AU],
+    });
+  });
+
+  it("falls back to the provided mbid when no candidate name matches exactly", () => {
+    const fuzzy: ArtistHit = { mbid: "x", name: "The Beatles", disambiguation: "" };
+    expect(resolveArtistChoice("Beatles", [fuzzy], "lastfm-mbid")).toEqual({
+      kind: "single",
+      mbid: "lastfm-mbid",
+    });
+  });
+
+  it("falls back to MB's top hit when there is no exact match and no fallback", () => {
+    const fuzzy: ArtistHit = { mbid: "top", name: "The Beatles", disambiguation: "" };
+    expect(resolveArtistChoice("Beatles", [fuzzy])).toEqual({ kind: "single", mbid: "top" });
+  });
+
+  it("is none when there are no candidates and no fallback", () => {
+    expect(resolveArtistChoice("Nobody Here", [])).toEqual({ kind: "none" });
+  });
+});

web/src/lib/artist-resolve.ts

@@ -0,0 +1,34 @@
+export type ArtistHit = { mbid: string; name: string; disambiguation: string };
+
+export type Resolution =
+  | { kind: "single"; mbid: string }
+  | { kind: "ambiguous"; candidates: ArtistHit[] }
+  | { kind: "none" };
+
+const norm = (s: string) => s.trim().toLowerCase();
+
+/**
+ * Decide which MusicBrainz artist a Last.fm row refers to, given MB's search candidates.
+ *
+ * The bug this guards against: Last.fm's own MBID for an ambiguous name can point at the
+ * wrong same-named band (it mapped "Looking Glass" to an Australian stoner band, not the US
+ * pop group). So we prefer MB's own resolution:
+ *   - 2+ candidates whose name EXACTLY matches (normalized) → ambiguous; let the user pick.
+ *   - exactly 1 exact-name match → use it, even if it disagrees with `fallbackMbid`.
+ *   - no exact-name match → keep prior behavior: the Last.fm `fallbackMbid`, else MB's top hit.
+ *
+ * `candidates` is assumed to arrive in MB relevance order (best first), so `candidates[0]`
+ * is the suggested default and exact-name matches keep that ordering.
+ */
+export function resolveArtistChoice(
+  name: string,
+  candidates: ArtistHit[],
+  fallbackMbid?: string | null,
+): Resolution {
+  const exact = candidates.filter((a) => norm(a.name) === norm(name));
+  if (exact.length >= 2) return { kind: "ambiguous", candidates: exact };
+  if (exact.length === 1) return { kind: "single", mbid: exact[0].mbid };
+  if (fallbackMbid) return { kind: "single", mbid: fallbackMbid };
+  if (candidates.length > 0) return { kind: "single", mbid: candidates[0].mbid };
+  return { kind: "none" };
+}

worker/lyra_worker/_musicbrainz.py

@@ -1,3 +1,4 @@
+import time
 from difflib import SequenceMatcher
 
 from lyra_worker.types import MBTarget
@@ -5,10 +6,104 @@ from lyra_worker.types import MBTarget
 _MIN_TITLE_RATIO = 0.6
 
 
+def _with_retry(fn, *, attempts: int = 4, base_delay: float = 1.0, sleep=time.sleep):
+    """Call `fn`, retrying transient MusicBrainz failures with exponential backoff.
+
+    A dropped TLS connection to the MusicBrainz webservice surfaces as
+    musicbrainzngs.NetworkError (`<urlopen error [SSL: UNEXPECTED_EOF_WHILE_READING]>`),
+    and MusicBrainz's own 503 rate-limit surfaces as ResponseError; both are
+    transient, so we retry rather than failing the whole pipeline for the job.
+    Any other exception (bug, non-transient) propagates immediately.
+
+    musicbrainzngs is imported lazily so importing this module stays offline.
+    """
+    import musicbrainzngs
+
+    transient = (musicbrainzngs.NetworkError, musicbrainzngs.ResponseError)
+    for i in range(attempts):
+        try:
+            return fn()
+        except transient:
+            if i == attempts - 1:
+                raise
+            sleep(base_delay * (2**i))
+
+
 def _ratio(a: str, b: str) -> float:
     return SequenceMatcher(None, a.strip().casefold(), b.strip().casefold()).ratio()
 
 
+def _is_album(g: dict) -> bool:
+    return (g.get("primary-type") or "").casefold() == "album"
+
+
+def _rg_year(g: dict) -> int:
+    """First-release year of a release-group, or 9999 when unknown (so undated groups lose ties)."""
+    frd = (g.get("first-release-date") or "")[:4]
+    return int(frd) if frd.isdigit() else 9999
+
+
+def _credited_artist(g: dict) -> str:
+    """Primary credited artist name of a release-group search result, or '' if absent."""
+    credit = g.get("artist-credit")
+    if isinstance(credit, list) and credit and isinstance(credit[0], dict):
+        return (credit[0].get("artist") or {}).get("name", "") or ""
+    return ""
+
+
+def _best_release_group(artist: str, album: str, groups: list) -> dict | None:
+    """Choose the best release-group match for `artist`/`album`.
+
+    Ranked, highest first, on `(artist_ratio, title_ratio, is_album)`:
+      * artist match dominates — MusicBrainz relevance ties same-titled release groups by
+        *different* artists (e.g. "Fun Machine" exists as a Lake Street Dive EP and an
+        unrelated band's Album), and the artist we followed is trustworthy, so a credited
+        artist that matches the request outranks everything else;
+      * title similarity is next;
+      * an Album primary-type only breaks ties *within* the same artist+title, so a famous
+        album (Michael Jackson's "Off the Wall") still isn't resolved to its same-named
+        single whose short tracklist would map positionally onto the album's files;
+      * finally, the EARLIEST release-group wins — MusicBrainz lists the original album and its
+        later reissues/anniversary editions as separate same-titled groups (Linkin Park "Hybrid
+        Theory": a 2000 original and a 2023 13-track reissue), which tie on everything above, so
+        MB's arbitrary relevance order used to pick a reissue whose longer tracklist/runtime then
+        failed completeness checks against the standard album a source delivers. Prefer the original.
+    Artist is a ranking signal, never a hard filter — a low match sinks a candidate but
+    never drops the release, so credited-name variations (feat., punctuation) still resolve.
+    Returns None below the title threshold. Pure — no network I/O."""
+    if not groups:
+        return None
+    best = max(
+        groups,
+        key=lambda g: (
+            _ratio(artist, _credited_artist(g)),
+            _ratio(album, g.get("title", "")),
+            _is_album(g),
+            -_rg_year(g),  # tie-break: the original (earliest) group over later reissues
+        ),
+    )
+    if _ratio(album, best.get("title", "")) < _MIN_TITLE_RATIO:
+        return None
+    return best
+
+
+def _pick_release(releases: list) -> dict | None:
+    """Choose which release of a release-group to read the tracklist from. MB returns them in an
+    arbitrary order and ``releases[0]`` is often a later reissue/deluxe/anniversary edition with
+    more tracks and a longer runtime than the standard album a source actually delivers — which
+    then makes a *complete* download fail the pipeline's track-count/duration completeness checks
+    and re-download forever (Linkin Park "Hybrid Theory": MB's arbitrary pick was a 2023 13-track
+    reissue at ~52 min vs the 2000 original's 12 tracks / ~38 min). Prefer the ORIGINAL standard
+    edition: an Official release, then the earliest date, then a dated one. Pure — unit-tested."""
+    if not releases:
+        return None
+    def _key(r: dict) -> tuple:
+        official = 0 if (r.get("status") == "Official") else 1  # official editions first
+        date = r.get("date") or ""
+        return (official, 0 if date else 1, date)  # then dated, then earliest date
+    return min(releases, key=_key)
+
+
 class MusicBrainzResolver:
     """Real MbResolver using the MusicBrainz webservice via musicbrainzngs.
 
@@ -26,12 +121,10 @@ class MusicBrainzResolver:
 
         musicbrainzngs.set_useragent(self._app, self._version, self._contact)
 
-        res = musicbrainzngs.search_release_groups(query=album, artist=artist, limit=5)
+        res = _with_retry(lambda: musicbrainzngs.search_release_groups(query=album, artist=artist, limit=5))
         groups = res.get("release-group-list", [])
-        if not groups:
-            return None
-        rg = max(groups, key=lambda g: _ratio(album, g.get("title", "")))
-        if _ratio(album, rg.get("title", "")) < _MIN_TITLE_RATIO:
+        rg = _best_release_group(artist, album, groups)
+        if rg is None:
             return None
 
         canonical_album = rg.get("title", album)
@@ -53,11 +146,12 @@ class MusicBrainzResolver:
         titles: tuple[str, ...] = ()
         rgid = rg.get("id")
         if rgid:
-            rgfull = musicbrainzngs.get_release_group_by_id(rgid, includes=["releases"])
+            rgfull = _with_retry(lambda: musicbrainzngs.get_release_group_by_id(rgid, includes=["releases"]))
             releases = rgfull.get("release-group", {}).get("release-list", [])
-            if releases:
-                rel = musicbrainzngs.get_release_by_id(
-                    releases[0]["id"], includes=["recordings"]
+            chosen = _pick_release(releases)
+            if chosen:
+                rel = _with_retry(
+                    lambda: musicbrainzngs.get_release_by_id(chosen["id"], includes=["recordings"])
                 ).get("release", {})
                 tracks = []
                 for medium in rel.get("medium-list", []):

worker/lyra_worker/_mutagen.py

@@ -1,21 +1,48 @@
 import os
+import re
 
 from lyra_worker.library import safe_name
 
 _AUDIO_EXT = {".flac", ".mp3", ".m4a", ".opus", ".ogg", ".aac", ".wav"}
 
+# Leading "13.", "13 -", "13)" style track-number prefix streamrip bakes into a bonus-track
+# filename. Requires an explicit delimiter so a real title that merely starts with a number
+# (e.g. "99 Luftballons") is left alone.
+_NUM_PREFIX = re.compile(r"^\s*\d+\s*[.):\-]\s*")
 
-def plan_track_names(names: list[str], tracklist: tuple[str, ...]) -> list[tuple[str, str, str, int]]:
+
+def _clean_extra_title(stem: str, artist: str = "") -> str:
+    """Derive a display title for a track BEYOND the MusicBrainz tracklist (a bonus track that
+    streamrip named itself, e.g. `13. John Mayer - St. Patrick's Day (Album Version)`). Strip a
+    leading track-number prefix, then a leading `<artist> - ` if present, so the example yields
+    `St. Patrick's Day (Album Version)`. Falls back to the stem if nothing strips."""
+    s = _NUM_PREFIX.sub("", stem).strip()
+    if artist:
+        s = re.sub(r"^" + re.escape(artist) + r"\s*-\s*", "", s, flags=re.IGNORECASE).strip()
+    return s or stem
+
+
+def plan_track_names(
+    names: list[str], tracklist: tuple[str, ...], artist: str = ""
+) -> list[tuple[str, str, str, int]]:
     """Map audio filenames -> (old_name, new_name, title, track_number).
 
-    Files are sorted; titles come from `tracklist` by position when available,
-    else the filename stem. Pure — no filesystem or tag I/O.
+    Files are sorted; titles come from `tracklist` by position when available. A file BEYOND
+    the tracklist (a bonus track streamrip named itself) is canonicalized via `_clean_extra_title`
+    so it gets a clean `## Title.ext` name too; with no tracklist at all the on-disk stem is kept.
+    Pure — no filesystem or tag I/O.
     """
     plan: list[tuple[str, str, str, int]] = []
     audio = sorted(n for n in names if os.path.splitext(n)[1].lower() in _AUDIO_EXT)
     for i, name in enumerate(audio):
         ext = os.path.splitext(name)[1].lower()
-        title = tracklist[i] if i < len(tracklist) else os.path.splitext(name)[0]
+        stem = os.path.splitext(name)[0]
+        if i < len(tracklist):
+            title = tracklist[i]
+        elif tracklist:  # a genuine bonus track past a known tracklist -> canonicalize its name
+            title = _clean_extra_title(stem, artist)
+        else:  # no tracklist resolved at all -> keep the on-disk stem (renumbered)
+            title = stem
         new_name = f"{i + 1:02d} {safe_name(title)}{ext}"
         plan.append((name, new_name, title, i + 1))
     return plan
@@ -29,7 +56,7 @@ class MutagenTagger:
         import mutagen
 
         names = os.listdir(album_dir)
-        for old_name, new_name, title, number in plan_track_names(names, target.tracklist):
+        for old_name, new_name, title, number in plan_track_names(names, target.tracklist, target.artist):
             try:
                 path = os.path.join(album_dir, old_name)
                 audio = mutagen.File(path, easy=True)

worker/lyra_worker/adapters/_slskd.py

@@ -2,6 +2,7 @@ import json
 import os
 import shutil
 import time
+import unicodedata
 from typing import Callable
 
 import requests
@@ -10,9 +11,26 @@ _DEFAULT_DOWNLOADS_ROOT = "/slskd-downloads"
 
 _AUDIO_EXT = {"flac", "mp3", "m4a", "ogg", "opus", "wav", "aac"}
 _LOSSLESS_EXT = {"flac", "wav"}
-_SEARCH_POLLS = 40   # * 3s ≈ 2 min
-_XFER_POLLS = 200    # * 3s ≈ 10 min
+_SEARCH_POLLS = 40    # * 3s ≈ 2 min
+_STALL_POLLS = 30     # * 3s ≈ 90s of no byte progress → abandon a queued/dead peer and fall through
 _POLL_SECONDS = 3
+# Absolute backstop for a slow-but-progressing peer. Scaled to the album's total bytes (a flat
+# cap starved big albums: a 500MB set from a ~400KB/s serial peer needs ~60-70min and was cut at
+# a flat 60min, then re-downloaded from scratch — never finishing). Budget at a conservative floor
+# throughput, clamped. The 90s stall timeout still abandons a truly dead/queued peer quickly.
+_MIN_XFER_POLLS = 400        # * 3s ≈ 20 min floor (small albums)
+_MAX_XFER_POLLS = 2400       # * 3s ≈ 2 h ceiling (bounds worst-case slot hold)
+_BACKSTOP_FLOOR_BPS = 64_000  # budget the backstop assuming >= ~64 KB/s sustained
+
+
+def _max_polls(total_bytes: int) -> int:
+    """Poll budget (each _POLL_SECONDS) for the absolute download backstop, scaled to album size at
+    a conservative floor throughput so a large slow-but-healthy transfer isn't abandoned mid-flight.
+    Clamped to [_MIN_XFER_POLLS, _MAX_XFER_POLLS]. Pure — unit-tested."""
+    if total_bytes <= 0:
+        return _MIN_XFER_POLLS
+    budget = int(total_bytes / (_BACKSTOP_FLOOR_BPS * _POLL_SECONDS))
+    return max(_MIN_XFER_POLLS, min(budget, _MAX_XFER_POLLS))
 
 
 def _basename(path: str) -> str:
@@ -27,6 +45,114 @@ def _dirname(path: str) -> str:
     return ""
 
 
+_ETA_CAP_SECONDS = 359_999  # ~100h; keeps the value well inside Job.downloadEtaSeconds (int4)
+
+
+def _eta_seconds(total_bytes: int, bytes_now: int, speed_bps: float) -> int | None:
+    """Seconds until the transfer finishes at the current measured rate, or None when it can't be
+    estimated (no throughput sample yet, unknown total, or already complete). Pure — unit-tested.
+    A near-stalled peer drives speed_bps toward zero, so the estimate is capped: an uncapped value
+    can exceed int4 and overflow the downloadEtaSeconds column write (poisoning the pipeline txn)."""
+    if speed_bps <= 0 or total_bytes <= 0 or bytes_now >= total_bytes:
+        return None
+    return min(int((total_bytes - bytes_now) / speed_bps), _ETA_CAP_SECONDS)
+
+
+def _norm(s: str) -> str:
+    """Lowercase, strip diacritics, and reduce to alphanumeric tokens, so 'Beyoncé' matches a
+    peer's 'Beyonce' folder and punctuation ('I Am… Sasha Fierce') doesn't defeat the match."""
+    s = unicodedata.normalize("NFKD", s or "")
+    s = "".join(ch for ch in s if not unicodedata.combining(ch))
+    s = "".join(ch if ch.isalnum() else " " for ch in s.lower())
+    return " ".join(s.split())
+
+
+def _path_matches(needle: str, path: str) -> bool:
+    """True if `needle` (an artist or album name) appears in a file path, tolerant of accents and
+    punctuation. Every significant token (≥3 chars) must be present; for a very short needle
+    (e.g. 'U2') the whole normalized string must appear."""
+    npath = _norm(path)
+    tokens = [t for t in _norm(needle).split() if len(t) >= 3]
+    if not tokens:
+        whole = _norm(needle)
+        return bool(whole) and whole in npath
+    return all(t in npath for t in tokens)
+
+
+def _keep_matching(responses: list, needle: str) -> list:
+    """Drop files — then now-empty responses — whose path doesn't match `needle`. Makes an album-
+    or artist-only fallback search precise so we never grab a same-titled release by a different
+    artist. Pure — unit-tested offline."""
+    out = []
+    for resp in responses:
+        files = [f for f in (resp.get("files") or []) if _path_matches(needle, f.get("filename", ""))]
+        if files:
+            out.append({**resp, "files": files})
+    return out
+
+
+def _parse_search_responses(responses: list, artist: str = "") -> list[dict]:
+    """Turn slskd search responses into album candidates (one per peer+directory), ordered so the
+    peers that will FINISH first come first: a free upload slot, then the shortest estimated
+    transfer time (album bytes ÷ the peer's advertised speed), then a shorter queue. Ranking on
+    estimated time — not raw speed — matters because Soulseek FLAC rips of the same album vary
+    ~2x in size (a 24-bit/bloated rip vs a standard CD rip), and Lyra scores them the SAME quality
+    class (the search exposes no bit-depth), so a peer advertising high speed but serving huge
+    files can lose a race — and blow past the download backstop — against a peer with a smaller
+    standard rip. The pipeline tries candidates in this order, so this is what makes the Soulseek
+    fall-through converge on a copy that actually completes. Pure — no I/O — unit-tested offline."""
+    scored: list[tuple] = []
+    for resp in responses:
+        username = resp.get("username", "")
+        has_slot = 0 if resp.get("hasFreeUploadSlot") else 1  # 0 sorts first (free slot = starts now)
+        speed = int(resp.get("uploadSpeed") or 0)
+        queue = int(resp.get("queueLength") or 0)
+        by_dir: dict[str, list] = {}
+        for f in resp.get("files", []) or []:
+            name = f.get("filename", "")
+            ext = name.rsplit(".", 1)[-1].lower() if "." in name else ""
+            if ext not in _AUDIO_EXT:
+                continue
+            by_dir.setdefault(_dirname(name), []).append(
+                {"filename": name, "size": f.get("size", 0), "ext": ext}
+            )
+        for directory, dfiles in by_dir.items():
+            if not dfiles:
+                continue
+            # Estimated seconds to pull this album from this peer at its advertised rate. Unknown/
+            # zero speed → a low nominal (1 B/s) so total size still orders those peers last.
+            total_bytes = sum(int(f["size"] or 0) for f in dfiles)
+            est_seconds = total_bytes / (speed if speed > 0 else 1)
+            lossless = all(f["ext"] in _LOSSLESS_EXT for f in dfiles)
+            dirname = _basename(directory)
+            # Trust the REQUESTED artist when it appears anywhere in the path — many rips put the
+            # artist in a parent folder ("Rihanna\(2009) Rated R\…") or bury it ("2009 - Rihanna -
+            # Rated R"), which the "Artist - Album" folder parse gets wrong; that tanked the
+            # confidence score and lost real FLACs to clean-named MP3s. album stays the folder name
+            # (fuzzy-matched downstream); artist is what the parse gets wrong.
+            if artist and _path_matches(artist, dfiles[0]["filename"]):
+                guess_artist, guess_album = artist, dirname
+            elif " - " in dirname:
+                guess_artist, guess_album = (p.strip() for p in dirname.split(" - ", 1))
+            else:
+                guess_artist, guess_album = "", dirname
+            candidate = {
+                "source_ref": json.dumps({
+                    "username": username,
+                    "files": [{"filename": f["filename"], "size": f["size"]} for f in dfiles],
+                }),
+                "title": guess_album,
+                "artist": guess_artist,
+                "track_count": len(dfiles),
+                "format": "FLAC" if lossless else "MP3",
+                "bitrate": None,
+            }
+            scored.append((has_slot, est_seconds, queue, candidate))
+    # ascending: free slot first, then shortest estimated transfer, then shortest queue
+    scored.sort(key=lambda t: (t[0], t[1], t[2]))
+    return [c for *_rest, c in scored]
+
+
 class SlskdClient:
     """Real Soulseek client talking to an slskd daemon's REST API.
 
