Audible: HTML refresh, multi-narrator & works dedup

Switch nightly discovery refresh to scrape Audible's curated HTML storefronts (popular, new releases, category pages) while keeping real-time user paths on the JSON catalog API. Add robust HTML resilience knobs (increased retries, capped jittered backoff, AdaptivePacer changes and per-batch cooldowns) to avoid failing nightly jobs during 503 storms. Implement multi-narrator capture via a new extractAllNarrators helper and update parsers to preserve all narrator anchors. Introduce two-pass dedup: in-memory deduplicateAndCollectGroups + collapseByExistingWorks that consults the works table, export metadataScore for consistent representative selection, and persist dedup groups (fire-and-forget). Wire collapseByExistingWorks into search/author/series routes and make defensive dedup in the refresh processor. Add HTML parsing helpers, runtime/lang-aware parsing, jitteredBackoff cap, and tests for the new behaviors.

documentation/TABLEOFCONTENTS.md

@@ -45,6 +45,8 @@
 - **Web scraping (popular, new releases)** → [integrations/audible.md](integrations/audible.md)
 - **Database caching, real-time matching** → [integrations/audible.md](integrations/audible.md)
 - **Book covers API for login page** → [frontend/pages/login.md](frontend/pages/login.md)
+- **Dedup & works table (cross-ASIN identity)** → [integrations/audible.md](integrations/audible.md#dedup--works-table)
+- **Multi-narrator capture in HTML scrapers** → [integrations/audible.md](integrations/audible.md#narrator-capture-in-html-scrapers)
 
 ## E-book Support (First-Class)
 - **First-class ebook requests, separate tracking** → [integrations/ebook-sidecar.md](integrations/ebook-sidecar.md)

documentation/integrations/audible.md

@@ -1,29 +1,40 @@
 # Audible Integration
 
-**Status:** Implemented | Unauthenticated Audible JSON catalog API (primary) + Audnexus API (per-ASIN details)
+**Status:** Implemented | Hybrid — curated HTML for discovery refresh + Audible JSON catalog API for user-facing real-time + Audnexus for per-ASIN details
 
 ## Overview
 
-Audiobook metadata for discovery, search, and detail pages. All catalog operations (search, popular, new releases, categories, category books, author books, single-product details) now call Audible's unauthenticated public JSON catalog API (`api.audible.<tld>/1.0/catalog/*`). Per-ASIN detail lookups prefer Audnexus; the catalog API is used as fallback.
+Audiobook metadata for discovery, search, and detail pages. Split by access pattern:
+
+- **Nightly discovery refresh** (popular / new releases / category lists) — scraped from Audible's **curated HTML storefronts** (`www.audible.<tld>/adblbestsellers`, `/newreleases`, `/search?node=<id>`). The HTML pages reflect Audible's own editorial picks.
+- **User-facing real-time** (search, author books, categories listing, per-ASIN details) — Audible's unauthenticated public **JSON catalog API** (`api.audible.<tld>/1.0/catalog/*`).
+- **Per-ASIN detail lookups** — Audnexus (`api.audnex.us/books/{asin}`) primary; catalog API used as fallback when Audnexus returns 404.
 
 ## Architecture
 
-- **Primary data source:** Audible JSON catalog API, same endpoint used by the official Audible mobile apps. No authentication, no API key, no user credentials, no special headers.
-- **Per-ASIN details:** Audnexus (`api.audnex.us/books/{asin}`) remains primary; catalog API (`/1.0/catalog/products/{asin}`) is the fallback when Audnexus returns 404.
-- **HTML scraping:** Removed from `audible.service.ts`. The only remaining HTML path is `audible-series.ts` (series-page scraping, out of scope).
-- **`www.audible.<tld>`:** Still used by `audible-series.ts` and by `getBaseUrl()` for "View on Audible" link generation. Not used for any catalog operation.
+- **Curated HTML (refresh job only):** the three methods called solely by `audible-refresh.processor.ts` (`getPopularAudiobooks`, `getNewReleases`, `getCategoryBooks`) scrape Audible's storefront HTML to inherit editorial curation. Beefed-up retry/backoff knobs (12 retries, 3-min jittered cap) handle 503 storms patiently on the nightly job without slowing healthy users.
+- **JSON catalog API (real-time):** `search`, `searchByAuthorAsin`, `getCategories` (categories listing), and `fetchAudibleDetailsFromApi` (per-ASIN fallback). Same endpoint used by the official Audible mobile apps. No authentication, no API key, no user credentials, no special headers.
+- **Audnexus (per-ASIN):** `getAudiobookDetails` and `getRuntime` prefer Audnexus, with catalog API fallback for `getAudiobookDetails`.
+- **`www.audible.<tld>`:** Used by HTML refresh scraping, by `audible-series.ts`, and by `getBaseUrl()` for "View on Audible" link generation.
 