@@ -58,11 +184,12 @@ class SlskdClient:
         r.raise_for_status()
         return r.json()
 
-    def search_album(self, artist: str, album: str) -> list[dict]:
+    def _run_search(self, text: str) -> list:
+        """POST a search, wait for it to complete, and return the raw responses (a list)."""
         r = requests.post(
             f"{self._url}/api/v0/searches",
             headers=self._headers(),
-            json={"searchText": f"{artist} {album}"},
+            json={"searchText": text},
             timeout=30,
         )
         r.raise_for_status()
@@ -75,43 +202,27 @@ class SlskdClient:
                 break
 
         responses = self._get(f"/api/v0/searches/{search_id}/responses")
-        candidates: list[dict] = []
-        for resp in responses:
-            username = resp.get("username", "")
-            by_dir: dict[str, list] = {}
-            for f in resp.get("files", []) or []:
-                name = f.get("filename", "")
-                ext = name.rsplit(".", 1)[-1].lower() if "." in name else ""
-                if ext not in _AUDIO_EXT:
+        return responses if isinstance(responses, list) else []
+
+    def search_album(self, artist: str, album: str) -> list[dict]:
+        responses = self._run_search(f"{artist} {album}")
+        if not responses:
+            # The Soulseek server silently drops searches containing certain blocklisted terms
+            # (major-label artists / album titles like Adele, Rihanna, Beyoncé, "Lemonade") →
+            # 0 responses. The blocked term poisons the whole query, but the OTHER field usually
+            # isn't blocked, so retry with each field alone (normalized) and keep only results
+            # whose path still matches the field we dropped. That recovers e.g. Beyoncé's
+            # "I Am… Sasha Fierce" via the album title, without ever grabbing a same-titled album
+            # by someone else.
+            for query, needle in ((album, artist), (artist, album)):
+                q = _norm(query)
+                if not q:
                     continue
-                by_dir.setdefault(_dirname(name), []).append(
-                    {"filename": name, "size": f.get("size", 0), "ext": ext}
-                )
-            for directory, dfiles in by_dir.items():
-                if not dfiles:
-                    continue
-                lossless = all(f["ext"] in _LOSSLESS_EXT for f in dfiles)
-                dirname = _basename(directory)
-                if " - " in dirname:
-                    guess_artist, guess_album = (p.strip() for p in dirname.split(" - ", 1))
-                else:
-                    guess_artist, guess_album = "", dirname
-                candidates.append(
-                    {
-                        "source_ref": json.dumps(
-                            {
-                                "username": username,
-                                "files": [{"filename": f["filename"], "size": f["size"]} for f in dfiles],
-                            }
-                        ),
-                        "title": guess_album,
-                        "artist": guess_artist,
-                        "track_count": len(dfiles),
-                        "format": "FLAC" if lossless else "MP3",
-                        "bitrate": None,
-                    }
-                )
-        return candidates
+                alt = _keep_matching(self._run_search(q), needle)
+                if alt:
+                    responses = alt
+                    break
+        return _parse_search_responses(responses, artist)
 
     def download(self, source_ref: str, dest: str, on_progress: Callable[[float], None]) -> dict:
         ref = json.loads(source_ref)
@@ -119,7 +230,17 @@ class SlskdClient:
         files = ref["files"]
         wanted = {f["filename"] for f in files}
         os.makedirs(dest, exist_ok=True)
+        xfer_ids: set[str] = set()  # slskd transfer ids we started, for cancel-on-abandon
+        try:
+            return self._download(username, files, wanted, dest, on_progress, xfer_ids)
+        except Exception:
+            # Abandoning this peer (timeout / error / the pipeline moving on): cancel the transfers
+            # we enqueued so they don't keep running and pile up in slskd as the job falls through
+            # to other peers/sources for the same album.
+            self._cancel_transfers(username, xfer_ids)
+            raise
 
+    def _download(self, username, files, wanted, dest, on_progress, xfer_ids: set) -> dict:
         r = requests.post(
             f"{self._url}/api/v0/transfers/downloads/{username}",
             headers=self._headers(),
@@ -130,23 +251,49 @@ class SlskdClient:
         on_progress(0.0)
 
         total = len(files)
+        total_bytes = sum(int(f.get("size") or 0) for f in files)
         completed = False
-        for _ in range(_XFER_POLLS):
+        last_bytes = -1
+        stall = 0
+        speed = 0.0  # bytes/sec, EWMA-smoothed so the ETA doesn't jitter poll to poll
+        prev_bytes = 0
+        prev_t = time.monotonic()
+        for _ in range(_max_polls(total_bytes)):
             time.sleep(_POLL_SECONDS)
             data = self._get(f"/api/v0/transfers/downloads/{username}")
             states: list[str] = []
+            bytes_now = 0
             for directory in data.get("directories", []) if isinstance(data, dict) else []:
                 for f in directory.get("files", []):
                     if f.get("filename") in wanted:  # only the files we enqueued
                         states.append(str(f.get("state", "")))
+                        bytes_now += int(f.get("bytesTransferred") or 0)
+                        if f.get("id"):  # remember the transfer so we can cancel it if abandoned
+                            xfer_ids.add(str(f["id"]))
             if states:
                 if any(bad in s for s in states for bad in ("Errored", "Rejected", "Cancelled")):
                     raise RuntimeError("soulseek transfer failed")
+                now = time.monotonic()
+                dt = now - prev_t
+                if dt > 0 and bytes_now >= prev_bytes:
+                    inst = (bytes_now - prev_bytes) / dt
+                    speed = inst if speed == 0.0 else 0.6 * speed + 0.4 * inst
+                prev_bytes, prev_t = bytes_now, now
                 done = sum(1 for s in states if "Completed" in s)
-                on_progress(min(1.0, done / max(total, 1)))
+                on_progress(min(1.0, done / max(total, 1)), _eta_seconds(total_bytes, bytes_now, speed))
                 if done >= total:
                     completed = True
                     break
+                # Stall detection: a peer making byte progress keeps its slot (up to the absolute
+                # backstop above) so a slow-but-working transfer can finish; one that sends no bytes
+                # for _STALL_POLLS (~90s) is queued/dead — abandon it fast and fall through.
+                if bytes_now > last_bytes:
+                    last_bytes = bytes_now
+                    stall = 0
+                else:
+                    stall += 1
+                    if stall >= _STALL_POLLS:
+                        raise TimeoutError(f"soulseek transfer stalled for {username}")
 
         if not completed:
             raise TimeoutError(f"soulseek download did not complete for {username}")
@@ -162,6 +309,19 @@ class SlskdClient:
             )
         return {"track_count": moved, "path": dest}
 
+    def _cancel_transfers(self, username: str, xfer_ids: set) -> None:
+        """Cancel the given in-flight slskd transfers for a peer. Best-effort: a cleanup failure
+        must never mask the original download failure."""
+        for tid in xfer_ids:
+            try:
+                requests.delete(
+                    f"{self._url}/api/v0/transfers/downloads/{username}/{tid}",
+                    headers=self._headers(),
+                    timeout=15,
+                )
+            except Exception:
+                pass
+
     def _retrieve(self, files: list[dict], dest: str) -> int:
         """Move the just-finished slskd files out of its downloads dir into `dest`.
 

worker/lyra_worker/adapters/_streamrip.py

@@ -34,6 +34,53 @@ def _run(coro):
     return _loop.run_until_complete(coro)
 
 
+class _ProgressAggregator:
+    """Turn streamrip's per-track byte callbacks into one album-level 0..1 fraction.
+
+    streamrip downloads an album's tracks concurrently and reports progress only per track
+    (a byte delta per chunk), so the file-count poller sees 0 complete files until they all
+    land at once — the 0%→100% jump. Summing downloaded bytes over an *estimated* album total
+    (average started-track size × track count) gives a smooth, generally-climbing fraction.
+    Track sizes are similar, so the estimate is stable once a track or two has started; capped
+    at 0.99 because the pipeline sets 1.0 once the whole download returns. All streamrip
+    coroutines run on one event loop on the calling thread, so no locking is needed."""
+
+    def __init__(self, track_count: int, on_progress: Callable[[float], None]):
+        self._n = track_count or 0
+        self._on_progress = on_progress
+        self._seen_total = 0   # summed byte-size of tracks that have started downloading
+        self._seen_count = 0   # number of tracks that have started
+        self._done_bytes = 0   # summed bytes downloaded across all started tracks
+
+    def start_track(self, total) -> None:
+        self._seen_total += max(int(total or 0), 0)
+        self._seen_count += 1
+
+    def advance(self, delta) -> None:
+        self._done_bytes += max(int(delta or 0), 0)
+        if self._seen_count == 0:
+            return
+        avg = self._seen_total / self._seen_count
+        n = self._n if self._n > 0 else self._seen_count
+        est_total = avg * n
+        if est_total > 0:
+            self._on_progress(min(self._done_bytes / est_total, 0.99))
+
+
+class _ProgressHandle:
+    """streamrip's `Track.download` uses `with get_progress_callback(...) as cb:`; `cb` must be
+    the per-chunk byte callback. This stand-in yields that callback and no-ops on exit."""
+
+    def __init__(self, update: Callable[[int], None]):
+        self.update = update
+
+    def __enter__(self):
+        return self.update
+
+    def __exit__(self, *_):
+        return False
+
+
 def _flatten_audio(dest: str) -> int:
     """streamrip writes into a nested `Artist - Album [...]/` subfolder. Move any audio
     files up into `dest`, remove the now-empty subfolders, and return the count of audio
@@ -136,6 +183,7 @@ class StreamripClient:
 
     async def _download(self, album_id: str, dest: str, on_progress: Callable[[float], None]) -> None:
         os.makedirs(dest, exist_ok=True)
+        import streamrip.media.track as _track_mod
         from streamrip.client import QobuzClient
         from streamrip.db import Database, Dummy
         from streamrip.media import PendingAlbum
@@ -153,7 +201,21 @@ class StreamripClient:
             album = await pending.resolve()
             if album is None:
                 raise RuntimeError(f"could not resolve Qobuz album {album_id}")
-            await album.rip()
+            # Hook streamrip's per-track byte progress into a single album-level fraction so the
+            # Floor bar climbs smoothly instead of jumping 0→100. `Track.download` binds
+            # get_progress_callback at import, so patch it on the track module (not progress.py).
+            agg = _ProgressAggregator(len(getattr(album, "tracks", []) or []), on_progress)
+            _orig_cb = _track_mod.get_progress_callback
+
+            def _hook(_enabled, total, _desc):
+                agg.start_track(total)
+                return _ProgressHandle(agg.advance)
+
+            _track_mod.get_progress_callback = _hook
+            try:
+                await album.rip()
+            finally:
+                _track_mod.get_progress_callback = _orig_cb
             on_progress(1.0)
         finally:
             await client.session.close()

worker/lyra_worker/adapters/fakes.py

@@ -72,17 +72,24 @@ class FailingAdapter(_BaseFake):
 
 
 class FakeMbBrowser:
-    """In-memory MbBrowser for tests. `releases` maps artist mbid -> [ReleaseGroupInfo]."""
+    """In-memory MbBrowser for tests. `releases` maps artist mbid -> [ReleaseGroupInfo].
+    `default_core=True` returns a synthetic core Album for any artist not in `releases`, so
+    scoring tests don't trip the "no core discography → drop the suggestion" filter."""
 
-    def __init__(self, artists=None, releases=None):
+    def __init__(self, artists=None, releases=None, default_core=False):
         self._artists = list(artists or [])
         self._releases = dict(releases or {})
+        self._default_core = default_core
 
     def search_artist(self, name: str) -> list[ArtistHit]:
         return list(self._artists)
 
     def browse_release_groups(self, artist_mbid: str) -> list[ReleaseGroupInfo]:
-        return list(self._releases.get(artist_mbid, []))
+        if artist_mbid in self._releases:
+            return list(self._releases[artist_mbid])
+        if self._default_core:
+            return [ReleaseGroupInfo(f"rg-{artist_mbid}", "Album", "Album", (), "2000-01-01")]
+        return []
 
 
 class FakeAudioProbe:

worker/lyra_worker/claim.py

@@ -24,13 +24,50 @@ def reclaim_stuck_jobs(conn: psycopg.Connection) -> int:
     return n
 
 
+_MAX_ATTEMPTS = 5
+
+
+def requeue_or_fail(
+    conn: psycopg.Connection, job_id: str, error: str, max_attempts: int = _MAX_ATTEMPTS
+) -> None:
+    """Recover a job whose pipeline raised an *unhandled* exception (e.g. a dropped MusicBrainz
+    TLS connection surfacing as an SSL EOF). Without this the job is left in its mid-pipeline
+    'matching' state; since claim_next only picks 'requested' it is orphaned until the next
+    restart's reclaim_stuck_jobs — showing "Searching sources…" on the press indefinitely.
+
+    Such errors are usually transient, so requeue to 'requested' until the attempt cap, then fall
+    to 'needs_attention' so a persistently-failing job stays visible instead of looping forever.
+    `attempts` is bumped by claim_next on every claim, so it already counts tries."""
+    with conn.cursor() as cur:
+        cur.execute('SELECT attempts, "requestId" FROM "Job" WHERE id = %s', (job_id,))
+        row = cur.fetchone()
+        if row is None:
+            return
+        attempts, request_id = row
+        if attempts >= max_attempts:
+            cur.execute(
+                'UPDATE "Job" SET state = \'needs_attention\', error = %s, "claimedAt" = NULL, '
+                '"updatedAt" = now() WHERE id = %s',
+                (error, job_id),
+            )
+            cur.execute('UPDATE "Request" SET status = \'needs_attention\' WHERE id = %s', (request_id,))
+        else:
+            cur.execute(
+                'UPDATE "Job" SET state = \'requested\', "currentStage" = \'intake\', error = NULL, '
+                '"claimedAt" = NULL, "downloadProgress" = 0, "updatedAt" = now() WHERE id = %s',
+                (job_id,),
+            )
+            cur.execute('UPDATE "Request" SET status = \'pending\' WHERE id = %s', (request_id,))
+    conn.commit()
+
+
 def claim_next(conn: psycopg.Connection) -> str | None:
     """Atomically claim the oldest queued job. Returns its id or None."""
     with conn.cursor() as cur:
         cur.execute(
             """
             SELECT id FROM "Job"
-            WHERE state = 'requested'
+            WHERE state = 'requested' AND NOT paused
             ORDER BY "createdAt" ASC
             FOR UPDATE SKIP LOCKED
             LIMIT 1

worker/lyra_worker/discovery.py

@@ -204,6 +204,16 @@ def _derive_albums(conn: psycopg.Connection, browser: MbBrowser, surfaced: list[
         except Exception as e:  # one artist's browse failure must not abort the sweep
             print(f"worker: discovery album browse failed for {artist['mbid']}: {e}", flush=True)
             continue
+        # Drop suggestions for "artists" with no real discography of their own — producers,
+        # session players, band members, soundtrack-only credits. ListenBrainz surfaces them
+        # (they co-occur in listens) but they have no core release-group to follow or want, and
+        # their page is empty. Require at least one core release (Album/Single/EP, no secondary
+        # type); on browse failure we keep the suggestion rather than drop a real artist.
+        if not any(is_core_release(g.primary_type, g.secondary_types) for g in groups):
+            with conn.cursor() as cur:
+                cur.execute("DELETE FROM \"DiscoverySuggestion\" WHERE status = 'pending' "
+                            'AND "artistMbid" = %s', (artist["mbid"],))
+            continue
         core = [g for g in groups
                 if _album_kind_ok(g, cfg.albums_only) and g.rg_mbid not in have]
         core.sort(key=lambda g: g.first_release_date or "", reverse=True)
@@ -217,6 +227,16 @@ def _derive_albums(conn: psycopg.Connection, browser: MbBrowser, surfaced: list[
         if popular is not None:
             reasons[popular] = "newest,popular" if reasons.get(popular) == "newest" else "popular"
 
+        # Re-deriving replaces this artist's album set: drop any pending album suggestion of
+        # theirs that is no longer a pick (stale singles/EPs from older sweeps, or an album now
+        # owned), keeping wanted/dismissed rows untouched. Then upsert the current picks.
+        with conn.cursor() as cur:
+            cur.execute(
+                "DELETE FROM \"DiscoverySuggestion\" WHERE kind = 'album' AND status = 'pending' "
+                'AND "artistMbid" = %s AND NOT ("rgMbid" = ANY(%s))',
+                (artist["mbid"], list(reasons.keys())),
+            )
+
         for rg in core:  # preserve newest-first order among the chosen
             reason = reasons.get(rg.rg_mbid)
             if reason is not None:
@@ -305,6 +325,18 @@ def _recompute_suggestions(conn: psycopg.Connection, affected: set[str],
     return surfaced
 
 
+def reset_pending(conn: psycopg.Connection) -> None:
+    """Clear discovery state so the next sweep rebuilds from scratch with current logic. Drops
+    all PENDING suggestions (followed/wanted/dismissed user decisions are kept), every seed
+    contribution + library-seed throttle, and marks each watched artist due again."""
+    with conn.cursor() as cur:
+        cur.execute("DELETE FROM \"DiscoverySuggestion\" WHERE status = 'pending'")
+        cur.execute('DELETE FROM "DiscoverySeedContribution"')
+        cur.execute('DELETE FROM "DiscoverySeed"')
+        cur.execute('UPDATE "WatchedArtist" SET "lastDiscoveredAt" = NULL')
+    conn.commit()
+
+
 def run_discovery(conn: psycopg.Connection, sources: list[SimilaritySource],
                   browser: MbBrowser, cfg: DiscoveryConfig, popularity=None) -> DiscoveryResult:
     """Record each seed's similar-artist contributions, then recompute affected

worker/lyra_worker/floor.py

@@ -0,0 +1,42 @@
+import os
+
+from lyra_worker.config import get_config
+
+_TRUE = {"1", "true", "yes", "on"}
+
+
+def press_paused(config: dict[str, str]) -> bool:
+    """True when the whole press is paused — the worker stops claiming new jobs."""
+    return str(config.get("floor.paused", "")).strip().lower() in _TRUE
+
+
+def job_delay_seconds(config: dict[str, str]) -> float:
+    """Seconds to wait between finishing one download and claiming the next. Invalid or
+    absent -> 0.0 (back-to-back). Negatives are clamped to 0."""
+    try:
+        return max(0.0, float(int(config.get("floor.jobDelaySeconds", "") or 0)))
+    except (TypeError, ValueError):
+        return 0.0
+
+
+def kill_requested(config: dict[str, str]) -> bool:
+    """True when a force-stop of the in-flight download has been requested."""
+    return str(config.get("floor.killRequested", "")).strip().lower() in _TRUE
+
+
+def consume_kill_switch(conn, exit_fn=os._exit) -> bool:
+    """If a force-stop is pending, clear the flag and terminate the process (default
+    os._exit) so the in-flight download is aborted. The container restart policy brings the
+    worker back; startup clears staging and requeues the aborted job. Returns whether it
+    fired. exit_fn is injectable for tests."""
+    config = get_config(conn)
+    if not kill_requested(config):
+        return False
+    with conn.cursor() as cur:
+        cur.execute(
+            'UPDATE "Config" SET value = \'false\', "updatedAt" = now() WHERE key = %s',
+            ("floor.killRequested",),
+        )
+    conn.commit()
+    exit_fn(0)
+    return True

worker/lyra_worker/main.py

@@ -6,11 +6,12 @@ import uuid
 
 import psycopg
 
-from lyra_worker.claim import claim_next, reclaim_stuck_jobs
+from lyra_worker.claim import claim_next, reclaim_stuck_jobs, requeue_or_fail
 from lyra_worker.config import get_config
 from lyra_worker.crypto import secret_key_problem
 from lyra_worker.db import wait_for_db
-from lyra_worker.discovery import DiscoveryConfig, DiscoveryResult, run_discovery
+from lyra_worker.discovery import DiscoveryConfig, DiscoveryResult, reset_pending, run_discovery
+from lyra_worker.floor import press_paused, job_delay_seconds, consume_kill_switch
 from lyra_worker.library import clear_staging_root
 from lyra_worker.monitor import MonitorConfig, fulfill_owned_releases, reconcile, sweep
 from lyra_worker.pipeline import run_pipeline
@@ -147,6 +148,10 @@ def maybe_run_discovery(conn, sources, browser, config, dcfg, now, last_discover
 
     starting = not in_progress
     if starting:  # begin a fresh sweep
+        if str(config.get("discover.resetRequested", "")).strip().lower() in _TRUE:
+            reset_pending(conn)  # wipe stale pending suggestions before rebuilding
+            _set_config(conn, "discover.resetRequested", "false")
+            print("worker: discovery reset — cleared pending suggestions, rebuilding", flush=True)
         _set_config(conn, "discover.inProgress", "true")
         _set_config(conn, "discover.requested", "false")
 
@@ -222,9 +227,30 @@ def _liveness_heartbeat() -> None:
         time.sleep(HEARTBEAT_SECONDS)
 
 
+KILL_POLL_SECONDS = 1.5
+
+
+def _kill_switch_watch() -> None:
+    """Poll floor.killRequested on a background daemon thread (its own DB connection) so a
+    force-stop aborts the process even while the main loop is blocked in a download. On any
+    DB hiccup, reconnect and keep watching."""
+    conn = wait_for_db()
+    while True:
+        try:
+            consume_kill_switch(conn)
+        except Exception:
+            try:
+                conn.close()
+            except Exception:
+                pass
+            conn = wait_for_db()
+        time.sleep(KILL_POLL_SECONDS)
+
+
 def run_forever() -> None:
     _require_secret_key()
     threading.Thread(target=_liveness_heartbeat, daemon=True).start()
+    threading.Thread(target=_kill_switch_watch, daemon=True).start()
     conn = wait_for_db()
     clear_staging_root(STAGING_ROOT)  # sweep any staging dirs orphaned by a prior crash
     reclaimed = reclaim_stuck_jobs(conn)  # requeue jobs a prior crash left mid-pipeline
@@ -248,6 +274,7 @@ def run_forever() -> None:
     last_discover_tick = 0.0
     last_heartbeat = 0.0
     last_prune = 0.0
+    next_job_allowed = 0.0  # monotonic time before which no new job is claimed (inter-job delay)
     try:
         while True:
             # A transient DB connection loss (e.g. AdminShutdown when Postgres restarts) raises
@@ -279,6 +306,12 @@ def run_forever() -> None:
 
                 maybe_run_scan(conn, resolver, probe, browser, config)
 