 ## Data Sources
 
-All catalog operations are HTTP GET against `{apiBaseUrl}` (region-dependent, e.g. `https://api.audible.com`):
+### Nightly refresh (HTML — `htmlClient`, baseURL `www.audible.<tld>`)
+
+| Operation | Endpoint | Key params |
+|---|---|---|
+| Popular | `/adblbestsellers` | `pageSize=50`, `page=<n>` (omitted on first page) |
+| New releases | `/newreleases` | `pageSize=50`, `page=<n>` (omitted on first page) |
+| Category books | `/search` | `node=<categoryId>&pageSize=50&sort=popularity-rank&page=<n>` |
+
+Parsed via cheerio. Selectors: `.productListItem` (popular/new releases), `.s-result-item, .productListItem` (categories).
+
+### Real-time (JSON catalog API — `apiClient`, baseURL `api.audible.<tld>`)
 
 | Operation | Endpoint | Key params |
 |---|---|---|
 | Search | `/1.0/catalog/products` | `keywords=<q>` |
 | Author books | `/1.0/catalog/products` | `author=<name>` (name, NOT ASIN) |
-| Popular | `/1.0/catalog/products` | `products_sort_by=BestSellers` |
-| New releases | `/1.0/catalog/products` | `products_sort_by=-ReleaseDate` |
-| Category books | `/1.0/catalog/products` | `category_id=<id>&products_sort_by=BestSellers` |
 | Categories listing | `/1.0/catalog/categories` | (none) |
 | Single product | `/1.0/catalog/products/{asin}` | — |
 | Audnexus (per-ASIN) | `https://api.audnex.us/books/{asin}` | `region={audnexusParam}` |
@@ -48,20 +59,20 @@ Populates every `AudibleAudiobook` field. Covered:
 
 ## Gotchas
 
+- **Catalog API cannot filter preorders or surface curated bestsellers.** The API's `BestSellers` sort is a right-now velocity rank that spikes on launch-day promos and preorder windows; the `-ReleaseDate` sort returns 100% future preorders. There is no server-side `release_time`, `released-only`, `customer_rights`, or alternate sort (`Reviewed`, `MostListened`, etc.) — every plausible variant was tested and silently ignored. This is why the nightly refresh job uses the curated HTML storefront pages instead.
 - **`author=` takes a name, not an ASIN.** The catalog API has no ASIN-based author param. `searchByAuthorAsin()` queries by name, then filters client-side: keeps only products where `products[].authors[].asin === authorAsin`. Preserves ASIN-authoritative author identity. Also filters by `product.language` via `isAcceptedLanguage()` for the configured region.
 - **Invalid ASIN returns HTTP 200 with stub body.** `/1.0/catalog/products/{asin}` responds 200 with `{product: {asin: INPUT}}` and no other fields. `fetchAudibleDetailsFromApi()` detects this via missing `product.title` and returns `null`.
 - **`publisher_summary` is HTML.** Service strips tags via inline `stripHtml()` helper (regex-based, no cheerio) before populating `description`. Falls back to `merchandising_summary` (plain text) if `publisher_summary` missing.
 - **Series is an array.** `products[].series[]` — a book may belong to multiple series. Service picks the first entry with non-empty `sequence`, else the first entry. `sequence` is cleaned by extracting first `/\d+(?:\.\d+)?/` match for numeric ordering.
 - **Stub `product_images`:** cover URL reads from `product_images['500']`; missing keys fall back to `undefined`.
-- **`page` is 0-indexed.** Despite the default value appearing to be 1, the API returns items `(page * num_results)` through `((page + 1) * num_results - 1)`. So `page=1` fetches items 51–100, not 1–50. All service methods accept a 1-indexed `page` and subtract 1 at the axios call. The symptom of getting this wrong is silent: queries whose `total_results ≤ num_results` return an empty `products` array while `total_results` is populated (e.g. author searches for small catalogues).
+- **`page` is 0-indexed (catalog API only).** Despite the default value appearing to be 1, the API returns items `(page * num_results)` through `((page + 1) * num_results - 1)`. So `page=1` fetches items 51–100, not 1–50. All catalog-API service methods accept a 1-indexed `page` and subtract 1 at the axios call. The symptom of getting this wrong is silent: queries whose `total_results ≤ num_results` return an empty `products` array while `total_results` is populated (e.g. author searches for small catalogues). HTML paths use Audible's native 1-indexed `page` query param and omit it on the first page.
 