+                # Press paused (downloads only) or the inter-job delay hasn't elapsed: keep the
+                # loop alive (monitor/scan/discovery above still run) but claim nothing.
+                if press_paused(config) or now < next_job_allowed:
+                    time.sleep(IDLE_SLEEP)
+                    continue
+
                 job_id = claim_next(conn)
                 if job_id is None:
                     time.sleep(IDLE_SLEEP)
@@ -290,8 +323,12 @@ def run_forever() -> None:
                 except Exception as e:  # one bad job must not take down the worker loop
                     print(f"worker: pipeline failed for job {job_id}: {e}", flush=True)
                     conn.rollback()
+                    # Don't orphan the job in its mid-pipeline 'matching' state (claim_next only
+                    # picks 'requested'): requeue for a bounded retry, then needs_attention.
+                    requeue_or_fail(conn, job_id, str(e))
                 finish_job(conn, job_id, mcfg)
                 print(f"worker: finished job {job_id}", flush=True)
+                next_job_allowed = time.monotonic() + job_delay_seconds(config)
             except psycopg.OperationalError as e:
                 print(f"worker: db connection lost ({e}); reconnecting...", flush=True)
                 try:

worker/lyra_worker/monitor.py

@@ -86,16 +86,40 @@ def _set_state(conn: psycopg.Connection, rel_id: str, state: str) -> None:
         cur.execute('UPDATE "MonitoredRelease" SET state = %s WHERE id = %s', (state, rel_id))
 
 
-def _in_library_at_cutoff(conn: psycopg.Connection, artist: str, album: str, cutoff: int) -> bool:
+def _in_library_at_cutoff(
+    conn: psycopg.Connection, artist: str, album: str, rg_mbid: str, cutoff: int
+) -> bool:
+    """True if the album is owned at/above `cutoff`. Matches on the release-group MBID when the
+    library row has one — robust against MusicBrainz artist-name canonicalization (e.g. the
+    followed 'Kanye West' imports as 'Ye') — and falls back to exact artist+album for scanned
+    rows that predate MBID capture."""
     with conn.cursor() as cur:
         cur.execute(
-            'SELECT 1 FROM "LibraryItem" WHERE artist = %s AND album = %s '
-            'AND "qualityClass" >= %s LIMIT 1',
-            (artist, album, cutoff),
+            'SELECT 1 FROM "LibraryItem" WHERE "qualityClass" >= %s AND ('
+            '  ("rgMbid" IS NOT NULL AND "rgMbid" = %s) '
+            '  OR ("rgMbid" IS NULL AND artist = %s AND album = %s)) LIMIT 1',
+            (cutoff, rg_mbid, artist, album),
         )
         return cur.fetchone() is not None
 
 
+def _retire_failed_attempts(conn: psycopg.Connection, rel_id: str) -> None:
+    """Delete prior failed (needs_attention) Request rows for this release (Job/Candidate cascade).
+
+    A hard-to-match release is re-pressed every retry interval; each failed attempt used to leave
+    a needs_attention Request behind, piling up phantom "needs attention" cards on the press even
+    after a later attempt imported the album. Retiring them when the release is fulfilled or
+    re-attempted keeps the press showing one live attempt at a time. Manual (non-monitor) requests
+    have no monitoredReleaseId and are never touched."""
+    with conn.cursor() as cur:
+        cur.execute(
+            'DELETE FROM "Request" r USING "Job" j '
+            'WHERE j."requestId" = r.id AND r."monitoredReleaseId" = %s '
+            "AND j.state = 'needs_attention'",
+            (rel_id,),
+        )
+
+
 def _enqueue_one(conn: psycopg.Connection, rel_id: str, artist: str, album: str) -> None:
     with conn.cursor() as cur:
         cur.execute(
@@ -118,7 +142,8 @@ def enqueue_due(conn: psycopg.Connection, cfg: MonitorConfig) -> int:
         cur.execute(
             'SELECT mr.id, mr."artistName", mr.album, mr.state, mr."currentQualityClass", '
             '  (mr."firstGrabbedAt" IS NOT NULL '
-            '   AND mr."firstGrabbedAt" < now() - make_interval(days => %s)) AS window_expired '
+            '   AND mr."firstGrabbedAt" < now() - make_interval(days => %s)) AS window_expired, '
+            '  mr."rgMbid" '
             'FROM "MonitoredRelease" mr '
             'WHERE mr.monitored = true '
             '  AND NOT mr.ignored '  # user-ignored releases are never enqueued
@@ -134,8 +159,11 @@ def enqueue_due(conn: psycopg.Connection, cfg: MonitorConfig) -> int:
         rows = cur.fetchall()
 
     enqueued = 0
-    for rel_id, artist, album, state, quality, window_expired in rows:
-        if _in_library_at_cutoff(conn, artist, album, cfg.quality_cutoff):
+    for rel_id, artist, album, state, quality, window_expired, rg_mbid in rows:
+        # A due release is about to be fulfilled or re-attempted; either way any earlier failed
+        # attempt is superseded, so retire it before deciding.
+        _retire_failed_attempts(conn, rel_id)
+        if _in_library_at_cutoff(conn, artist, album, rg_mbid, cfg.quality_cutoff):
             _set_state(conn, rel_id, "fulfilled")
             continue
         if state == "grabbed" and (window_expired or (quality is not None and quality >= cfg.quality_cutoff)):
@@ -154,12 +182,19 @@ def fulfill_owned_releases(conn: psycopg.Connection, cfg: MonitorConfig) -> int:
     with conn.cursor() as cur:
         cur.execute(
             'UPDATE "MonitoredRelease" mr SET state = \'fulfilled\', '
-            '  "currentQualityClass" = li.q, '
+            '  "currentQualityClass" = owned.q, '
             '  "firstGrabbedAt" = COALESCE(mr."firstGrabbedAt", now()) '
-            'FROM (SELECT artist, album, max("qualityClass") AS q FROM "LibraryItem" '
-            '      GROUP BY artist, album) li '
-            'WHERE li.artist = mr."artistName" AND li.album = mr.album '
-            "  AND mr.monitored = true AND mr.state <> 'fulfilled' AND li.q >= %s",
+            'FROM ('
+            '  SELECT mr2.id AS mid, max(li."qualityClass") AS q '
+            '  FROM "MonitoredRelease" mr2 JOIN "LibraryItem" li ON ('
+            # match on rgMbid where the library row has one (survives artist-name canonicalization),
+            # else fall back to exact artist+album for scanned rows lacking an MBID
+            '       (li."rgMbid" IS NOT NULL AND li."rgMbid" = mr2."rgMbid") '
+            '    OR (li."rgMbid" IS NULL AND li.artist = mr2."artistName" AND li.album = mr2.album)) '
+            "  WHERE mr2.monitored = true AND mr2.state <> 'fulfilled' "
+            '  GROUP BY mr2.id HAVING max(li."qualityClass") >= %s'
+            ') owned '
+            'WHERE mr.id = owned.mid',
             (cfg.quality_cutoff,),
         )
         fixed = cur.rowcount
@@ -171,21 +206,28 @@ def reconcile(conn: psycopg.Connection, job_id: str, cfg: MonitorConfig) -> None
     """Feed a finished job's outcome back into its MonitoredRelease (if any)."""
     with conn.cursor() as cur:
         cur.execute(
-            'SELECT j.state, r."monitoredReleaseId", r.artist, r.album '
-            'FROM "Job" j JOIN "Request" r ON r.id = j."requestId" WHERE j.id = %s',
+            'SELECT j.state, r."monitoredReleaseId", r.artist, r.album, mr."rgMbid" '
+            'FROM "Job" j JOIN "Request" r ON r.id = j."requestId" '
+            'LEFT JOIN "MonitoredRelease" mr ON mr.id = r."monitoredReleaseId" '
+            'WHERE j.id = %s',
             (job_id,),
         )
         row = cur.fetchone()
     if row is None:
         return
-    state, rel_id, artist, album = row
+    state, rel_id, artist, album, rg_mbid = row
     if rel_id is None or state != "imported":
         return  # needs_attention (or in-flight) leaves the release wanted
 
+    # A successful import supersedes any earlier failed attempts for this release.
+    _retire_failed_attempts(conn, rel_id)
+
     with conn.cursor() as cur:
         cur.execute(
-            'SELECT "qualityClass" FROM "LibraryItem" WHERE artist = %s AND album = %s',
-            (artist, album),
+            'SELECT max("qualityClass") FROM "LibraryItem" WHERE '
+            '  ("rgMbid" IS NOT NULL AND "rgMbid" = %s) '
+            '  OR ("rgMbid" IS NULL AND artist = %s AND album = %s)',
+            (rg_mbid, artist, album),
         )
         lib = cur.fetchone()
     quality = lib[0] if lib else None

worker/lyra_worker/pipeline.py

@@ -1,7 +1,9 @@
 import os
+import random
 import shutil
 import threading
 import time
+from concurrent.futures import ThreadPoolExecutor
 from dataclasses import replace
 from typing import Callable, Sequence
 
@@ -10,6 +12,7 @@ import psycopg
 from lyra_worker.adapters.base import SourceAdapter
 from lyra_worker.confidence import score_confidence
 from lyra_worker.library import album_dir, import_album, list_audio_files, staging_dir
+from lyra_worker.qobuz_gate import gate_open, record_download
 from lyra_worker.quality import quality_class
 from lyra_worker.ranker import rank_candidates
 from lyra_worker.types import Candidate, MBTarget
@@ -19,8 +22,19 @@ _AUDIO_EXT = {".flac", ".mp3", ".m4a", ".opus", ".ogg", ".aac", ".wav"}
 # Reject a download whose total playtime falls below this fraction of MusicBrainz's expected
 # total — catches truncated files / preview clips substituted for real tracks that keep the
 # track COUNT right. Generous, so edition/encoding differences don't false-positive; bonus
-# tracks (over-long) are always fine.
-_DURATION_MIN_RATIO = 0.85
+# tracks (over-long) are always fine. Kept well below 1.0 because a source legitimately delivers a
+# different edition than MB's resolved release (shorter radio edits, no bonus/live tracks), and only
+# a genuinely truncated/preview-filled download (playtime a fraction of expected) should be caught.
+_DURATION_MIN_RATIO = 0.65
+# Cap how many ranked candidates a single job will actually download+verify before giving up.
+# Without this, the incomplete-download fall-through can grind through dozens of sources (a popular
+# album returns ~200 Soulseek candidates), each a slow attempt that also leaves an abandoned slskd
+# transfer. The best candidates rank first, so 6 attempts is plenty.
+_MAX_DOWNLOAD_ATTEMPTS = 6
+# The best quality any non-Qobuz source offers (Soulseek FLAC = CD-lossless). A library copy at or
+# above this can only be improved by hi-res Qobuz, so when Qobuz is gated by pacing an upgrade of
+# such a copy is deferred without even searching.
+_CD_LOSSLESS_CLASS = 2
 
 
 def _measure_staged_duration_s(staging: str) -> float | None:
@@ -69,6 +83,40 @@ def _count_staged_audio(staging: str) -> int:
     return n
 
 
+def _download_problem(result, staging: str, target, winner, measure_duration) -> str | None:
+    """Why a completed download is unusable — ``"incomplete download"`` / ``"download too short
+    (...)"`` — or None if it's good. Completeness is judged against the CHOSEN SOURCE's own track
+    count (``winner.track_count``), NOT MusicBrainz's canonical release: MB picks an arbitrary
+    releases[0] that is often a deluxe/expanded edition with more tracks (and runtime) than the
+    standard album a source legitimately delivers, so gating on ``target.track_count`` falsely
+    rejects complete standard-edition downloads (The Script "No Sound Without Silence" 11 vs MB 12,
+    Skillet "Unleashed" 12 vs MB's 20-track deluxe). Pure — no DB I/O — so the download loop can
+    call it per candidate and fall through to the next source on a problem."""
+    expected = winner.track_count
+    # (a) Truncated: fewer files than the source promised (a track failed mid-download). Count the
+    # NON-EMPTY audio on disk too: a silently-skipped track leaves a 0-byte placeholder the adapter
+    # still counts. (`staged and ...` keeps fake adapters that stage nothing on the count path.)
+    staged = _count_staged_audio(staging)
+    truncated = result.track_count < expected or bool(staged and staged < expected)
+    # (b) Implausibly small vs MB's album — a lone-track "full album" video / wrong match, as
+    # distinct from a legitimately smaller edition. A real edition keeps more than half MB's tracks.
+    too_small = target.track_count is not None and expected * 2 <= target.track_count
+    if truncated or too_small:
+        return "incomplete download"
+    # Duration: a truncated file or short preview substituted for a real track shows up as playtime
+    # well under the expected total. Scale MB's total to the DELIVERED edition's size so a smaller
+    # edition isn't judged against a larger one's runtime. Only when the total is known (measured is
+    # None for the offline fakes).
+    if target.total_duration_s:
+        expected_total = target.total_duration_s
+        if target.track_count:
+            expected_total *= winner.track_count / target.track_count
+        measured = measure_duration(staging)
+        if measured is not None and measured < expected_total * _DURATION_MIN_RATIO:
+            return f"download too short ({measured / 60:.0f} of ~{expected_total / 60:.0f} min)"
+    return None
+
+
 def _poll_download_progress(job_id: str, staging: str, expected: int, stop: "threading.Event",
                             reports: "threading.Event") -> None:
     """Until `stop`, periodically write Job.downloadProgress = files-in-staging / expected
@@ -94,6 +142,16 @@ def _poll_download_progress(job_id: str, staging: str, expected: int, stop: "thr
         conn.close()
 
 
+def _search_adapter(adapter: SourceAdapter, target: MBTarget) -> list[Candidate]:
+    """Search one adapter, swallowing failures — a down source contributes no candidates rather
+    than crashing the job. Runs on a worker thread (searches are parallelized across sources)."""
+    try:
+        return list(adapter.search(target))
+    except Exception as e:
+        print(f"pipeline: adapter {adapter.name} search failed: {e}", flush=True)
+        return []
+
+
 def _set_state(conn: psycopg.Connection, job_id: str, state: str, stage: str) -> None:
     with conn.cursor() as cur:
         cur.execute(
@@ -103,30 +161,42 @@ def _set_state(conn: psycopg.Connection, job_id: str, state: str, stage: str) ->
     conn.commit()
 
 
-def _set_download_progress(conn: psycopg.Connection, job_id: str, frac: float) -> None:
+def _set_download_progress(
+    conn: psycopg.Connection, job_id: str, frac: float, eta_seconds: int | None = None
+) -> None:
     with conn.cursor() as cur:
-        cur.execute('UPDATE "Job" SET "downloadProgress" = %s WHERE id = %s', (frac, job_id))
+        cur.execute(
+            'UPDATE "Job" SET "downloadProgress" = %s, "downloadEtaSeconds" = %s WHERE id = %s',
+            (frac, eta_seconds, job_id),
+        )
     conn.commit()
 
 
 def _make_on_progress(conn: psycopg.Connection, job_id: str, reports: "threading.Event"):
-    """Build the on_progress(pct) callback threaded into adapter.download. It marks the
-    adapter as reporting real byte-level progress (so the file-count poller yields) and
-    throttle-writes Job.downloadProgress. Called synchronously from adapter.download on the
-    pipeline's own thread, so it safely reuses `conn`."""
+    """Build the on_progress(pct, eta_seconds=None) callback threaded into adapter.download. It
+    marks the adapter as reporting real byte-level progress (so the file-count poller yields) and
+    throttle-writes Job.downloadProgress (+ the download ETA when the adapter measures one — slskd
+    does; others pass None). Called synchronously from adapter.download on the pipeline's own
+    thread, so it safely reuses `conn`."""
     state = {"frac": -1.0, "t": 0.0}
 
-    def on_progress(pct: float) -> None:
+    def on_progress(pct: float, eta_seconds: int | None = None) -> None:
         reports.set()
         frac = 0.0 if pct < 0 else 1.0 if pct > 1 else float(pct)
         now = time.monotonic()
-        # throttle DB writes: on a >=1% move or every 0.5s, not once per received byte
-        if frac - state["frac"] >= 0.01 or (now - state["t"]) >= 0.5:
+        # throttle DB writes: on a >=1% move, a new ETA, or every 0.5s — not once per received byte
+        if frac - state["frac"] >= 0.01 or eta_seconds is not None or (now - state["t"]) >= 0.5:
             state["frac"] = frac
             state["t"] = now
             try:
-                _set_download_progress(conn, job_id, frac)
+                _set_download_progress(conn, job_id, frac, eta_seconds)
             except Exception as e:  # a progress write must never fail the download
+                # Roll back so a failed write can't leave the shared pipeline conn in an aborted
+                # txn (which would cascade "current transaction is aborted" into every later query).
+                try:
+                    conn.rollback()
+                except Exception:
+                    pass
                 print(f"pipeline: on_progress write failed: {e}", flush=True)
 
     return on_progress
@@ -214,6 +284,17 @@ def _library_quality(conn: psycopg.Connection, target: MBTarget) -> int | None:
         return row[0] if row else None
 
 
+def _keep_existing_and_complete(conn: psycopg.Connection, job_id: str) -> None:
+    """Finish a job without importing — keep the existing library copy — and mark its request
+    completed. Used when a (deferred) upgrade has nothing better to offer right now; the release
+    stays below the cutoff and is re-attempted later."""
+    _set_state(conn, job_id, "imported", "import")
+    with conn.cursor() as cur:
+        cur.execute("UPDATE \"Request\" SET status = 'completed' WHERE id = %s",
+                    (_request_id(conn, job_id),))
+    conn.commit()
+
+
 def _persist_candidates(conn: psycopg.Connection, job_id: str, candidates: list[Candidate]) -> None:
     with conn.cursor() as cur:
         for c in candidates:
@@ -306,19 +387,43 @@ def run_pipeline(
         conn.commit()
         return
 
-    # 2. match — search all adapters; score + persist EVERY candidate found
+    # Pre-search upgrade skip: a >= CD-lossless copy can only be improved by hi-res Qobuz. If Qobuz
+    # is gated by pacing, don't even search (no other source can beat it) — defer to when Qobuz
+    # frees up. This avoids re-searching every already-owned album (~2 min Soulseek search each) on
+    # every monitor cycle. A user-forced re-acquire always proceeds.
+    _existing_q = _library_quality(conn, target)
+    if _existing_q is not None and _existing_q >= _CD_LOSSLESS_CLASS \
+            and not _is_force_job(conn, job_id) and not gate_open(conn):
+        _keep_existing_and_complete(conn, job_id)
+        return
+
+    # 2. match — search all adapters CONCURRENTLY (Soulseek's search polls up to ~2 min; run it
+    # alongside Qobuz's instant search so total match time = the slowest source, not their sum).
+    # Only Qobuz touches streamrip's shared asyncio loop and there's a single Qobuz adapter, so no
+    # two threads drive that loop at once. Results are collected in adapter order (map preserves
+    # input order) so candidate ordering stays deterministic.
     _set_state(conn, job_id, "matching", "match")
     found: list[Candidate] = []
-    for adapter in adapters:
-        try:
-            results = adapter.search(target)
-        except Exception as e:  # a down source contributes no candidates, never crashes the job
-            print(f"pipeline: adapter {adapter.name} search failed: {e}", flush=True)
-            continue
+    with ThreadPoolExecutor(max_workers=max(1, len(adapters))) as pool:
+        per_adapter = list(pool.map(lambda a: (a, _search_adapter(a, target)), adapters))
+    for adapter, results in per_adapter:
         for c in results:
             found.append(replace(c, source_tier=adapter.tier, confidence=score_confidence(target, c)))
+    # Drop candidates that can't be the full album — a result with far fewer tracks than the release
+    # has (e.g. a single-song Soulseek folder named like the album). Same threshold as the
+    # completeness 'too small' guard, applied up front so we never waste a download attempt on it.
+    # Unknown counts (0) and legitimately smaller editions (e.g. 11 vs 12) are kept.
+    if target.track_count:
+        found = [c for c in found if c.track_count == 0 or c.track_count * 2 > target.track_count]
     _persist_candidates(conn, job_id, found)
 
+    # Qobuz pacing: when the budget gate is closed (off-hours / daily cap / spacing), drop Qobuz
+    # candidates so the job falls through to another source and Qobuz stays under the radar — UNLESS
+    # Qobuz is the only option (nothing to fall through to, so don't strand the album).
+    if any(c.source == "qobuz" for c in found) and any(c.source != "qobuz" for c in found):
+        if not gate_open(conn):
+            found = [c for c in found if c.source != "qobuz"]
+
     # 3. rank
     _set_state(conn, job_id, "matched", "rank")
     ranked = rank_candidates(target, found, min_confidence)
@@ -326,12 +431,23 @@ def run_pipeline(
         _fail(conn, job_id, "no candidate above confidence threshold")
         return
 