 ## Rate Limiting & Resilience
 
-- 503s still possible but dramatically less frequent than the HTML surface.
-- `fetchWithRetry()` — jittered exponential backoff, 5 retries, retries on 503/429/5xx.
-- `AdaptivePacer` circuit-breaker preserved.
-- Inter-page base delay on API paths: **500–1500ms** (down from 2000–4000ms for HTML).
-- API responses include `Cache-Control: private, max-age=1800`.
+- **Real-time JSON API paths:** 503s are uncommon. `fetchWithRetry()` uses jittered exponential backoff, 5 retries, retries on 503/429/5xx. API responses include `Cache-Control: private, max-age=1800`.
+- **Nightly HTML refresh paths:** 503s are more likely (HTML storefront is more rate-sensitive). Same `fetchWithRetry()`, but with `HTML_MAX_RETRIES=12` and `HTML_MAX_BACKOFF_MS=180_000` (3-minute cap on jittered backoff). Healthy refreshes still complete fast (per-page success on attempt 0); users hit by sustained 503 storms grind through patiently rather than abandoning the refresh.
+- **`AdaptivePacer`** — inter-page delay 2–4 s baseline, scales up multiplicatively under retry pressure, with a 45–60 s circuit-breaker cooldown after 3 consecutive retry-pages.
+- **Per-batch cooldowns** in `audible-refresh.processor.ts` — 15–30 s between popular/new-releases, 10–20 s between categories.
 
 ## Region Configuration
 
@@ -101,8 +112,8 @@ Configurable Audible region for accurate metadata matching across international
 - Automatic refresh: Region change triggers `audible_refresh` job.
 
 **Per-region HTTP clients (on init):**
-- `apiClient` — `baseURL=apiBaseUrl`, `Accept: application/json`, `User-Agent: ReadMeABook/1.0`, no language/ipRedirect params.
-- `htmlClient` — `baseURL=baseUrl`, browser headers, default params `ipRedirectOverride=true` + `language=<audibleLocaleParam>`. Used only by `audible-series.ts` and `getBaseUrl()`-based link generation.
+- `apiClient` — `baseURL=apiBaseUrl`, `Accept: application/json`, `User-Agent: ReadMeABook/1.0`, no language/ipRedirect params. Used for the real-time JSON catalog operations (search, author books, categories listing, per-ASIN details fallback).
+- `htmlClient` — `baseURL=baseUrl`, rotating browser headers (`pickUserAgent` + `getBrowserHeaders`), default params `ipRedirectOverride=true` + `language=<audibleLocaleParam>`. Used by the nightly discovery refresh (`/adblbestsellers`, `/newreleases`, `/search?node=...`), by `audible-series.ts`, and by `getBaseUrl()`-based link generation.
 - Audnexus calls include `region=<audnexusParam>`.
 