+    # Upgrade guard: for an album already in the library, skip the download when no currently-
+    # available source beats the copy we have (ranked[0] is the highest-quality candidate). This
+    # happens when Qobuz is gated by pacing and only a same-quality Soulseek copy is left — the
+    # download would be discarded by keep-if-better, so defer it. The release stays below the cutoff
+    # and is re-attempted when a better source frees up. A user-forced re-acquire always proceeds.
+    existing_q = _library_quality(conn, target)
+    if existing_q is not None and not _is_force_job(conn, job_id) \
+            and quality_class(ranked[0].quality) <= existing_q:
+        _keep_existing_and_complete(conn, job_id)  # keep the existing copy; nothing better available
+        return
+
     # 4. download (fall-through) into an isolated per-job staging dir
     _set_state(conn, job_id, "downloading", "download")
     by_source = {a.name: a for a in adapters}
     staging = staging_dir(staging_root or f"{dest_root}/.staging", job_id)
     winner = None
-    result = None
+    last_problem = None  # why the most recent downloaded-but-rejected candidate was unusable
     _set_download_progress(conn, job_id, 0.0)  # reset for this run
     expected = target.track_count or (ranked[0].track_count if ranked else 0)
     _stop = threading.Event()
@@ -342,46 +458,46 @@ def run_pipeline(
     _poller.start()
     try:
         try:
+            attempts = 0
             for candidate in ranked:
                 adapter = by_source.get(candidate.source)
                 if adapter is None:
                     continue
+                if attempts >= _MAX_DOWNLOAD_ATTEMPTS:
+                    break  # don't grind through every peer/source of a popular album
+                attempts += 1
                 _mark_chosen(conn, job_id, candidate.source_ref)  # reflect the source now in flight
                 shutil.rmtree(staging, ignore_errors=True)  # clean slate per attempt
                 _reports.clear()  # this attempt hasn't reported yet → poller estimates until it does
                 result = adapter.download(candidate, staging, _make_on_progress(conn, job_id, _reports))
-                if result.ok:
+                if not result.ok:
+                    continue
+                # Verify completeness on the staging copy. A download that lands incomplete or too
+                # short falls through to the NEXT-ranked source (e.g. Qobuz repeatedly fails one
+                # track → try the deluxe edition or Soulseek) rather than failing the whole job on
+                # the first source's shortfall.
+                last_problem = _download_problem(result, staging, target, candidate, measure_duration)
+                if last_problem is None:
                     winner = candidate
                     break
         finally:
             _stop.set()
             _poller.join(timeout=3)
-        if winner is None or result is None or not result.ok:
-            _fail(conn, job_id, "all downloads failed")
+        if winner is None:
+            # No source produced a complete album: surface why the last one was rejected (incomplete
+            # / too short), or "all downloads failed" if none even downloaded.
+            _fail(conn, job_id, last_problem or "all downloads failed")
             return
         _set_download_progress(conn, job_id, 1.0)  # download done — UI shows 100% into Finishing
+        if winner.source == "qobuz":
+            # Count this Qobuz download against today's budget and set the next randomized spacing.
+            try:
+                record_download(conn, random.random())
+            except Exception as e:  # pacing bookkeeping must never fail a good download
+                print(f"pipeline: qobuz pacing record failed for job {job_id}: {e}", flush=True)
 
-        # 5. verify completeness on the staging copy, then promote + tag
+        # 5. promote + tag the verified winner
         _set_state(conn, job_id, "tagging", "tag")
-        expected = target.track_count if target.track_count is not None else winner.track_count
-        # Also count the NON-EMPTY audio actually on disk: a silently-skipped track leaves a
-        # 0-byte placeholder that the adapter still counts, so trust the files when any are
-        # staged. (`staged and ...` keeps fake adapters that stage nothing on the count path.)
-        staged = _count_staged_audio(staging)
-        if result.track_count < expected or (staged and staged < expected):
-            _fail(conn, job_id, "incomplete download")
-            return
-
-        # Duration check: even with the right track count, a truncated file or a short preview
-        # substituted for a real track shows up as a total playtime well under MB's expected
-        # total. Only applies when both are known (measured is None for the offline fakes).
-        if target.total_duration_s:
-            measured = measure_duration(staging)
-            if measured is not None and measured < target.total_duration_s * _DURATION_MIN_RATIO:
-                _fail(conn, job_id,
-                      f"download too short ({measured / 60:.0f} of ~{target.total_duration_s / 60:.0f} min)")
-                return
-
         final = album_dir(dest_root, target)
         existing_q = _library_quality(conn, target)
         new_q = quality_class(winner.quality)
@@ -395,12 +511,7 @@ def run_pipeline(
                     print(f"pipeline: tagging failed for job {job_id}: {e}", flush=True)
             _import(conn, job_id, target, winner, final)
         else:
-            # an existing copy is already at >= this quality — keep it untouched, just complete the request
-            with conn.cursor() as cur:
-                cur.execute(
-                    "UPDATE \"Request\" SET status = 'completed' WHERE id = %s",
-                    (_request_id(conn, job_id),),
-                )
-            conn.commit()
+            # an existing copy is already at >= this quality — keep it untouched, just complete
+            _keep_existing_and_complete(conn, job_id)
     finally:
         shutil.rmtree(staging, ignore_errors=True)  # a partial/failed download never persists

worker/lyra_worker/qobuz_gate.py

@@ -0,0 +1,159 @@
+"""Qobuz pacing ("budget gate").
+
+Keep using Qobuz for hi-res, but at a human-looking volume and cadence so the account isn't
+flagged for bulk downloading (which got it 403 USER_BLOCKED). One gate decides, per album,
+whether Qobuz is an eligible source right now: within active hours, under today's cap (a warm-up
+ramp of a steady cap), and past a randomized spacing since the last Qobuz download. When the gate
+is closed the pipeline drops Qobuz candidates and falls through to the next source; monitored
+albums upgrade to Qobuz later once budget frees up.
+
+Pure decision logic (`today_cap`, `is_allowed`, `next_gap_seconds`) is separated from the two DB
+helpers (`gate_open`, `record_download`) so the policy is unit-tested without a clock or DB.
+"""
+from dataclasses import dataclass
+from datetime import date
+
+_TRUE = {"1", "true", "yes", "on"}
+_JITTER = 0.3   # ± fraction around the spread interval
+_GAP_FLOOR_S = 60.0
+
+
+@dataclass(frozen=True)
+class QobuzPacing:
+    enabled: bool = True
+    active_start: int = 8
+    active_end: int = 23
+    daily_cap: int = 40
+    warmup_start_date: str = ""  # ISO "YYYY-MM-DD", or "" for no warm-up ramp
+
+    @classmethod
+    def from_config(cls, config: dict) -> "QobuzPacing":
+        def _int(key, default):
+            try:
+                return int(config[key])
+            except (KeyError, TypeError, ValueError):
+                return default
+
+        return cls(
+            enabled=str(config.get("qobuz.pacing.enabled", "true")).strip().lower() in _TRUE,
+            active_start=_int("qobuz.pacing.activeStartHour", 8),
+            active_end=_int("qobuz.pacing.activeEndHour", 23),
+            daily_cap=_int("qobuz.pacing.dailyCap", 40),
+            warmup_start_date=str(config.get("qobuz.pacing.warmupStartDate", "") or "").strip(),
+        )
+
+
+def today_cap(pacing: QobuzPacing, today: date) -> int:
+    """Today's Qobuz download cap: the steady cap, ramped up by account age (days since
+    ``warmup_start_date``). No warm-up date -> steady cap. A small steady cap is never exceeded."""
+    steady = max(0, pacing.daily_cap)
+    if not pacing.warmup_start_date:
+        return steady
+    try:
+        start = date.fromisoformat(pacing.warmup_start_date)
+    except ValueError:
+        return steady
+    days = max(0, (today - start).days)
+    if days < 3:
+        return min(5, steady)
+    if days < 7:
+        return min(10, steady)
+    if days < 14:
+        return min(20, steady)
+    return steady
+
+
+def is_allowed(pacing: QobuzPacing, now_hour: int, count_today: int, cap: int,
+               now_epoch: float, next_allowed_epoch: float) -> bool:
+    """Pure gate decision. Qobuz is eligible only within active hours, under today's cap, and past
+    the randomized spacing since the last download. Pacing disabled -> always allowed."""
+    if not pacing.enabled:
+        return True
+    if not (pacing.active_start <= now_hour < pacing.active_end):
+        return False
+    if count_today >= cap:
+        return False
+    if now_epoch < next_allowed_epoch:
+        return False
+    return True
+
+
+def next_gap_seconds(pacing: QobuzPacing, cap: int, rand01: float) -> float:
+    """Seconds until the next Qobuz download is allowed: the day's budget spread across the active
+    window (``activeWindow / cap``), ± _JITTER, floored at _GAP_FLOOR_S. ``rand01`` in [0,1] is the
+    injected randomness (so tests are deterministic; callers pass random.random())."""
+    window_s = max(1, pacing.active_end - pacing.active_start) * 3600
+    base = window_s / max(1, cap)
+    lo, hi = base * (1 - _JITTER), base * (1 + _JITTER)
+    gap = lo + (hi - lo) * rand01
+    return max(_GAP_FLOOR_S, gap)
+
+
+def _now_parts(conn):
+    """(hour:int, today:date, epoch:float) from the DB clock — avoids container-TZ ambiguity."""
+    with conn.cursor() as cur:
+        cur.execute("SELECT extract(hour from now())::int, current_date, extract(epoch from now())")
+        hour, today, epoch = cur.fetchone()
+    return int(hour), today, float(epoch)
+
+
+def _read(conn, keys):
+    with conn.cursor() as cur:
+        cur.execute('SELECT key, value FROM "Config" WHERE key = ANY(%s)', (list(keys),))
+        return dict(cur.fetchall())
+
+
+def gate_open(conn) -> bool:
+    """Whether Qobuz may be used for a download right now (reads pacing config + counter/spacing
+    state from Config and the DB clock). A counter from a previous day is treated as 0."""
+    from lyra_worker.config import get_config
+
+    pacing = QobuzPacing.from_config(get_config(conn))
+    hour, today, epoch = _now_parts(conn)
+    state = _read(conn, ("qobuz.pacing.countDay", "qobuz.pacing.countToday", "qobuz.pacing.nextAllowedAt"))
+    count = 0
+    if state.get("qobuz.pacing.countDay") == today.isoformat():
+        try:
+            count = int(state.get("qobuz.pacing.countToday") or 0)
+        except (TypeError, ValueError):
+            count = 0
+    try:
+        next_allowed = float(state.get("qobuz.pacing.nextAllowedAt") or 0)
+    except (TypeError, ValueError):
+        next_allowed = 0.0
+    return is_allowed(pacing, hour, count, today_cap(pacing, today), epoch, next_allowed)
+
+
+def record_download(conn, rand01: float) -> None:
+    """Record a successful Qobuz download: bump today's counter (resetting it if the stored day is
+    stale) and set the randomized next-allowed time. ``rand01`` is injected randomness."""
+    from lyra_worker.config import get_config
+
+    pacing = QobuzPacing.from_config(get_config(conn))
+    _hour, today, epoch = _now_parts(conn)
+    state = _read(conn, ("qobuz.pacing.countDay", "qobuz.pacing.countToday"))
+    if state.get("qobuz.pacing.countDay") == today.isoformat():
+        try:
+            count = int(state.get("qobuz.pacing.countToday") or 0)
+        except (TypeError, ValueError):
+            count = 0
+    else:
+        count = 0
+    count += 1
+    next_allowed = epoch + next_gap_seconds(pacing, today_cap(pacing, today), rand01)
+    _write(conn, {
+        "qobuz.pacing.countDay": today.isoformat(),
+        "qobuz.pacing.countToday": str(count),
+        "qobuz.pacing.nextAllowedAt": repr(next_allowed),
+    })
+
+
+def _write(conn, kv: dict) -> None:
+    with conn.cursor() as cur:
+        for key, value in kv.items():
+            cur.execute(
+                'INSERT INTO "Config"(key,value,secret,"updatedAt") VALUES (%s,%s,false,now()) '
+                'ON CONFLICT (key) DO UPDATE SET value=EXCLUDED.value, "updatedAt"=now()',
+                (key, value),
+            )
+    conn.commit()

worker/lyra_worker/similarity/_lastfm.py

@@ -31,6 +31,7 @@ class LastfmSource:
             params={"method": method, "api_key": self._key, "format": "json", **params},
             timeout=self._timeout,
         )
+        res.raise_for_status()  # a 4xx/5xx (rate limit, auth) → RequestException, handled by callers
         return res.json()
 
     def health(self) -> bool:

worker/lyra_worker/similarity/_lastfm_albums.py

@@ -41,6 +41,7 @@ class LastfmAlbumPopularity:
                 "api_key": self._key, "format": "json", "limit": str(self._limit),
                 "autocorrect": "1",
             }, timeout=self._timeout)
+            res.raise_for_status()
             return _parse_top_albums(res.json())
         except Exception:  # a Last.fm outage must never abort the sweep
             return []

worker/tests/test_candidate_filter.py

@@ -0,0 +1,57 @@
+"""A search result with far fewer tracks than the release (e.g. a single-song Soulseek folder
+named like the album) can't be the full album — it must be dropped before we waste a download."""
+import os
+
+from lyra_worker.claim import claim_next
+from lyra_worker.pipeline import run_pipeline
+from lyra_worker.types import Candidate, DownloadResult, MBTarget, Quality
+from tests.conftest import insert_request
+
+_Q = Quality(fmt="FLAC", lossless=True, bit_depth=24, sample_rate=96000)
+
+
+class _TwoCandQobuz:
+    name = "qobuz"
+    tier = 0
+
+    def __init__(self):
+        self.downloaded = []
+
+    def health(self):
+        return True
+
+    def search(self, target):
+        return [
+            Candidate(source="qobuz", source_ref="single", matched_artist=target.artist,
+                      matched_album=target.album, quality=_Q, track_count=1, source_tier=0),
+            Candidate(source="qobuz", source_ref="full", matched_artist=target.artist,
+                      matched_album=target.album, quality=_Q, track_count=12, source_tier=0),
+        ]
+
+    def download(self, candidate, dest, on_progress):
+        self.downloaded.append(candidate.source_ref)
+        os.makedirs(dest, exist_ok=True)
+        for i in range(candidate.track_count):
+            with open(os.path.join(dest, f"{i + 1:02d}.flac"), "wb") as fh:
+                fh.write(b"a")
+        return DownloadResult(ok=True, path=dest, track_count=candidate.track_count)
+
+
+class _Resolver:
+    def __init__(self, n):
+        self.n = n
+
+    def resolve(self, artist, album):
+        return MBTarget(artist=artist, album=album, track_count=self.n)
+
+
+def test_single_song_candidate_is_dropped(conn):
+    job_id = insert_request(conn, artist="A", album="B")
+    claim_next(conn)
+    adapter = _TwoCandQobuz()
+    run_pipeline(conn, job_id, [adapter], resolver=_Resolver(12), dest_root="/tmp/lib-cf")
+
+    assert adapter.downloaded == ["full"]  # only the full album was ever attempted
+    with conn.cursor() as cur:
+        cur.execute('SELECT "trackCount" FROM "Candidate" WHERE "jobId" = %s', (job_id,))
+        assert [r[0] for r in cur.fetchall()] == [12]  # the single-song folder wasn't even persisted

worker/tests/test_claim.py

@@ -24,3 +24,17 @@ def test_claim_does_not_repick_claimed_job(conn):
     insert_request(conn)
     assert claim_next(conn) is not None
     assert claim_next(conn) is None
+
+
+def test_claim_skips_paused_jobs(conn):
+    paused_job = insert_request(conn, album="Paused Album")
+    with conn.cursor() as cur:
+        cur.execute('UPDATE "Job" SET paused = true WHERE id = %s', (paused_job,))
+    conn.commit()
+    unpaused_job = insert_request(conn, album="Live Album")
+
+    claimed = claim_next(conn)
+    assert claimed == unpaused_job  # the paused one is skipped
+
+    # With only the paused job left, nothing is claimable.
+    assert claim_next(conn) is None

worker/tests/test_discovery.py

@@ -26,7 +26,7 @@ def test_seeds_discovery_from_owned_album_artists(conn):
     # An artist you own an album by but don't follow becomes a discovery seed.
     _insert_library_item(conn, "Owned Artist", "lib-seed")
     src = FakeSimilaritySource(similar={"lib-seed": [SimilarArtist("c1", "New Find", 0.9)]})
-    result = run_discovery(conn, [src], FakeMbBrowser(), DiscoveryConfig())
+    result = run_discovery(conn, [src], FakeMbBrowser(default_core=True), DiscoveryConfig())
     assert result.seeds == 1
     assert [r[0] for r in _artist_rows(conn)] == ["c1"]  # surfaced from the library seed
     with conn.cursor() as cur:  # library seed throttled via DiscoverySeed, not WatchedArtist
@@ -37,7 +37,7 @@ def test_seeds_discovery_from_owned_album_artists(conn):
 def test_library_seeding_can_be_disabled(conn):
     _insert_library_item(conn, "Owned Artist", "lib-seed")
     src = FakeSimilaritySource(similar={"lib-seed": [SimilarArtist("c1", "New Find", 0.9)]})
-    result = run_discovery(conn, [src], FakeMbBrowser(), DiscoveryConfig(seed_from_library=False))
+    result = run_discovery(conn, [src], FakeMbBrowser(default_core=True), DiscoveryConfig(seed_from_library=False))
     assert result.seeds == 0 and _artist_rows(conn) == []  # library not seeded
 
 
@@ -49,7 +49,7 @@ def test_owned_artist_is_not_suggested_as_a_candidate(conn):
         "s1": [SimilarArtist("c1", "Owned", 0.9), SimilarArtist("c2", "New", 0.8)],
         "c1": [],  # the library seed itself finds nothing new here
     })
-    run_discovery(conn, [src], FakeMbBrowser(), DiscoveryConfig())
+    run_discovery(conn, [src], FakeMbBrowser(default_core=True), DiscoveryConfig())
     assert [r[0] for r in _artist_rows(conn)] == ["c2"]  # c1 excluded (already owned)
 
 
@@ -60,7 +60,7 @@ def test_aggregates_scores_across_seeds(conn):
         "s1": [SimilarArtist("c1", "Shared", 0.6), SimilarArtist("c2", "Only1", 0.4)],
         "s2": [SimilarArtist("c1", "Shared", 0.5)],
     })
-    result = run_discovery(conn, [src], FakeMbBrowser(), DiscoveryConfig())
+    result = run_discovery(conn, [src], FakeMbBrowser(default_core=True), DiscoveryConfig())
     assert result.artists == 2
     rows = _artist_rows(conn)
     # c1 recommended by both seeds; each seed max-normalizes its top score to 1.0, so
@@ -81,7 +81,7 @@ def test_normalizes_across_sources_of_different_scales(conn):
     lfm = FakeSimilaritySource(similar={"s1": [
         SimilarArtist("c_lfm", "LFM Only", 0.9), SimilarArtist("c_shared", "Shared", 0.5)]})
     lfm.name = "lastfm"
-    run_discovery(conn, [lb, lfm], FakeMbBrowser(), DiscoveryConfig())
+    run_discovery(conn, [lb, lfm], FakeMbBrowser(default_core=True), DiscoveryConfig())
 
     lb_only = _artist_score(conn, "c_lb")[0]    # 5000/5000 = 1.0
     lfm_only = _artist_score(conn, "c_lfm")[0]  # 0.9/0.9  = 1.0
@@ -95,7 +95,7 @@ def test_skips_already_followed_artists(conn):
     insert_watched_artist(conn, mbid="c1", name="Already Followed")
     src = FakeSimilaritySource(similar={"s1": [SimilarArtist("c1", "X", 0.9),
                                                 SimilarArtist("c2", "New", 0.8)]})
-    run_discovery(conn, [src], FakeMbBrowser(), DiscoveryConfig())
+    run_discovery(conn, [src], FakeMbBrowser(default_core=True), DiscoveryConfig())
     assert [r[0] for r in _artist_rows(conn)] == ["c2"]  # c1 filtered (followed)
 
 
@@ -103,7 +103,7 @@ def test_min_score_filters_weak_candidates(conn):
     insert_watched_artist(conn, mbid="s1", name="Seed")
     src = FakeSimilaritySource(similar={"s1": [SimilarArtist("c1", "Strong", 0.9),
                                                 SimilarArtist("c2", "Weak", 0.1)]})
-    run_discovery(conn, [src], FakeMbBrowser(), DiscoveryConfig(min_score=0.5))
+    run_discovery(conn, [src], FakeMbBrowser(default_core=True), DiscoveryConfig(min_score=0.5))
     assert [r[0] for r in _artist_rows(conn)] == ["c1"]
 
 
@@ -111,7 +111,7 @@ def test_dismissed_suggestion_is_never_revived(conn):
     insert_watched_artist(conn, mbid="s1", name="Seed")
     insert_discovery_suggestion(conn, kind="artist", artist_mbid="c1", status="dismissed", score=0.0)
     src = FakeSimilaritySource(similar={"s1": [SimilarArtist("c1", "X", 0.9)]})
-    run_discovery(conn, [src], FakeMbBrowser(), DiscoveryConfig())
+    run_discovery(conn, [src], FakeMbBrowser(default_core=True), DiscoveryConfig())
     with conn.cursor() as cur:
         cur.execute('SELECT status::text, score FROM "DiscoverySuggestion" WHERE "artistMbid" = %s', ("c1",))
         assert cur.fetchone() == ("dismissed", 0.0)  # untouched — WHERE status='pending' blocked the update
@@ -120,18 +120,18 @@ def test_dismissed_suggestion_is_never_revived(conn):
 def test_stamps_last_discovered_and_throttles(conn):
     insert_watched_artist(conn, mbid="s1", name="Seed")
     src = FakeSimilaritySource(similar={"s1": [SimilarArtist("c1", "X", 0.9)]})
-    assert run_discovery(conn, [src], FakeMbBrowser(), DiscoveryConfig()).artists == 1
+    assert run_discovery(conn, [src], FakeMbBrowser(default_core=True), DiscoveryConfig()).artists == 1
     with conn.cursor() as cur:
         cur.execute('SELECT "lastDiscoveredAt" IS NOT NULL FROM "WatchedArtist" WHERE mbid = %s', ("s1",))
         assert cur.fetchone()[0] is True
     # second run: seed not due (lastDiscoveredAt recent) -> no seeds -> nothing new
-    assert run_discovery(conn, [src], FakeMbBrowser(), DiscoveryConfig()).artists == 0
+    assert run_discovery(conn, [src], FakeMbBrowser(default_core=True), DiscoveryConfig()).artists == 0
 
 
 def test_unhealthy_source_contributes_nothing(conn):
     insert_watched_artist(conn, mbid="s1", name="Seed")
     src = FakeSimilaritySource(similar={"s1": [SimilarArtist("c1", "X", 0.9)]}, healthy=False)
-    assert run_discovery(conn, [src], FakeMbBrowser(), DiscoveryConfig()).artists == 0
+    assert run_discovery(conn, [src], FakeMbBrowser(default_core=True), DiscoveryConfig()).artists == 0
     assert _artist_rows(conn) == []
 
 
@@ -146,7 +146,7 @@ def test_source_with_raising_health_is_skipped_not_fatal(conn):
             return []
 
     good = FakeSimilaritySource(similar={"s1": [SimilarArtist("c1", "X", 0.9)]})
-    result = run_discovery(conn, [BoomSource(), good], FakeMbBrowser(), DiscoveryConfig())
+    result = run_discovery(conn, [BoomSource(), good], FakeMbBrowser(default_core=True), DiscoveryConfig())
     assert result.artists == 1  # good source still processed; boom skipped, no crash
 
 
@@ -176,8 +176,8 @@ def test_score_sums_across_separate_chunks(conn):
         "s2": [SimilarArtist("c1", "Shared", 0.5)],
     })
     cfg = DiscoveryConfig(chunk_size=1)
-    run_discovery(conn, [src], FakeMbBrowser(), cfg)   # processes s1 (older)
-    run_discovery(conn, [src], FakeMbBrowser(), cfg)   # processes s2
+    run_discovery(conn, [src], FakeMbBrowser(default_core=True), cfg)   # processes s1 (older)
+    run_discovery(conn, [src], FakeMbBrowser(default_core=True), cfg)   # processes s2
     # replace-not-accumulate would leave 0.5/seedCount1; contributions give the sum.
     assert _artist_score(conn, "c1") == (2.0, 2)  # each seed max-normalizes to 1.0
 
@@ -193,7 +193,7 @@ def test_partial_sweep_preserves_other_seeds_contribution(conn):
         "s1": [SimilarArtist("c1", "Shared", 0.6)],
         "s2": [SimilarArtist("c1", "Shared", 0.5)],
     })
-    run_discovery(conn, [src], FakeMbBrowser(), DiscoveryConfig())  # both -> 1.1/2
+    run_discovery(conn, [src], FakeMbBrowser(default_core=True), DiscoveryConfig())  # both -> 1.1/2
     assert _artist_score(conn, "c1") == (2.0, 2)  # each seed max-normalizes to 1.0
 
     # s1 due again (older than interval), s2 fresh; s1 now scores c1 at 0.7.
@@ -202,7 +202,7 @@ def test_partial_sweep_preserves_other_seeds_contribution(conn):
         cur.execute('UPDATE "WatchedArtist" SET "lastDiscoveredAt" = now() WHERE mbid = %s', ("s2",))
     conn.commit()
     src2 = FakeSimilaritySource(similar={"s1": [SimilarArtist("c1", "Shared", 0.7)]})
-    run_discovery(conn, [src2], FakeMbBrowser(), DiscoveryConfig())
+    run_discovery(conn, [src2], FakeMbBrowser(default_core=True), DiscoveryConfig())
     assert _artist_score(conn, "c1") == (2.0, 2)  # 1.0 (new s1) + 1.0 (kept s2), max-normalized
 
 
@@ -215,7 +215,7 @@ def test_seed_dropoff_lowers_seedcount(conn):
         "s1": [SimilarArtist("c1", "Shared", 0.6)],
         "s2": [SimilarArtist("c1", "Shared", 0.5)],
     })
-    run_discovery(conn, [src], FakeMbBrowser(), DiscoveryConfig())
+    run_discovery(conn, [src], FakeMbBrowser(default_core=True), DiscoveryConfig())
     assert _artist_score(conn, "c1") == (2.0, 2)  # each seed max-normalizes to 1.0
 
     # s1 re-swept but no longer finds c1 -> its contribution is removed, c1 recomputed.
@@ -224,7 +224,7 @@ def test_seed_dropoff_lowers_seedcount(conn):
         cur.execute('UPDATE "WatchedArtist" SET "lastDiscoveredAt" = now() WHERE mbid = %s', ("s2",))
     conn.commit()
     src2 = FakeSimilaritySource(similar={"s1": [SimilarArtist("c9", "Other", 0.9)]})
-    run_discovery(conn, [src2], FakeMbBrowser(), DiscoveryConfig())
+    run_discovery(conn, [src2], FakeMbBrowser(default_core=True), DiscoveryConfig())
     assert _artist_score(conn, "c1") == (1.0, 1)  # only s2 remains (max-normalized to 1.0)
 
 
@@ -241,7 +241,7 @@ def test_evicts_contributions_from_removed_seed(conn):
         "s1": [SimilarArtist("c1", "Shared", 0.6)],
         "s2": [SimilarArtist("c1", "Shared", 0.5)],
     })
-    run_discovery(conn, [src], FakeMbBrowser(), DiscoveryConfig())
+    run_discovery(conn, [src], FakeMbBrowser(default_core=True), DiscoveryConfig())
     assert _artist_score(conn, "c1") == (2.0, 2)  # each seed max-normalizes to 1.0
 
     # Unfollow s1 (no FK cascade on DiscoverySeedContribution).
@@ -254,7 +254,7 @@ def test_evicts_contributions_from_removed_seed(conn):
         cur.execute('UPDATE "WatchedArtist" SET "lastDiscoveredAt" = now() - interval \'200 hours\' WHERE mbid = %s', ("s2",))
     conn.commit()
     src2 = FakeSimilaritySource(similar={"s2": [SimilarArtist("c1", "Shared", 0.5)]})
-    run_discovery(conn, [src2], FakeMbBrowser(), DiscoveryConfig())
+    run_discovery(conn, [src2], FakeMbBrowser(default_core=True), DiscoveryConfig())
 
     assert _artist_score(conn, "c1") == (1.0, 1)  # s1's stale contribution evicted; s2 alone -> 1.0
     with conn.cursor() as cur:
@@ -271,9 +271,38 @@ def test_run_discovery_processes_at_most_chunk_size_seeds(conn):
         insert_watched_artist(conn, mbid=f"s{i}", name=f"Seed{i}")
     src = FakeSimilaritySource(similar={f"s{i}": [SimilarArtist(f"c{i}", f"Cand{i}", 0.9)] for i in range(4)})
     cfg = DiscoveryConfig.from_config({"discover.chunkSize": "2"})
-    result = run_discovery(conn, [src], FakeMbBrowser(), cfg)
+    result = run_discovery(conn, [src], FakeMbBrowser(default_core=True), cfg)
     assert result.seeds == 2          # only chunk_size seeds this call
     # exactly 2 seeds now have lastDiscoveredAt set
     with conn.cursor() as cur:
         cur.execute('SELECT count(*) FROM "WatchedArtist" WHERE "lastDiscoveredAt" IS NOT NULL')
         assert cur.fetchone()[0] == 2
+
+
+def test_reset_pending_clears_state_and_marks_seeds_due(conn):
+    from lyra_worker.discovery import reset_pending
+    from tests.conftest import insert_discovery_suggestion
+    insert_watched_artist(conn, mbid="s1", name="Seed", last_discovered_sql="now()")
+    insert_discovery_suggestion(conn, kind="artist", artist_mbid="c1", status="pending")
+    insert_discovery_suggestion(conn, kind="artist", artist_mbid="c2", status="dismissed")
+    insert_discovery_suggestion(conn, kind="artist", artist_mbid="c3", status="wanted")
+    with conn.cursor() as cur:
+        cur.execute('INSERT INTO "DiscoverySeedContribution" (id, "candidateMbid", "candidateName", '
+                    '"seedMbid", score, sources, "updatedAt") VALUES '
+                    "(gen_random_uuid()::text, 'c1', 'C1', 's1', 1.0, '{}', now())")
+        cur.execute('INSERT INTO "DiscoverySeed" (mbid, "lastDiscoveredAt") VALUES (%s, now())', ("libseed",))
+    conn.commit()
+
+    reset_pending(conn)
+
+    with conn.cursor() as cur:
+        cur.execute("SELECT count(*) FROM \"DiscoverySuggestion\" WHERE status='pending'")
+        assert cur.fetchone()[0] == 0                       # pending cleared
+        cur.execute("SELECT count(*) FROM \"DiscoverySuggestion\" WHERE status IN ('dismissed','wanted')")
+        assert cur.fetchone()[0] == 2                       # user decisions kept
+        cur.execute('SELECT count(*) FROM "DiscoverySeedContribution"')
+        assert cur.fetchone()[0] == 0
+        cur.execute('SELECT count(*) FROM "DiscoverySeed"')
+        assert cur.fetchone()[0] == 0
+        cur.execute('SELECT count(*) FROM "WatchedArtist" WHERE "lastDiscoveredAt" IS NULL')
+        assert cur.fetchone()[0] == 1                       # seed due again

worker/tests/test_discovery_albums.py

@@ -2,7 +2,7 @@ from lyra_worker.adapters.fakes import FakeMbBrowser, FakeSimilaritySource
 from lyra_worker.browser import ReleaseGroupInfo
 from lyra_worker.discovery import DiscoveryConfig, run_discovery
 from lyra_worker.similarity.base import SimilarArtist
-from tests.conftest import insert_monitored_release, insert_watched_artist
+from tests.conftest import insert_discovery_suggestion, insert_monitored_release, insert_watched_artist
 
 
 def _album_rows(conn):
@@ -12,6 +12,13 @@ def _album_rows(conn):
         return cur.fetchall()
 
 
+def _artist_rows(conn):
+    with conn.cursor() as cur:
+        cur.execute("SELECT \"artistMbid\" FROM \"DiscoverySuggestion\" "
+                    "WHERE kind = 'artist' AND status = 'pending' ORDER BY \"artistMbid\"")
+        return cur.fetchall()
+
+
 def _reasons(conn):
     with conn.cursor() as cur:
         cur.execute('SELECT "rgMbid", "albumReason" FROM "DiscoverySuggestion" '
@@ -92,6 +99,56 @@ def test_popular_matches_by_title_when_lastfm_mbid_missing_and_skips_owned(conn)
     assert _reasons(conn) == {"rg-new": "newest", "rg-match": "popular"}
 
 
+def test_rederiving_replaces_stale_pending_album_suggestions(conn):
+    insert_watched_artist(conn, mbid="s1", name="Seed")
+    # a stale pending album suggestion for c1 that is no longer a pick (e.g. an old single)
+    insert_discovery_suggestion(conn, kind="album", artist_mbid="c1", rg_mbid="rg-old-single",
+                                album="Old Single", status="pending")
+    src = FakeSimilaritySource(similar={"s1": [SimilarArtist("c1", "Cand", 0.9)]})
+    browser = FakeMbBrowser(releases={"c1": [
+        ReleaseGroupInfo("rg-new", "New Album", "Album", (), "2022-01-01"),
+    ]})
+    run_discovery(conn, [src], browser, DiscoveryConfig(albums_per_artist=1))
+    # the stale single is gone; only the freshly-derived album remains
+    assert [r[1] for r in _album_rows(conn)] == ["rg-new"]
+
+
+def test_rederiving_keeps_wanted_album_suggestions(conn):
+    insert_watched_artist(conn, mbid="s1", name="Seed")
+    insert_discovery_suggestion(conn, kind="album", artist_mbid="c1", rg_mbid="rg-wanted",
+                                album="Wanted One", status="wanted")
+    src = FakeSimilaritySource(similar={"s1": [SimilarArtist("c1", "Cand", 0.9)]})
+    browser = FakeMbBrowser(releases={"c1": [
+        ReleaseGroupInfo("rg-new", "New Album", "Album", (), "2022-01-01"),
+    ]})
+    run_discovery(conn, [src], browser, DiscoveryConfig(albums_per_artist=1))
+    with conn.cursor() as cur:  # wanted row untouched (only pending stale rows are replaced)
+        cur.execute('SELECT count(*) FROM "DiscoverySuggestion" WHERE "rgMbid" = %s AND status = \'wanted\'', ("rg-wanted",))
+        assert cur.fetchone()[0] == 1
+
+
+def test_drops_artist_with_no_core_discography(conn):
+    # Two candidates: one real (has a core album), one a "producer" whose only release is a
+    # secondary-typed soundtrack (like Max Martin's & Juliet) — the latter must be dropped.
+    insert_watched_artist(conn, mbid="s1", name="Seed")
+    src = FakeSimilaritySource(similar={"s1": [
+        SimilarArtist("real", "Real Artist", 0.9), SimilarArtist("producer", "Producer", 0.8)]})
+    browser = FakeMbBrowser(releases={
+        "real": [ReleaseGroupInfo("rg-r", "Record", "Album", (), "2020-01-01")],
+        "producer": [ReleaseGroupInfo("rg-st", "A Musical", "Album", ("Soundtrack",), "2019-01-01")],
+    })
+    run_discovery(conn, [src], browser, DiscoveryConfig(albums_per_artist=1))
+    assert [r[0] for r in _artist_rows(conn)] == ["real"]  # producer dropped
+
+
+def test_drops_artist_with_empty_discography(conn):
+    insert_watched_artist(conn, mbid="s1", name="Seed")
+    src = FakeSimilaritySource(similar={"s1": [SimilarArtist("member", "Band Member", 0.9)]})
+    browser = FakeMbBrowser(releases={"member": []})  # no release groups at all
+    run_discovery(conn, [src], browser, DiscoveryConfig(albums_per_artist=1))
+    assert _artist_rows(conn) == []
+
+
 def test_no_popularity_provider_yields_newest_only(conn):
     insert_watched_artist(conn, mbid="s1", name="Seed")
     src = FakeSimilaritySource(similar={"s1": [SimilarArtist("c1", "Cand", 0.9)]})

worker/tests/test_floor.py

@@ -0,0 +1,50 @@
+from lyra_worker.floor import (
+    press_paused,
+    job_delay_seconds,
+    kill_requested,
+    consume_kill_switch,
+)
+
+
+def test_press_paused_reads_truthy_values():
+    assert press_paused({"floor.paused": "true"}) is True
+    assert press_paused({"floor.paused": "ON"}) is True
+    assert press_paused({"floor.paused": "false"}) is False
+    assert press_paused({}) is False
+
+
+def test_job_delay_seconds_parses_and_defaults():
+    assert job_delay_seconds({"floor.jobDelaySeconds": "30"}) == 30.0
+    assert job_delay_seconds({}) == 0.0
+    assert job_delay_seconds({"floor.jobDelaySeconds": "-5"}) == 0.0
+    assert job_delay_seconds({"floor.jobDelaySeconds": "junk"}) == 0.0
+
+
+def test_kill_requested_reads_flag():
+    assert kill_requested({"floor.killRequested": "true"}) is True
+    assert kill_requested({}) is False
+
+
+def test_consume_kill_switch_clears_flag_and_exits(conn):
+    with conn.cursor() as cur:
+        cur.execute(
+            'INSERT INTO "Config"(key, value, secret, "updatedAt") '
+            "VALUES ('floor.killRequested', 'true', false, now())"
+        )
+    conn.commit()
+
+    calls = []
+    fired = consume_kill_switch(conn, exit_fn=lambda code: calls.append(code))
+
+    assert fired is True
+    assert calls == [0]
+    with conn.cursor() as cur:
+        cur.execute("SELECT value FROM \"Config\" WHERE key = 'floor.killRequested'")
+        assert cur.fetchone()[0] == "false"
+
+
+def test_consume_kill_switch_noop_when_unset(conn):
+    calls = []
+    fired = consume_kill_switch(conn, exit_fn=lambda code: calls.append(code))
+    assert fired is False
+    assert calls == []

worker/tests/test_incomplete_fallthrough.py

@@ -0,0 +1,125 @@
+"""When the top-ranked source downloads but is incomplete (e.g. Qobuz consistently fails one
+track), the pipeline must fall through to the next-ranked source instead of failing the whole
+job — only giving up (needs_attention) when every candidate is incomplete."""
+import os
+
+from lyra_worker.claim import claim_next
+from lyra_worker.pipeline import run_pipeline
+from lyra_worker.types import Candidate, DownloadResult, Quality
+from tests.conftest import insert_request
+
+_HIRES = Quality(fmt="FLAC", lossless=True, bit_depth=24, sample_rate=96000)
+_CD = Quality(fmt="FLAC", lossless=True, bit_depth=16, sample_rate=44100)
+
+
+def _stage(dest, n):
+    os.makedirs(dest, exist_ok=True)
+    for i in range(n):
+        with open(os.path.join(dest, f"{i + 1:02d}.flac"), "wb") as fh:
+            fh.write(b"audio")
+
+
+class _Qobuz:
+    """Higher-quality (ranks first) Qobuz that promises 3 tracks but delivers `deliver`."""
+    name = "qobuz"
+    tier = 0
+
+    def __init__(self, deliver):
+        self._deliver = deliver
+
+    def health(self):
+        return True
+
+    def search(self, target):
+        return [Candidate(source="qobuz", source_ref="q", matched_artist=target.artist,
+                          matched_album=target.album, quality=_HIRES, track_count=3, source_tier=0)]
+
+    def download(self, candidate, dest, on_progress):
+        _stage(dest, self._deliver)
+        return DownloadResult(ok=True, path=dest, track_count=self._deliver)
+
+
+class _Soulseek:
+    """Lower-quality (ranks second) Soulseek that promises 3 tracks and delivers `deliver`."""
+    name = "soulseek"
+    tier = 1
+
+    def __init__(self, deliver):
+        self._deliver = deliver
+
+    def health(self):
+        return True
+
+    def search(self, target):
+        return [Candidate(source="soulseek", source_ref="s", matched_artist=target.artist,
+                          matched_album=target.album, quality=_CD, track_count=3, source_tier=1)]
+
+    def download(self, candidate, dest, on_progress):
+        _stage(dest, self._deliver)
+        return DownloadResult(ok=True, path=dest, track_count=self._deliver)
+
+
+def _state_error(conn, job_id):
+    with conn.cursor() as cur:
+        cur.execute('SELECT state, error FROM "Job" WHERE id = %s', (job_id,))
+        return cur.fetchone()
+
+
+def _chosen_source(conn, job_id):
+    with conn.cursor() as cur:
+        cur.execute('SELECT source FROM "Candidate" WHERE "jobId" = %s AND chosen = true', (job_id,))
+        row = cur.fetchone()
+        return row[0] if row else None
+
+
+def test_incomplete_top_source_falls_through_to_complete_next(conn, tmp_path):
+    # Qobuz (ranks first on quality) delivers only 2 of its promised 3 tracks -> incomplete.
+    # The pipeline must try the next source (Soulseek, complete) rather than failing the job.
+    job_id = insert_request(conn, artist="Radiohead", album="In Rainbows")
+    claim_next(conn)
+    run_pipeline(conn, job_id, [_Qobuz(deliver=2), _Soulseek(deliver=3)], dest_root=str(tmp_path))
+    assert _state_error(conn, job_id) == ("imported", None)
+    assert _chosen_source(conn, job_id) == "soulseek"
+
+
+def test_all_sources_incomplete_still_needs_attention(conn, tmp_path):
+    # Every candidate is incomplete -> no fallback succeeds -> needs_attention (incomplete).
+    job_id = insert_request(conn, artist="Radiohead", album="In Rainbows")
+    claim_next(conn)
+    run_pipeline(conn, job_id, [_Qobuz(deliver=2), _Soulseek(deliver=1)], dest_root=str(tmp_path))
+    assert _state_error(conn, job_id) == ("needs_attention", "incomplete download")
+    assert not (tmp_path / "Radiohead" / "In Rainbows").exists()  # nothing imported
+
+
+class _CountingIncompleteQobuz:
+    """Offers `n` candidates that all download but come back incomplete; counts download calls."""
+    name = "qobuz"
+    tier = 0
+
+    def __init__(self, n):
+        self.n = n
+        self.calls = 0
+
+    def health(self):
+        return True
+
+    def search(self, target):
+        q = _HIRES
+        return [Candidate(source="qobuz", source_ref=f"r{i}", matched_artist=target.artist,
+                          matched_album=target.album, quality=q, track_count=3, source_tier=0)
+                for i in range(self.n)]
+
+    def download(self, candidate, dest, on_progress):
+        self.calls += 1
+        _stage(dest, 1)  # 1 of 3 → always incomplete
+        return DownloadResult(ok=True, path=dest, track_count=1)
+
+
+def test_fall_through_is_capped(conn):
+    from lyra_worker.pipeline import _MAX_DOWNLOAD_ATTEMPTS
+    a = _CountingIncompleteQobuz(n=20)
+    job_id = insert_request(conn, artist="Radiohead", album="In Rainbows")
+    claim_next(conn)
+    run_pipeline(conn, job_id, [a], dest_root="/tmp/lib-cap")
+    assert _state_error(conn, job_id) == ("needs_attention", "incomplete download")
+    assert a.calls == _MAX_DOWNLOAD_ATTEMPTS  # capped, not all 20 peers tried

worker/tests/test_monitor_enqueue.py

@@ -72,6 +72,52 @@ def test_grabbed_past_window_is_fulfilled_not_enqueued(conn):
     assert _state(conn, rel_id) == "fulfilled"
 
 
+def test_owned_by_rgmbid_despite_name_mismatch_is_fulfilled(conn):
+    # The monitor followed "Kanye West" but the library stores MB's canonical "Ye"; matching on
+    # rgMbid recognises ownership so the release is fulfilled, not re-pressed.
+    rel_id = insert_monitored_release(conn, artist_name="Kanye West", album="Graduation",
+                                      rg_mbid="rg-grad", monitored=True, state="wanted")
+    with conn.cursor() as cur:
+        cur.execute(
+            'INSERT INTO "LibraryItem" (id, artist, album, path, source, format, '
+            '"qualityClass", "rgMbid", "importedAt") '
+            "VALUES (gen_random_uuid()::text, 'Ye', 'Graduation', '/m', 'soulseek', 'FLAC', 2, "
+            "'rg-grad', now())"
+        )
+    conn.commit()
+    assert enqueue_due(conn, CFG) == 0
+    assert _state(conn, rel_id) == "fulfilled"
+
+
+def test_re_enqueue_retires_prior_failed_attempt(conn):
+    # A hard-to-match release re-pressed each retry interval must not pile up needs_attention
+    # Request rows; the prior failed attempt is retired when the next one is enqueued.
+    rel_id = insert_monitored_release(conn, artist_name="X", album="Y", rg_mbid="rg-1",
+                                      monitored=True, state="wanted")
+    with conn.cursor() as cur:
+        cur.execute(
+            'INSERT INTO "Request" (id, artist, album, status, "monitoredReleaseId", "createdAt") '
+            "VALUES (gen_random_uuid()::text, 'X', 'Y', 'needs_attention', %s, now()) RETURNING id",
+            (rel_id,),
+        )
+        req = cur.fetchone()[0]
+        cur.execute(
+            'INSERT INTO "Job" (id, "requestId", state, "currentStage", attempts, "createdAt", "updatedAt") '
+            "VALUES (gen_random_uuid()::text, %s, 'needs_attention', 'rank', 1, now(), now())",
+            (req,),
+        )
+    conn.commit()
+    assert enqueue_due(conn, CFG) == 1
+    assert _job_count_for(conn, rel_id) == 1  # old failed attempt retired, one fresh attempt
+    with conn.cursor() as cur:
+        cur.execute(
+            'SELECT j.state FROM "Job" j JOIN "Request" r ON r.id = j."requestId" '
+            'WHERE r."monitoredReleaseId" = %s',
+            (rel_id,),
+        )
+        assert cur.fetchone()[0] == "requested"
+
+
 def test_already_in_library_at_cutoff_is_fulfilled(conn):
     rel_id = insert_monitored_release(conn, artist_name="A", album="B", rg_mbid="rg-1",
                                       monitored=True, state="wanted")

worker/tests/test_monitor_reconcile.py

@@ -22,17 +22,27 @@ def _make_job(conn, rel_id, artist, album, state):
     return req, job
 
 
-def _add_library(conn, req, artist, album, quality):
+def _add_library(conn, req, artist, album, quality, rg_mbid=None):
     with conn.cursor() as cur:
         cur.execute(
             'INSERT INTO "LibraryItem" (id, "requestId", artist, album, path, source, format, '
-            '"qualityClass", "importedAt") '
-            "VALUES (gen_random_uuid()::text, %s, %s, %s, '/m', 'qobuz', 'FLAC', %s, now())",
-            (req, artist, album, quality),
+            '"qualityClass", "rgMbid", "importedAt") '
+            "VALUES (gen_random_uuid()::text, %s, %s, %s, '/m', 'qobuz', 'FLAC', %s, %s, now())",
+            (req, artist, album, quality, rg_mbid),
         )
     conn.commit()
 
 
+def _needs_attention_count(conn, rel_id):
+    with conn.cursor() as cur:
+        cur.execute(
+            'SELECT count(*) FROM "Request" r JOIN "Job" j ON j."requestId" = r.id '
+            "WHERE r.\"monitoredReleaseId\" = %s AND j.state = 'needs_attention'",
+            (rel_id,),
+        )
+        return cur.fetchone()[0]
+
+
 def _release(conn, rel_id):
     with conn.cursor() as cur:
         cur.execute(
@@ -106,6 +116,42 @@ def test_fulfill_owned_releases_ignores_below_cutoff(conn):
     assert _release(conn, rel)[0] == "wanted"
 
 
+def test_reconcile_matches_by_rgmbid_when_artist_name_canonicalized(conn):
+    # Real 'Kanye West' -> 'Ye' bug: the MonitoredRelease keeps the followed name, but the
+    # imported LibraryItem carries MusicBrainz's canonical artist. Matching on rgMbid lets the
+    # release reconcile to fulfilled instead of looking un-owned and re-pressing forever.
+    rel = insert_monitored_release(conn, artist_name="Kanye West", album="Graduation",
+                                   rg_mbid="rg-grad", monitored=True, state="wanted")
+    req, job = _make_job(conn, rel, "Kanye West", "Graduation", "imported")
+    _add_library(conn, req, "Ye", "Graduation", 3, rg_mbid="rg-grad")
+    reconcile(conn, job, CFG)
+    assert _release(conn, rel) == ("fulfilled", 3, True)
+
+
+def test_reconcile_fulfilled_retires_failed_sibling_attempts(conn):
+    # 'V' by Maroon 5 phantom: earlier failed attempts leave needs_attention Request rows that
+    # keep showing on the press even after a later attempt imports. A successful import retires
+    # those failed siblings so the press stops showing an owned album as needing attention.
+    rel = insert_monitored_release(conn, artist_name="Maroon 5", album="V", rg_mbid="rg-v",
+                                   monitored=True, state="wanted")
+    _make_job(conn, rel, "Maroon 5", "V", "needs_attention")
+    _make_job(conn, rel, "Maroon 5", "V", "needs_attention")
+    req, job = _make_job(conn, rel, "Maroon 5", "V", "imported")
+    _add_library(conn, req, "Maroon 5", "V", 2)
+    reconcile(conn, job, CFG)
+    assert _release(conn, rel)[0] == "fulfilled"
+    assert _needs_attention_count(conn, rel) == 0  # failed siblings retired
+
+
+def test_fulfill_owned_releases_matches_by_rgmbid(conn):
+    # Ne-Yo -> Ne‐Yo (typographic hyphen) name drift: still owned by rgMbid.
+    rel = insert_monitored_release(conn, artist_name="Ne-Yo", album="Because of You",
+                                   rg_mbid="rg-ney", monitored=True, state="grabbed")
+    _add_library(conn, None, "Ne‐Yo", "Because of You", 2, rg_mbid="rg-ney")
+    assert fulfill_owned_releases(conn, MonitorConfig(quality_cutoff=2)) == 1
+    assert _release(conn, rel) == ("fulfilled", 2, True)
+
+
 def test_unlinked_job_is_ignored(conn):
     with conn.cursor() as cur:
         cur.execute(

worker/tests/test_musicbrainz.py

@@ -0,0 +1,174 @@
+import musicbrainzngs
+import pytest
+
+from lyra_worker._musicbrainz import _best_release_group, _pick_release, _with_retry
+
+# Offline unit tests for the pure release-group selection logic. The live resolve()
+# path (network) is covered by test_musicbrainz_live.py.
+
+
+def _rg(title, primary_type, artist=""):
+    g = {"id": title.lower(), "title": title, "primary-type": primary_type}
+    if artist:
+        g["artist-credit"] = [{"artist": {"name": artist, "id": artist.lower()}}]
+    return g
+
+
+def test_prefers_album_over_same_titled_single():
+    # Real-world "Off the Wall" bug: MusicBrainz returns the Single release-group
+    # first, tied on title with the Album. We must pick the Album so a 10-track
+    # download isn't tagged from the 2-track single's tracklist.
+    groups = [
+        _rg("Off the Wall", "Single", "Michael Jackson"),
+        _rg("Off the Wall", "Album", "Michael Jackson"),
+    ]
+    assert _best_release_group("Michael Jackson", "Off the Wall", groups)["primary-type"] == "Album"
+
+
+def test_album_preference_is_order_independent():
+    groups = [
+        _rg("Off the Wall", "Album", "Michael Jackson"),
+        _rg("Off the Wall", "Single", "Michael Jackson"),
+    ]
+    assert _best_release_group("Michael Jackson", "Off the Wall", groups)["primary-type"] == "Album"
+
+
+def test_closer_title_still_wins_over_album_type():
+    # Title similarity dominates the album-type preference (for a single artist).
+    groups = [
+        _rg("Continuum", "Single", "John Mayer"),
+        _rg("Continuum Deluxe Reissue", "Album", "John Mayer"),
+    ]
+    assert _best_release_group("John Mayer", "Continuum", groups)["title"] == "Continuum"
+
+
+def test_below_threshold_returns_none():
+    groups = [_rg("Something Totally Different", "Album", "Michael Jackson")]
+    assert _best_release_group("Michael Jackson", "Off the Wall", groups) is None
+
+
+def test_empty_groups_returns_none():
+    assert _best_release_group("Michael Jackson", "Off the Wall", []) is None
+
+
+def test_prefers_original_group_over_later_reissue():
+    # Real "Hybrid Theory" bug: MB lists a 2000 original and a 2023 reissue as separate same-titled
+    # Album groups (tie on artist/title/type). The reissue's longer tracklist fails completeness
+    # checks, so the original (earliest) must win regardless of MB's relevance order.
+    groups = [
+        {"id": "reissue", "title": "Hybrid Theory", "primary-type": "Album",
+         "first-release-date": "2023-07-21",
+         "artist-credit": [{"artist": {"name": "Linkin Park"}}]},
+        {"id": "original", "title": "Hybrid Theory", "primary-type": "Album",
+         "first-release-date": "2000-10-24",
+         "artist-credit": [{"artist": {"name": "Linkin Park"}}]},
+    ]
+    assert _best_release_group("Linkin Park", "Hybrid Theory", groups)["id"] == "original"
+
+
+def test_prefers_credited_artist_over_same_titled_other_artist_album():
+    # Real-world "Fun Machine" bug: Lake Street Dive's EP is what we followed, but two
+    # unrelated bands also have a release group literally titled "Fun Machine" as an Album.
+    # All three tie on title (1.0), so the old is_album tiebreak wrongly grabbed a foreign
+    # Album. The followed artist must win over title-tied albums by other artists.
+    groups = [
+        _rg("Fun Machine", "EP", "Lake Street Dive"),
+        _rg("Fun Machine", "Album", "Bastards of Melody"),
+        _rg("Fun Machine", "Album", "Teledildonics 5000"),
+    ]
+    best = _best_release_group("Lake Street Dive", "Fun Machine", groups)
+    assert best["artist-credit"][0]["artist"]["name"] == "Lake Street Dive"
+
+
+# --- _with_retry: transient MusicBrainz failures (SSL EOF, 503 rate-limit) ---
+
+
+def test_pick_release_prefers_original_official_over_later_reissue():
+    # Real "Hybrid Theory" bug: MB's arbitrary releases[0] was a 2023 13-track reissue; the 2000
+    # original is what sources deliver, so completeness checks must reference it.
+    releases = [
+        {"id": "reissue", "status": "Official", "date": "2023-06-01"},
+        {"id": "original", "status": "Official", "date": "2000-10-24"},
+        {"id": "deluxe", "status": "Official", "date": "2020-10-09"},
+    ]
+    assert _pick_release(releases)["id"] == "original"
+
+
+def test_pick_release_skips_non_official_and_undated():
+    releases = [
+        {"id": "promo", "status": "Promotion", "date": "1999-01-01"},  # earliest but not official
+        {"id": "undated", "status": "Official", "date": ""},
+        {"id": "official", "status": "Official", "date": "2001-03-03"},
+    ]
+    assert _pick_release(releases)["id"] == "official"
+
+
+def test_pick_release_empty_is_none():
+    assert _pick_release([]) is None
+
+
+def test_retry_returns_value_without_backoff_when_first_call_succeeds():
+    sleeps: list[float] = []
+    calls = {"n": 0}
+
+    def fn():
+        calls["n"] += 1
+        return "ok"
+
+    assert _with_retry(fn, sleep=sleeps.append) == "ok"
+    assert calls["n"] == 1
+    assert sleeps == []  # no backoff when nothing failed
+
+
+def test_retry_recovers_from_transient_network_error():
+    # The real symptom: <urlopen error [SSL: UNEXPECTED_EOF_WHILE_READING]> arrives
+    # as musicbrainzngs.NetworkError. A couple of retries clears it.
+    sleeps: list[float] = []
+    calls = {"n": 0}
+
+    def fn():
+        calls["n"] += 1
+        if calls["n"] < 3:
+            raise musicbrainzngs.NetworkError("SSL: UNEXPECTED_EOF_WHILE_READING")
+        return "recovered"
+
+    assert _with_retry(fn, sleep=sleeps.append) == "recovered"
+    assert calls["n"] == 3
+    assert sleeps == [1.0, 2.0]  # exponential backoff before each retry
+
+
+def test_retry_recovers_from_transient_response_error():
+    # MusicBrainz's own 503 rate-limit surfaces as ResponseError.
+    calls = {"n": 0}
+
+    def fn():
+        calls["n"] += 1
+        if calls["n"] < 2:
+            raise musicbrainzngs.ResponseError("503 Service Unavailable")
+        return "ok"
+
+    assert _with_retry(fn, sleep=lambda _: None) == "ok"
+    assert calls["n"] == 2
+
+
+def test_retry_reraises_after_exhausting_attempts():
+    sleeps: list[float] = []
+
+    def fn():
+        raise musicbrainzngs.NetworkError("still down")
+
+    with pytest.raises(musicbrainzngs.NetworkError):
+        _with_retry(fn, attempts=4, sleep=sleeps.append)
+    assert sleeps == [1.0, 2.0, 4.0]  # 3 backoffs across 4 attempts
+
+
+def test_retry_does_not_swallow_non_transient_errors():
+    calls = {"n": 0}
+
+    def fn():
+        calls["n"] += 1
+        raise ValueError("bug, not a network blip")
+
+    with pytest.raises(ValueError):
+        _with_retry(fn, sleep=lambda _: None)
+    assert calls["n"] == 1  # non-transient errors propagate immediately, no retry

worker/tests/test_mutagen.py

@@ -1,4 +1,34 @@
-from lyra_worker._mutagen import plan_track_names
+from lyra_worker._mutagen import _clean_extra_title, plan_track_names
+
+
+def test_clean_extra_title_strips_number_and_artist_prefix():
+    assert _clean_extra_title(
+        "13. John Mayer - St. Patrick's Day (Album Version)", "John Mayer"
+    ) == "St. Patrick's Day (Album Version)"
+
+
+def test_clean_extra_title_strips_number_without_artist_prefix():
+    assert _clean_extra_title("13. St. Patrick's Day", "John Mayer") == "St. Patrick's Day"
+
+
+def test_clean_extra_title_leaves_plain_title_and_numeric_titles():
+    assert _clean_extra_title("Hidden Track", "John Mayer") == "Hidden Track"
+    # "99 Luftballons" has no delimiter after the number -> not a track-number prefix
+    assert _clean_extra_title("99 Luftballons", "Nena") == "99 Luftballons"
+
+
+def test_plan_canonicalizes_bonus_track_beyond_tracklist():
+    names = [f"{i:02d} Track{i}.flac" for i in range(1, 13)] + [
+        "13. John Mayer - St. Patrick's Day (Album Version).flac"
+    ]
+    tracklist = tuple(f"Track{i}" for i in range(1, 13))
+    plan = plan_track_names(names, tracklist, artist="John Mayer")
+    assert plan[12] == (
+        "13. John Mayer - St. Patrick's Day (Album Version).flac",
+        "13 St. Patrick's Day (Album Version).flac",
+        "St. Patrick's Day (Album Version)",
+        13,
+    )
 
 
 def test_plan_uses_tracklist_titles_and_numbers():

worker/tests/test_pipeline.py

@@ -195,6 +195,33 @@ def test_below_threshold_candidates_are_persisted_for_diagnosis(conn):
         assert cur.fetchone()[0] >= 1  # found-but-rejected candidate retained for diagnosis
 
 
+def test_searches_run_concurrently_across_adapters(conn):
+    # Two sources that each take ~0.5s to search must run in parallel: total match time is ~the
+    # slowest single search (~0.5s), not the sum (~1.0s). Also verifies concurrency drops no
+    # candidates — both sources' contributions are persisted.
+    import time as _t
+
+    class _SlowQobuz(FakeQobuz):
+        def search(self, target):
+            _t.sleep(0.5)
+            return super().search(target)
+
+    class _SlowSoulseek(FakeSoulseek):
+        def search(self, target):
+            _t.sleep(0.5)
+            return super().search(target)
+
+    job_id = insert_request(conn, artist="Radiohead", album="In Rainbows")
+    claim_next(conn)
+    start = _t.monotonic()
+    run_pipeline(conn, job_id, [_SlowQobuz(), _SlowSoulseek()], dest_root="/tmp/lib")
+    elapsed = _t.monotonic() - start
+    assert elapsed < 0.9  # concurrent (~0.5s), not sequential (~1.0s)
+    with conn.cursor() as cur:
+        cur.execute('SELECT count(DISTINCT source) FROM "Candidate" WHERE "jobId" = %s', (job_id,))
+        assert cur.fetchone()[0] == 2  # both adapters contributed candidates
+
+
 def test_duplicate_adapter_names_raise(conn):
     import pytest
     from lyra_worker.adapters.fakes import FakeQobuz

worker/tests/test_qobuz_gate.py

@@ -0,0 +1,179 @@
+from datetime import date
+
+from lyra_worker.qobuz_gate import (
+    QobuzPacing,
+    today_cap,
+    is_allowed,
+    next_gap_seconds,
+    gate_open,
+    record_download,
+)
+
+
+# ---- QobuzPacing.from_config -------------------------------------------------
+
+def test_from_config_defaults():
+    p = QobuzPacing.from_config({})
+    assert p.enabled is True
+    assert p.active_start == 8 and p.active_end == 23
+    assert p.daily_cap == 40
+    assert p.warmup_start_date == ""
+
+
+def test_from_config_reads_values():
+    p = QobuzPacing.from_config({
+        "qobuz.pacing.enabled": "false",
+        "qobuz.pacing.activeStartHour": "9",
+        "qobuz.pacing.activeEndHour": "22",
+        "qobuz.pacing.dailyCap": "25",
+        "qobuz.pacing.warmupStartDate": "2026-07-15",
+    })
+    assert p.enabled is False
+    assert p.active_start == 9 and p.active_end == 22
+    assert p.daily_cap == 25
+    assert p.warmup_start_date == "2026-07-15"
+
+
+# ---- today_cap (warm-up ramp) ------------------------------------------------
+
+def test_today_cap_no_warmup_uses_steady():
+    p = QobuzPacing(daily_cap=40, warmup_start_date="")
+    assert today_cap(p, date(2026, 7, 20)) == 40
+
+
+def test_today_cap_ramp_boundaries():
+    p = QobuzPacing(daily_cap=40, warmup_start_date="2026-07-15")
+    assert today_cap(p, date(2026, 7, 15)) == 5    # day 0
+    assert today_cap(p, date(2026, 7, 17)) == 5    # day 2
+    assert today_cap(p, date(2026, 7, 18)) == 10   # day 3
+    assert today_cap(p, date(2026, 7, 21)) == 10   # day 6
+    assert today_cap(p, date(2026, 7, 22)) == 20   # day 7
+    assert today_cap(p, date(2026, 7, 28)) == 20   # day 13
+    assert today_cap(p, date(2026, 7, 29)) == 40   # day 14
+    assert today_cap(p, date(2026, 8, 30)) == 40   # long after
+
+
+def test_today_cap_steady_below_ramp_step_is_the_ceiling():
+    # A small steady cap is never exceeded by a ramp step.
+    p = QobuzPacing(daily_cap=3, warmup_start_date="2026-07-15")
+    assert today_cap(p, date(2026, 7, 29)) == 3
+
+
+# ---- is_allowed (the gate decision) -----------------------------------------
+
+_P = QobuzPacing(enabled=True, active_start=8, active_end=23, daily_cap=40)
+
+
+def test_disabled_always_allows():
+    p = QobuzPacing(enabled=False, active_start=8, active_end=23)
+    assert is_allowed(p, now_hour=3, count_today=999, cap=0, now_epoch=0, next_allowed_epoch=1e18) is True
+
+
+def test_blocked_outside_active_hours():
+    assert is_allowed(_P, now_hour=7, count_today=0, cap=40, now_epoch=100, next_allowed_epoch=0) is False
+    assert is_allowed(_P, now_hour=23, count_today=0, cap=40, now_epoch=100, next_allowed_epoch=0) is False
+
+
+def test_blocked_at_or_over_cap():
+    assert is_allowed(_P, now_hour=10, count_today=40, cap=40, now_epoch=100, next_allowed_epoch=0) is False
+    assert is_allowed(_P, now_hour=10, count_today=41, cap=40, now_epoch=100, next_allowed_epoch=0) is False
+
+
+def test_blocked_before_next_allowed():
+    assert is_allowed(_P, now_hour=10, count_today=0, cap=40, now_epoch=100, next_allowed_epoch=200) is False
+
+
+def test_allowed_when_all_conditions_met():
+    assert is_allowed(_P, now_hour=10, count_today=5, cap=40, now_epoch=300, next_allowed_epoch=200) is True
+
+
+# ---- next_gap_seconds (spread + jitter) -------------------------------------
+
+def test_next_gap_spreads_cap_across_active_window():
+    # 15h window (8..23) / cap 40 = 1350s base; ±30%; midpoint rand -> base.
+    p = QobuzPacing(active_start=8, active_end=23, daily_cap=40)
+    assert abs(next_gap_seconds(p, cap=40, rand01=0.5) - 1350.0) < 1.0
+    assert abs(next_gap_seconds(p, cap=40, rand01=0.0) - 1350.0 * 0.7) < 1.0   # low end
+    assert abs(next_gap_seconds(p, cap=40, rand01=1.0) - 1350.0 * 1.3) < 1.0   # high end
+
+
+def test_next_gap_has_a_floor():
+    # A huge cap would give a tiny gap; a 60s floor protects the request rate.
+    p = QobuzPacing(active_start=8, active_end=23, daily_cap=100000)
+    assert next_gap_seconds(p, cap=100000, rand01=0.0) == 60.0
+
+
+# ---- DB helpers: gate_open / record_download --------------------------------
+
+def _set(conn, key, value):
+    with conn.cursor() as cur:
+        cur.execute(
+            'INSERT INTO "Config"(key,value,secret,"updatedAt") VALUES (%s,%s,false,now()) '
+            'ON CONFLICT (key) DO UPDATE SET value=EXCLUDED.value, "updatedAt"=now()',
+            (key, value),
+        )
+    conn.commit()
+
+
+def _get(conn, key):
+    with conn.cursor() as cur:
+        cur.execute('SELECT value FROM "Config" WHERE key=%s', (key,))
+        row = cur.fetchone()
+        return row[0] if row else None
+
+
+def test_record_download_increments_and_sets_next_allowed(conn):
+    _set(conn, "qobuz.pacing.dailyCap", "40")
+    record_download(conn, rand01=0.5)
+    assert _get(conn, "qobuz.pacing.countToday") == "1"
+    assert _get(conn, "qobuz.pacing.countDay") is not None
+    assert _get(conn, "qobuz.pacing.nextAllowedAt") is not None
+    record_download(conn, rand01=0.5)
+    assert _get(conn, "qobuz.pacing.countToday") == "2"
+
+
+def test_record_download_resets_counter_on_new_day(conn):
+    # A stale count from a previous day must reset to 1, not accumulate.
+    _set(conn, "qobuz.pacing.dailyCap", "40")
+    _set(conn, "qobuz.pacing.countDay", "2000-01-01")
+    _set(conn, "qobuz.pacing.countToday", "39")
+    record_download(conn, rand01=0.5)
+    assert _get(conn, "qobuz.pacing.countToday") == "1"
+
+
+def test_gate_open_with_always_open_window(conn):
+    # An all-day window (0..24) removes the time-of-day dependency: no prior downloads and no
+    # next-allowed set → the gate is open regardless of when the test runs.
+    _set(conn, "qobuz.pacing.activeStartHour", "0")
+    _set(conn, "qobuz.pacing.activeEndHour", "24")
+    assert gate_open(conn) is True
+
+
+def test_gate_closed_by_empty_active_window(conn):
+    # start == end is an empty window (h >= 5 AND h < 5 is never true) → always closed.
+    _set(conn, "qobuz.pacing.activeStartHour", "5")
+    _set(conn, "qobuz.pacing.activeEndHour", "5")
+    assert gate_open(conn) is False
+
+
+def test_gate_closed_when_cap_reached_today(conn):
+    _set(conn, "qobuz.pacing.activeStartHour", "0")
+    _set(conn, "qobuz.pacing.activeEndHour", "24")
+    _set(conn, "qobuz.pacing.dailyCap", "3")
+    # Simulate 3 downloads already today.
+    with conn.cursor() as cur:
+        cur.execute("SELECT current_date::text")
+        today = cur.fetchone()[0]
+    _set(conn, "qobuz.pacing.countDay", today)
+    _set(conn, "qobuz.pacing.countToday", "3")
+    assert gate_open(conn) is False
+
+
+def test_gate_ignores_stale_count_from_a_previous_day(conn):
+    # A high count from yesterday must not keep the gate closed today.
+    _set(conn, "qobuz.pacing.activeStartHour", "0")
+    _set(conn, "qobuz.pacing.activeEndHour", "24")
+    _set(conn, "qobuz.pacing.dailyCap", "3")
+    _set(conn, "qobuz.pacing.countDay", "2000-01-01")
+    _set(conn, "qobuz.pacing.countToday", "99")
+    assert gate_open(conn) is True

worker/tests/test_qobuz_gate_pipeline.py

@@ -0,0 +1,85 @@
+"""The Qobuz budget gate, wired into the pipeline: when the gate is closed the job falls through
+to another source; when open it uses Qobuz and records the download; a Qobuz-only setup is never
+gated (nothing to fall through to)."""
+from lyra_worker.adapters.fakes import FakeQobuz, FakeSoulseek
+from lyra_worker.claim import claim_next
+from lyra_worker.pipeline import run_pipeline
+from tests.conftest import insert_request
+
+
+def _cfg(conn, key, value):
+    with conn.cursor() as cur:
+        cur.execute(
+            'INSERT INTO "Config"(key,value,secret,"updatedAt") VALUES (%s,%s,false,now()) '
+            'ON CONFLICT (key) DO UPDATE SET value=EXCLUDED.value, "updatedAt"=now()',
+            (key, value),
+        )
+    conn.commit()
+
+
+def _open_gate(conn):
+    _cfg(conn, "qobuz.pacing.activeStartHour", "0")
+    _cfg(conn, "qobuz.pacing.activeEndHour", "24")
+
+
+def _closed_gate(conn):
+    _cfg(conn, "qobuz.pacing.activeStartHour", "5")   # empty window (5..5) → always closed
+    _cfg(conn, "qobuz.pacing.activeEndHour", "5")
+
+
+def _state(conn, job_id):
+    with conn.cursor() as cur:
+        cur.execute('SELECT state FROM "Job" WHERE id = %s', (job_id,))
+        return cur.fetchone()[0]
+
+
+def _chosen(conn, job_id):
+    with conn.cursor() as cur:
+        cur.execute('SELECT source FROM "Candidate" WHERE "jobId" = %s AND chosen = true', (job_id,))
+        row = cur.fetchone()
+        return row[0] if row else None
+
+
+def _cfg_get(conn, key):
+    with conn.cursor() as cur:
+        cur.execute('SELECT value FROM "Config" WHERE key = %s', (key,))
+        row = cur.fetchone()
+        return row[0] if row else None
+
+
+def test_closed_gate_drops_qobuz_and_falls_through(conn):
+    _closed_gate(conn)
+    job_id = insert_request(conn, artist="Radiohead", album="In Rainbows")
+    claim_next(conn)
+    run_pipeline(conn, job_id, [FakeQobuz(), FakeSoulseek()], dest_root="/tmp/lib-gate")
+    assert _state(conn, job_id) == "imported"
+    assert _chosen(conn, job_id) == "soulseek"  # Qobuz dropped by the closed gate
+
+
+def test_open_gate_uses_qobuz_and_records_the_download(conn):
+    _open_gate(conn)
+    job_id = insert_request(conn, artist="Radiohead", album="In Rainbows")
+    claim_next(conn)
+    run_pipeline(conn, job_id, [FakeQobuz(), FakeSoulseek()], dest_root="/tmp/lib-gate")
+    assert _state(conn, job_id) == "imported"
+    assert _chosen(conn, job_id) == "qobuz"
+    assert _cfg_get(conn, "qobuz.pacing.countToday") == "1"  # recorded against today's budget
+    assert _cfg_get(conn, "qobuz.pacing.nextAllowedAt") is not None
+
+
+def test_qobuz_only_setup_is_never_gated(conn):
+    _closed_gate(conn)
+    job_id = insert_request(conn, artist="Radiohead", album="In Rainbows")
+    claim_next(conn)
+    run_pipeline(conn, job_id, [FakeQobuz()], dest_root="/tmp/lib-gate")
+    assert _state(conn, job_id) == "imported"
+    assert _chosen(conn, job_id) == "qobuz"  # not dropped — nothing to fall through to
+
+
+def test_non_qobuz_download_does_not_touch_the_counter(conn):
+    _closed_gate(conn)  # forces Soulseek
+    job_id = insert_request(conn, artist="Radiohead", album="In Rainbows")
+    claim_next(conn)
+    run_pipeline(conn, job_id, [FakeQobuz(), FakeSoulseek()], dest_root="/tmp/lib-gate")
+    assert _chosen(conn, job_id) == "soulseek"
+    assert _cfg_get(conn, "qobuz.pacing.countToday") is None  # only Qobuz downloads count

worker/tests/test_requeue.py

@@ -0,0 +1,69 @@
+from lyra_worker.claim import claim_next, requeue_or_fail
+from tests.conftest import insert_request
+
+
+def _set_attempts(conn, job_id, attempts, state="matching", stage="match"):
+    with conn.cursor() as cur:
+        cur.execute(
+            'UPDATE "Job" SET attempts = %s, state = %s, "currentStage" = %s, "claimedAt" = now() '
+            'WHERE id = %s',
+            (attempts, state, stage, job_id),
+        )
+    conn.commit()
+
+
+def _job(conn, job_id):
+    with conn.cursor() as cur:
+        cur.execute('SELECT state, "currentStage", "claimedAt", error FROM "Job" WHERE id = %s', (job_id,))
+        return cur.fetchone()
+
+
+def _request_status(conn, job_id):
+    with conn.cursor() as cur:
+        cur.execute(
+            'SELECT status FROM "Request" WHERE id = (SELECT "requestId" FROM "Job" WHERE id = %s)',
+            (job_id,),
+        )
+        return cur.fetchone()[0]
+
+
+def test_below_cap_requeues_to_requested(conn):
+    # A transient failure (e.g. MB SSL EOF) mid-'matching' must not strand the job — it returns
+    # to 'requested' so claim_next picks it up again instead of orphaning it.
+    jid = insert_request(conn)
+    _set_attempts(conn, jid, 2)
+    requeue_or_fail(conn, jid, "SSL: UNEXPECTED_EOF_WHILE_READING")
+    st, stage, claimed_at, err = _job(conn, jid)
+    assert st == "requested"
+    assert stage == "intake"
+    assert claimed_at is None
+    assert err is None
+    assert _request_status(conn, jid) == "pending"
+    assert claim_next(conn) == jid  # immediately re-claimable
+
+
+def test_at_cap_marks_needs_attention_with_error(conn):
+    jid = insert_request(conn)
+    _set_attempts(conn, jid, 5)  # hit the cap
+    requeue_or_fail(conn, jid, "SSL: UNEXPECTED_EOF_WHILE_READING")
+    st, _stage, claimed_at, err = _job(conn, jid)
+    assert st == "needs_attention"
+    assert err == "SSL: UNEXPECTED_EOF_WHILE_READING"
+    assert claimed_at is None
+    assert _request_status(conn, jid) == "needs_attention"
+    assert claim_next(conn) is None  # not re-claimed
+
+
+def test_requeues_until_cap_then_fails(conn):
+    jid = insert_request(conn)
+    for attempts in (1, 2, 3, 4):
+        _set_attempts(conn, jid, attempts)
+        requeue_or_fail(conn, jid, "boom")
+        assert _job(conn, jid)[0] == "requested"
+    _set_attempts(conn, jid, 5)
+    requeue_or_fail(conn, jid, "boom")
+    assert _job(conn, jid)[0] == "needs_attention"
+
+
+def test_missing_job_is_noop(conn):
+    requeue_or_fail(conn, "does-not-exist", "boom")  # must not raise

worker/tests/test_slskd_cancel.py

@@ -0,0 +1,218 @@
+"""slskd download behaviour: abandoned transfers are cancelled (no pile-up), a stalled peer is
+dropped fast, a slow-but-progressing peer is kept, and search candidates are ranked fast/free
+peers first."""
+import json
+
+import pytest
+
+import lyra_worker.adapters._slskd as slskd_mod
+from lyra_worker.adapters._slskd import (
+    _ETA_CAP_SECONDS,
+    _MAX_XFER_POLLS,
+    _MIN_XFER_POLLS,
+    _POLL_SECONDS,
+    SlskdClient,
+    _eta_seconds,
+    _max_polls,
+    _parse_search_responses,
+)
+
+
+class _Resp:
+    def __init__(self, payload=None):
+        self._p = payload if payload is not None else {}
+
+    def raise_for_status(self):
+        pass
+
+    def json(self):
+        return self._p
+
+
+class _SeqRequests:
+    """post OK; get() returns successive poll payloads (last one repeats); records DELETEs."""
+    def __init__(self, polls):
+        self._polls = list(polls)
+        self._i = 0
+        self.deleted = []
+
+    def post(self, url, **kw):
+        return _Resp({})
+
+    def get(self, url, **kw):
+        p = self._polls[min(self._i, len(self._polls) - 1)]
+        self._i += 1
+        return _Resp(p)
+
+    def delete(self, url, **kw):
+        self.deleted.append(url)
+        return _Resp({})
+
+
+def _poll(state, bytes_transferred, fid="t1"):
+    return {"directories": [{"files": [
+        {"filename": "f1.flac", "size": 100, "state": state,
+         "bytesTransferred": bytes_transferred, "id": fid},
+    ]}]}
+
+
+def _client(tmp_path):
+    return SlskdClient({"slskd.url": "http://x", "slskd.api_key": "k",
+                        "slskd.downloadsDir": str(tmp_path)})
+
+
+_REF = json.dumps({"username": "peerA", "files": [{"filename": "f1.flac", "size": 100}]})
+
+
+def test_stalled_peer_is_abandoned_and_cancelled(tmp_path, monkeypatch):
+    # No byte progress across polls → stall → TimeoutError → the transfer is cancelled.
+    monkeypatch.setattr(slskd_mod, "requests", _SeqRequests([_poll("InProgress", 0)]))
+    monkeypatch.setattr(slskd_mod.time, "sleep", lambda _s: None)
+    monkeypatch.setattr(slskd_mod, "_STALL_POLLS", 3)
+    fake = slskd_mod.requests
+
+    with pytest.raises(TimeoutError):
+        _client(tmp_path).download(_REF, str(tmp_path / "dest"), lambda *_a: None)
+    assert any(u.endswith("/api/v0/transfers/downloads/peerA/t1") for u in fake.deleted)
+
+
+def test_error_state_is_cancelled(tmp_path, monkeypatch):
+    monkeypatch.setattr(slskd_mod, "requests", _SeqRequests([_poll("Completed, Errored", 0)]))
+    monkeypatch.setattr(slskd_mod.time, "sleep", lambda _s: None)
+    fake = slskd_mod.requests
+    with pytest.raises(RuntimeError):
+        _client(tmp_path).download(_REF, str(tmp_path / "dest"), lambda *_a: None)
+    assert any(u.endswith("/peerA/t1") for u in fake.deleted)
+
+
+def test_slow_but_progressing_peer_is_not_abandoned(tmp_path, monkeypatch):
+    # Bytes climb for more polls than _STALL_POLLS, then complete → success, not a premature stall.
+    (tmp_path / "01.flac").write_bytes(b"audio")  # so _retrieve finds the finished file
+    ref = json.dumps({"username": "peerA", "files": [{"filename": r"d\Album\01.flac", "size": 5}]})
+    polls = [
+        {"directories": [{"files": [{"filename": r"d\Album\01.flac", "size": 5,
+                                     "state": "InProgress", "bytesTransferred": b, "id": "t1"}]}]}
+        for b in (10, 20, 30, 40, 50)
+    ] + [
+        {"directories": [{"files": [{"filename": r"d\Album\01.flac", "size": 5,
+                                     "state": "Completed, Succeeded", "bytesTransferred": 5, "id": "t1"}]}]}
+    ]
+    monkeypatch.setattr(slskd_mod, "requests", _SeqRequests(polls))
+    monkeypatch.setattr(slskd_mod.time, "sleep", lambda _s: None)
+    monkeypatch.setattr(slskd_mod, "_STALL_POLLS", 2)  # would fire early if progress were ignored
+
+    out = _client(tmp_path).download(ref, str(tmp_path / "dest"), lambda *_a: None)
+    assert out["track_count"] == 1
+    assert slskd_mod.requests.deleted == []  # completed cleanly → nothing cancelled
+
+
+def test_parse_ranks_free_and_fast_peers_first():
+    responses = [
+        {"username": "slow", "hasFreeUploadSlot": True, "uploadSpeed": 100, "queueLength": 0,
+         "files": [{"filename": r"a\Artist - Album\01.flac", "size": 5}]},
+        {"username": "fast", "hasFreeUploadSlot": True, "uploadSpeed": 9000, "queueLength": 0,
+         "files": [{"filename": r"b\Artist - Album\01.flac", "size": 5}]},
+        {"username": "queued", "hasFreeUploadSlot": False, "uploadSpeed": 99999, "queueLength": 40,
+         "files": [{"filename": r"c\Artist - Album\01.flac", "size": 5}]},
+    ]
+    peers = [json.loads(c["source_ref"])["username"] for c in _parse_search_responses(responses)]
+    assert peers == ["fast", "slow", "queued"]  # free+fast first; no-slot peer last despite speed
+
+
+def test_parse_uses_requested_artist_when_present_in_path():
+    # Real "Rated R" MP3-over-FLAC bug: the artist is in a PARENT folder ("Rihanna\(2009) Rated
+    # R\…"), so the "Artist - Album" folder parse yields no artist and confidence tanks. When the
+    # requested artist is in the path, trust it so the candidate scores as Rihanna.
+    responses = [
+        {"username": "p", "hasFreeUploadSlot": True, "uploadSpeed": 1000, "queueLength": 0,
+         "files": [{"filename": r"@@x\Music\Rihanna\(2009) Rated R\01. Mad House.flac", "size": 5}]},
+    ]
+    c = _parse_search_responses(responses, artist="Rihanna")[0]
+    assert c["artist"] == "Rihanna"
+
+
+def test_parse_without_artist_falls_back_to_folder_parse():
+    responses = [
+        {"username": "p", "hasFreeUploadSlot": True, "uploadSpeed": 1000, "queueLength": 0,
+         "files": [{"filename": r"x\Some Band - Some Album\01.flac", "size": 5}]},
+    ]
+    c = _parse_search_responses(responses)[0]  # no artist arg → folder parse
+    assert c["artist"] == "Some Band"
+
+
+def test_parse_prefers_shorter_transfer_over_raw_speed():
+    # The Hybrid Theory case: a peer advertising higher speed but serving a ~2x-larger rip (24-bit
+    # / bloated FLAC) should lose to a peer with lower speed but a smaller standard rip that
+    # finishes sooner — both are the same quality class to Lyra, so faster-to-finish wins.
+    responses = [
+        {"username": "bloated", "hasFreeUploadSlot": True, "uploadSpeed": 20000, "queueLength": 0,
+         "files": [{"filename": rf"x\LP - Album\{i:02}.flac", "size": 40_000_000} for i in range(12)]},
+        {"username": "lean", "hasFreeUploadSlot": True, "uploadSpeed": 15000, "queueLength": 0,
+         "files": [{"filename": rf"y\LP - Album\{i:02}.flac", "size": 24_000_000} for i in range(12)]},
+    ]
+    # bloated est = 480MB/20000 = 24000s; lean est = 288MB/15000 = 19200s → lean finishes first
+    peers = [json.loads(c["source_ref"])["username"] for c in _parse_search_responses(responses)]
+    assert peers[0] == "lean"
+
+
+def test_eta_seconds_from_measured_rate():
+    assert _eta_seconds(1000, 250, 250.0) == 3  # 750 bytes left / 250 B/s
+    assert _eta_seconds(1000, 0, 100.0) == 10
+
+
+def test_eta_seconds_unknown_cases():
+    assert _eta_seconds(1000, 250, 0.0) is None      # no speed sample yet
+    assert _eta_seconds(0, 0, 100.0) is None          # unknown total size
+    assert _eta_seconds(1000, 1000, 100.0) is None    # already complete
+
+
+def test_max_polls_scales_backstop_to_album_size():
+    # A 521MB album (Drake "Scorpion", 25 FLACs) from a slow serial peer must get far more than
+    # the old flat 1200-poll/60min budget, or it gets abandoned mid-download and restarts from zero.
+    scorpion_bytes = 521_000_000
+    polls = _max_polls(scorpion_bytes)
+    assert polls > 1200                              # more headroom than the old flat cap
+    assert polls * _POLL_SECONDS >= 90 * 60          # at least ~90 min of wall-clock budget
+
+    # A tiny single stays on the floor, not zero.
+    assert _max_polls(3_000_000) == _MIN_XFER_POLLS
+    assert _max_polls(0) == _MIN_XFER_POLLS
+
+    # An enormous set is clamped to the ceiling (bounds worst-case slot hold).
+    assert _max_polls(50_000_000_000) == _MAX_XFER_POLLS
+
+    # Monotonic: bigger album never gets a smaller budget.
+    assert _max_polls(200_000_000) <= _max_polls(400_000_000)
+
+
+def test_eta_seconds_capped_when_peer_nearly_stalls():
+    # A near-stalled peer (tiny speed) would yield an ETA far past int4, overflowing the
+    # downloadEtaSeconds column write. The estimate must be clamped to a safe ceiling.
+    eta = _eta_seconds(30_000_000, 0, 0.001)  # 30MB / 0.001 B/s ~= 3e10s uncapped
+    assert eta == _ETA_CAP_SECONDS
+    assert eta < 2_147_483_647  # fits in int4
+
+
+def test_download_reports_shrinking_eta(tmp_path, monkeypatch):
+    # A 900-byte file arriving 300 bytes per 3s poll → measured speed 100 B/s → ETA counts down
+    # (6s, 3s) then None on completion. on_progress gets (fraction, eta_seconds).
+    (tmp_path / "01.flac").write_bytes(b"x" * 900)  # so _retrieve finds the finished file
+    ref = json.dumps({"username": "peerA", "files": [{"filename": r"d\Album\01.flac", "size": 900}]})
+    polls = [
+        {"directories": [{"files": [{"filename": r"d\Album\01.flac", "size": 900,
+                                     "state": "InProgress", "bytesTransferred": b, "id": "t1"}]}]}
+        for b in (300, 600)
+    ] + [
+        {"directories": [{"files": [{"filename": r"d\Album\01.flac", "size": 900,
+                                     "state": "Completed, Succeeded", "bytesTransferred": 900, "id": "t1"}]}]}
+    ]
+    monkeypatch.setattr(slskd_mod, "requests", _SeqRequests(polls))
+    monkeypatch.setattr(slskd_mod.time, "sleep", lambda _s: None)
+    clock = iter(float(i) * 3 for i in range(100))  # 0,3,6,… deterministic monotonic time
+    monkeypatch.setattr(slskd_mod.time, "monotonic", lambda: next(clock))
+
+    calls: list = []
+    _client(tmp_path).download(ref, str(tmp_path / "dest"), lambda *a: calls.append(a))
+
+    etas = [a[1] for a in calls if len(a) > 1]
+    assert etas == [6, 3, None]  # counts down at 100 B/s, then None once complete

worker/tests/test_slskd_search_fallback.py

@@ -0,0 +1,150 @@
+"""Soulseek search fallback: the Soulseek server drops searches containing blocklisted terms
+(major-label artists / album titles) → 0 responses. search_album retries with album-only then
+artist-only, keeping only path-matching results so it never grabs a same-titled foreign album."""
+import json
+
+import lyra_worker.adapters._slskd as slskd_mod
+from lyra_worker.adapters._slskd import (
+    SlskdClient,
+    _keep_matching,
+    _norm,
+    _path_matches,
+)
+
+
+# --- pure helpers -----------------------------------------------------------------------------
+
+def test_norm_strips_diacritics_and_punctuation():
+    assert _norm("Beyoncé") == "beyonce"
+    assert _norm("I Am… Sasha Fierce") == "i am sasha fierce"
+    assert _norm("AC/DC") == "ac dc"
+
+
+def test_path_matches_is_accent_and_punctuation_tolerant():
+    p = r"@@bckfj\All\Beyonce\I Am... Sasha Fierce CD1\03 Halo.flac"
+    assert _path_matches("Beyoncé", p)              # accent folded
+    assert _path_matches("I Am… Sasha Fierce", p)   # ellipsis / punctuation ignored
+    assert not _path_matches("Rihanna", p)
+
+
+def test_path_matches_short_needle_requires_whole_string():
+    assert _path_matches("U2", r"x\U2 - Achtung Baby\01.flac")
+    assert not _path_matches("U2", r"x\Blur - 13\01.flac")
+
+
+def test_keep_matching_drops_foreign_and_empty_responses():
+    responses = [
+        {"username": "a", "files": [{"filename": r"x\Rihanna - Unapologetic\01.flac"}]},
+        {"username": "b", "files": [{"filename": r"x\Someone Else - Unapologetic\01.flac"}]},
+    ]
+    kept = _keep_matching(responses, "Rihanna")
+    assert [r["username"] for r in kept] == ["a"]
+
+
+# --- fallback orchestration (HTTP faked) ------------------------------------------------------
+
+class _Resp:
+    def __init__(self, payload):
+        self._p = payload
+
+    def raise_for_status(self):
+        pass
+
+    def json(self):
+        return self._p
+
+
+class _FakeSearch:
+    """Routes slskd search HTTP by URL: POST returns an id, GET on the search returns a completed
+    state, GET on /responses returns the canned responses for that query. Records queries sent."""
+    def __init__(self, by_query):
+        self.by_query = by_query          # {searchText: [responses]}
+        self.queries = []                 # searchText values POSTed, in order
+        self._id_query = {}
+        self._n = 0
+
+    def post(self, url, **kw):
+        text = kw["json"]["searchText"]
+        self.queries.append(text)
+        self._n += 1
+        sid = f"s{self._n}"
+        self._id_query[sid] = text
+        return _Resp({"id": sid})
+
+    def get(self, url, **kw):
+        if url.endswith("/responses"):
+            sid = url.rsplit("/", 2)[-2]
+            return _Resp(self.by_query.get(self._id_query[sid], []))
+        return _Resp({"state": "Completed, TimedOut"})
+
+
+def _client():
+    return SlskdClient({"slskd.url": "http://x", "slskd.api_key": "k", "slskd.downloadsDir": "/x"})
+
+
+def _install(monkeypatch, fake):
+    monkeypatch.setattr(slskd_mod, "requests", fake)
+    monkeypatch.setattr(slskd_mod.time, "sleep", lambda _s: None)
+
+
+def test_primary_hit_does_not_trigger_fallback(monkeypatch):
+    fake = _FakeSearch({
+        "Rihanna Unapologetic": [
+            {"username": "a", "hasFreeUploadSlot": True, "uploadSpeed": 1, "queueLength": 0,
+             "files": [{"filename": r"x\Rihanna - Unapologetic\01.flac", "size": 5}]},
+        ],
+    })
+    _install(monkeypatch, fake)
+    out = _client().search_album("Rihanna", "Unapologetic")
+    assert len(out) == 1
+    assert fake.queries == ["Rihanna Unapologetic"]  # no fallback searches
+
+
+def test_album_only_fallback_recovers_blocklisted_artist(monkeypatch):
+    # Primary (artist blocked) → 0; album-only returns the artist's folder → recovered.
+    fake = _FakeSearch({
+        "Beyoncé I Am… Sasha Fierce": [],  # blocked
+        "i am sasha fierce": [
+            {"username": "a", "hasFreeUploadSlot": True, "uploadSpeed": 1, "queueLength": 0,
+             "files": [{"filename": r"x\Beyonce - I Am... Sasha Fierce\01.flac", "size": 5}]},
+            {"username": "b", "hasFreeUploadSlot": True, "uploadSpeed": 1, "queueLength": 0,
+             "files": [{"filename": r"x\Other Artist - I Am Sasha Fierce\01.flac", "size": 5}]},
+        ],
+    })
+    _install(monkeypatch, fake)
+    out = _client().search_album("Beyoncé", "I Am… Sasha Fierce")
+    assert fake.queries == ["Beyoncé I Am… Sasha Fierce", "i am sasha fierce"]
+    peers = [json.loads(c["source_ref"])["username"] for c in out]
+    assert peers == ["a"]  # foreign same-titled album dropped
+
+
+def test_artist_only_fallback_when_album_title_blocked(monkeypatch):
+    # Album-only also blocked/empty → fall through to artist-only, filtered by album in path.
+    fake = _FakeSearch({
+        "Foo Lemonade": [],
+        "lemonade": [],  # album title itself blocked
+        "foo": [
+            {"username": "a", "hasFreeUploadSlot": True, "uploadSpeed": 1, "queueLength": 0,
+             "files": [{"filename": r"x\Foo - Lemonade\01.flac", "size": 5}]},
+        ],
+    })
+    _install(monkeypatch, fake)
+    out = _client().search_album("Foo", "Lemonade")
+    assert fake.queries == ["Foo Lemonade", "lemonade", "foo"]
+    assert len(out) == 1
+
+
+def test_fallback_yields_nothing_when_title_too_generic(monkeypatch):
+    # 'Adele 21': artist blocked, album title '21' returns lots of non-Adele content → filtered
+    # to nothing rather than grabbing a wrong album.
+    fake = _FakeSearch({
+        "Adele 21": [],
+        "21": [
+            {"username": "z", "hasFreeUploadSlot": True, "uploadSpeed": 1, "queueLength": 0,
+             "files": [{"filename": r"x\21 Savage - Savage Mode\01.flac", "size": 5}]},
+        ],
+        "adele": [],  # artist blocked
+    })
+    _install(monkeypatch, fake)
+    out = _client().search_album("Adele", "21")
+    assert out == []

worker/tests/test_smaller_edition_completeness.py

@@ -0,0 +1,101 @@
+"""MusicBrainz picks an arbitrary releases[0] that is often a deluxe/expanded edition with more
+tracks (and a longer runtime) than the standard album a source legitimately delivers. Completeness
+and the duration guard must be judged against what the CHOSEN SOURCE offered, not MB's canonical
+edition — otherwise a complete standard-edition download is falsely rejected as 'incomplete'
+(The Script "No Sound Without Silence": 11 delivered vs MB 12) or 'too short' (Skillet "Unleashed":
+12 delivered ~40 min vs MB's 20-track deluxe ~73 min). A genuinely short download must still fail.
+"""
+import os
+from dataclasses import replace
+
+from lyra_worker.adapters.fakes import FakeQobuz
+from lyra_worker.claim import claim_next
+from lyra_worker.pipeline import run_pipeline
+from lyra_worker.types import DownloadResult, MBTarget
+from tests.conftest import insert_request
+
+
+class _StandardEditionQobuz(FakeQobuz):
+    """Qobuz-like source offering a fixed-size *standard* edition (source_tracks), independent of
+    MB's possibly-larger deluxe count; stages that many real (non-empty) files and reports them."""
+    def __init__(self, source_tracks):
+        super().__init__()
+        self._n = source_tracks
+
+    def search(self, target):
+        return super().search(replace(target, track_count=self._n))
+
+    def download(self, candidate, dest, on_progress):
+        os.makedirs(dest, exist_ok=True)
+        for i in range(candidate.track_count):
+            with open(os.path.join(dest, f"{i + 1:02d} Track.flac"), "wb") as fh:
+                fh.write(b"audio")
+        on_progress(1.0)
+        return DownloadResult(ok=True, path=dest, track_count=candidate.track_count)
+
+
+class _PartialQobuz(FakeQobuz):
+    """Promises `promised` tracks but one fails mid-download: stages and reports `promised - 1`
+    (the real John Mayer "Roll it on Home" IncompleteRead case)."""
+    def __init__(self, promised):
+        super().__init__()
+        self._n = promised
+
+    def search(self, target):
+        return super().search(replace(target, track_count=self._n))
+
+    def download(self, candidate, dest, on_progress):
+        os.makedirs(dest, exist_ok=True)
+        delivered = candidate.track_count - 1
+        for i in range(delivered):
+            with open(os.path.join(dest, f"{i + 1:02d} Track.flac"), "wb") as fh:
+                fh.write(b"audio")
+        return DownloadResult(ok=True, path=dest, track_count=delivered)
+
+
+class _DeluxeResolver:
+    """Resolver whose canonical release is a larger (deluxe) edition than the source delivers."""
+    def __init__(self, mb_tracks, mb_total_s):
+        self._t, self._s = mb_tracks, mb_total_s
+
+    def resolve(self, artist, album):
+        return MBTarget(artist=artist, album=album, track_count=self._t, total_duration_s=self._s)
+
+
+def _state_error(conn, job_id):
+    with conn.cursor() as cur:
+        cur.execute('SELECT state, error FROM "Job" WHERE id = %s', (job_id,))
+        return cur.fetchone()
+
+
+def test_standard_edition_not_rejected_as_incomplete(conn):
+    # The Script: source delivers the 11-track standard edition; MB's canonical release is 12.
+    job_id = insert_request(conn, artist="The Script", album="No Sound Without Silence")
+    claim_next(conn)
+    run_pipeline(conn, job_id, [_StandardEditionQobuz(source_tracks=11)],
+                 resolver=_DeluxeResolver(mb_tracks=12, mb_total_s=2943),
+                 measure_duration=lambda _s: 2700.0, dest_root="/tmp/lib-se")
+    assert _state_error(conn, job_id) == ("imported", None)
+
+
+def test_smaller_edition_not_rejected_as_too_short(conn):
+    # Skillet: source delivers the 12-track standard (~40 min); MB's release is a 20-track deluxe
+    # (~73 min). The unscaled duration guard falsely rejects; scaling to the delivered edition
+    # (12/20) must let it pass.
+    job_id = insert_request(conn, artist="Skillet", album="Unleashed")
+    claim_next(conn)
+    run_pipeline(conn, job_id, [_StandardEditionQobuz(source_tracks=12)],
+                 resolver=_DeluxeResolver(mb_tracks=20, mb_total_s=4380),
+                 measure_duration=lambda _s: 2400.0, dest_root="/tmp/lib-se")
+    assert _state_error(conn, job_id) == ("imported", None)
+
+
+def test_genuinely_incomplete_still_rejected(conn):
+    # John Mayer: the source promised 12 tracks but one failed mid-download → only 11 land.
+    # This is a real partial and must still be rejected (the fix must not swallow it).
+    job_id = insert_request(conn, artist="John Mayer", album="The Search for Everything")
+    claim_next(conn)
+    run_pipeline(conn, job_id, [_PartialQobuz(promised=12)],
+                 resolver=_DeluxeResolver(mb_tracks=12, mb_total_s=2580),
+                 measure_duration=lambda _s: 2400.0, dest_root="/tmp/lib-se")
+    assert _state_error(conn, job_id) == ("needs_attention", "incomplete download")

worker/tests/test_streamrip_progress.py

@@ -0,0 +1,41 @@
+from lyra_worker.adapters._streamrip import _ProgressAggregator
+
+
+def _record():
+    seen: list[float] = []
+    return seen, seen.append
+
+
+def test_equal_size_tracks_aggregate_monotonically_to_near_one():
+    # 4 tracks of 100 bytes each, all started up front (concurrency covers the album), then
+    # each downloaded in 25-byte chunks. Fraction must climb monotonically 0→~1.
+    seen, cb = _record()
+    agg = _ProgressAggregator(4, cb)
+    for _ in range(4):
+        agg.start_track(100)
+    for _ in range(16):  # 16 chunks * 25 bytes = 400 bytes = full album
+        agg.advance(25)
+
+    assert seen == sorted(seen)                 # monotonic non-decreasing
+    assert all(0.0 <= f <= 0.99 for f in seen)  # never negative, capped at 0.99
+    assert seen[-1] == 0.99                      # reaches the cap when all bytes are in
+    assert seen[len(seen) // 2] > 0.3           # genuinely climbs mid-download (no 0→100 jump)
+
+
+def test_estimate_uses_track_count_before_all_tracks_start():
+    # Only the first of 4 tracks has started; with a known track count the estimate already
+    # spans the whole album, so finishing track 1 reports ~1/4, not ~1/1.
+    seen, cb = _record()
+    agg = _ProgressAggregator(4, cb)
+    agg.start_track(100)
+    agg.advance(100)  # track 1 fully downloaded
+    assert abs(seen[-1] - 0.25) < 0.01
+
+
+def test_zero_or_missing_sizes_never_emit_and_never_divide_by_zero():
+    seen, cb = _record()
+    agg = _ProgressAggregator(0, cb)
+    agg.advance(50)          # no track started yet -> no emit
+    agg.start_track(0)       # a track with unknown size
+    agg.advance(50)          # est_total still 0 -> still no emit, no crash
+    assert seen == []

worker/tests/test_upgrade_guard.py

@@ -0,0 +1,133 @@
+"""Upgrade guard: when an album is already in the library and no currently-available source can
+beat it (e.g. Qobuz is gated by pacing and only a same-quality Soulseek copy is left), the pipeline
+must NOT waste a download — keep-if-better would discard it anyway. It should skip and let the
+release be re-attempted later (when Qobuz frees up). A genuinely better source still upgrades."""
+import os
+
+from lyra_worker.claim import claim_next
+from lyra_worker.pipeline import run_pipeline
+from lyra_worker.types import Candidate, DownloadResult, Quality
+
+_CD = Quality(fmt="FLAC", lossless=True, bit_depth=16, sample_rate=44100)     # class 2
+_HIRES = Quality(fmt="FLAC", lossless=True, bit_depth=24, sample_rate=96000)  # class 3
+
+
+def _cfg(conn, key, value):
+    with conn.cursor() as cur:
+        cur.execute(
+            'INSERT INTO "Config"(key,value,secret,"updatedAt") VALUES (%s,%s,false,now()) '
+            'ON CONFLICT (key) DO UPDATE SET value=EXCLUDED.value, "updatedAt"=now()',
+            (key, value),
+        )
+    conn.commit()
+
+
+class _Counting:
+    def __init__(self, name, tier, quality):
+        self.name, self.tier, self._q = name, tier, quality
+        self.calls = 0
+        self.searches = 0
+
+    def health(self):
+        return True
+
+    def search(self, target):
+        self.searches += 1
+        return [Candidate(source=self.name, source_ref=self.name, matched_artist=target.artist,
+                          matched_album=target.album, quality=self._q, track_count=3, source_tier=self.tier)]
+
+    def download(self, candidate, dest, on_progress):
+        self.calls += 1
+        os.makedirs(dest, exist_ok=True)
+        for i in range(3):
+            with open(os.path.join(dest, f"0{i}.flac"), "wb") as fh:
+                fh.write(b"a")
+        return DownloadResult(ok=True, path=dest, track_count=3)
+
+
+def _seed_library(conn, artist, album, quality_class, path):
+    with conn.cursor() as cur:
+        cur.execute(
+            'INSERT INTO "LibraryItem" (id, artist, album, path, source, format, "qualityClass", "importedAt") '
+            "VALUES (gen_random_uuid()::text, %s, %s, %s, 'soulseek', 'FLAC', %s, now())",
+            (artist, album, path, quality_class),
+        )
+    conn.commit()
+
+
+def _insert_upgrade_job(conn, artist, album):
+    with conn.cursor() as cur:
+        cur.execute(
+            'INSERT INTO "MonitoredRelease" (id, "artistMbid", "artistName", "rgMbid", album, '
+            'monitored, state, "secondaryTypes", "createdAt") '
+            "VALUES (gen_random_uuid()::text, 'm', %s, 'rg', %s, true, 'grabbed', '{}', now()) RETURNING id",
+            (artist, album),
+        )
+        rel_id = cur.fetchone()[0]
+        cur.execute(
+            'INSERT INTO "Request" (id, artist, album, status, "monitoredReleaseId", "createdAt") '
+            "VALUES (gen_random_uuid()::text, %s, %s, 'pending', %s, now()) RETURNING id",
+            (artist, album, rel_id),
+        )
+        req = cur.fetchone()[0]
+        cur.execute(
+            'INSERT INTO "Job" (id, "requestId", state, "currentStage", attempts, "createdAt", "updatedAt") '
+            "VALUES (gen_random_uuid()::text, %s, 'requested', 'intake', 0, now(), now()) RETURNING id",
+            (req,),
+        )
+        job = cur.fetchone()[0]
+    conn.commit()
+    return job
+
+
+def _lib(conn, artist, album):
+    with conn.cursor() as cur:
+        cur.execute('SELECT path, "qualityClass" FROM "LibraryItem" WHERE artist=%s AND album=%s', (artist, album))
+        return cur.fetchone()
+
+
+def _state(conn, job_id):
+    with conn.cursor() as cur:
+        cur.execute('SELECT state FROM "Job" WHERE id=%s', (job_id,))
+        return cur.fetchone()[0]
+
+
+def test_upgrade_skipped_pre_search_when_qobuz_gated(conn):
+    # Qobuz gated (empty active window) + existing >= CD copy: only hi-res Qobuz could improve it,
+    # so defer WITHOUT even searching — avoids re-searching every owned album each monitor cycle.
+    _cfg(conn, "qobuz.pacing.activeStartHour", "5")
+    _cfg(conn, "qobuz.pacing.activeEndHour", "5")  # empty window → gate closed
+    _seed_library(conn, "A", "B", 2, "/m/existing")
+    job_id = _insert_upgrade_job(conn, "A", "B")
+    claim_next(conn)
+    soulseek = _Counting("soulseek", 1, _CD)
+    run_pipeline(conn, job_id, [soulseek], dest_root="/tmp/lib-ug", upgrade_cutoff=3)
+    assert soulseek.searches == 0                   # no wasted search
+    assert soulseek.calls == 0                      # no wasted download
+    assert _lib(conn, "A", "B") == ("/m/existing", 2)
+    assert _state(conn, job_id) == "imported"
+
+
+def test_upgrade_skipped_post_search_when_only_same_quality_found(conn):
+    # Gate OPEN, so it searches, but the only source found is same-quality → skip the download.
+    _cfg(conn, "qobuz.pacing.activeStartHour", "0")
+    _cfg(conn, "qobuz.pacing.activeEndHour", "24")  # gate open
+    _seed_library(conn, "A", "B", 2, "/m/existing")
+    job_id = _insert_upgrade_job(conn, "A", "B")
+    claim_next(conn)
+    soulseek = _Counting("soulseek", 1, _CD)
+    run_pipeline(conn, job_id, [soulseek], dest_root="/tmp/lib-ug", upgrade_cutoff=3)
+    assert soulseek.searches == 1                   # it did search (gate open)
+    assert soulseek.calls == 0                      # but skipped the useless download
+    assert _lib(conn, "A", "B") == ("/m/existing", 2)
+
+
+def test_upgrade_proceeds_when_a_better_source_is_available(conn):
+    # Existing class-2 copy; a class-3 (hi-res) source IS available → download + upgrade.
+    _seed_library(conn, "A", "B", 2, "/m/existing")
+    job_id = _insert_upgrade_job(conn, "A", "B")
+    claim_next(conn)
+    qobuz = _Counting("qobuz", 0, _HIRES)
+    run_pipeline(conn, job_id, [qobuz], dest_root="/tmp/lib-ug2", upgrade_cutoff=3)
+    assert qobuz.calls == 1                          # downloaded the better copy
+    assert _lib(conn, "A", "B")[1] == 3              # upgraded to hi-res