 **Files:**
@@ -130,6 +141,44 @@ Single matching algorithm used everywhere (search, popular, new-releases, jobs).
 
 **Note:** Fuzzy matching (70% threshold) is preserved in `ranking-algorithm.ts` for Prowlarr torrent ranking. Library availability checks require exact ASIN matches only.
 
+## Dedup & Works Table
+
+**Status:** ✅ Implemented | Two-pass dedup on every discovery view + cross-batch identity via works table
+
+Discovery views (search, author books, series detail) collapse duplicate Audible listings for the same recording (publisher re-listings, regional re-issues, full-cast vs single-narrator productions) into a single card. Two passes run in sequence:
+
+1. **Local pass — `deduplicateAndCollectGroups()`** (`src/lib/utils/deduplicate-audiobooks.ts`)
+   - Stateless, in-memory. Keys books by normalized title + sorted narrator set + duration (±max(5%, 10 min) tolerance), with subtitle compatibility to keep distinct series entries separate.
+   - Picks a canonical representative per group by `metadataScore()` (cover + rating + duration + description + narrator + release date + genres).
+   - Emits `DedupGroup[]` describing every multi-ASIN collapse → handed to `persistDedupGroups()` for the works table.
+
+2. **Works pass — `collapseByExistingWorks()`** (`src/lib/services/works.service.ts`)
+   - Async DB lookup. Reads `work_asins` for every ASIN in the local-passed list and collapses any books sharing a `workId` to one representative (same `metadataScore()` ranking).
+   - Catches duplicates the local pass misses: source-metadata divergence (e.g. HTML scraper captured different narrators), cross-page splits (paginated series), or non-matching field shapes.
+   - Degrades gracefully — returns the input unchanged on DB failure (view still renders).
+
+### Works Table Schema
+- `Work { id, title, author }` — one row per logical book
+- `WorkAsin { id, workId, asin, narrator?, durationMinutes?, isCanonical, source, createdAt }` — many ASINs per Work
+
+### Population Layers
+- **Layer 1 (auto):** `persistDedupGroups()` writes whenever the local pass finds a duplicate. Merges across pre-existing works when a new group spans them.
+- **Layer 2 (seed):** `seedAsin()` writes a single-ASIN work at request creation time, ensuring every requested ASIN has an entry to grow from.
+
+### Read Paths
+- **`collapseByExistingWorks()`** — view-level collapse (this section).
+- **`getSiblingAsins()`** — library availability matching (`audiobook-matcher.ts`), request-creation duplicate prevention (`request-creator.service.ts`), ignored-audiobook expansion. Returns sibling ASINs grouped by input ASIN.
+
+### Narrator Capture in HTML Scrapers
+- HTML scrapers (`audible-series.ts`, the two `parse*Items` parsers in `audible.service.ts`) capture **all** narrator anchors via `extractAllNarrators()` (`src/lib/utils/extract-narrator.ts`). Multi-narrator productions render each name as its own `<a href="?searchNarrator=...">` link; capturing only the first (prior bug) made co-narrated audiobooks fail to dedup. Order is not significant — `normalizeNarrator()` sorts before comparison.
+
+### Wired Routes
+- `src/app/api/audiobooks/search/route.ts`
+- `src/app/api/authors/[asin]/books/route.ts`
+- `src/app/api/series/[asin]/route.ts`
+
+Watched-list background jobs (`watched-lists.service.ts`) run the local pass only — they don't render a view, and the downstream `request-creator.service.ts` already does sibling-aware dedup at request creation time.
+
 ## Database-First Approach
 
 **Status:** Implemented
@@ -137,12 +186,12 @@ Single matching algorithm used everywhere (search, popular, new-releases, jobs).
 Discovery APIs serve cached data from DB with real-time matching.
 
 **Flow:**
-1. `audible_refresh` cron runs daily → fetches 200 popular + 200 new releases + user-configured categories via catalog API.
+1. `audible_refresh` cron runs daily → fetches 200 popular + 200 new releases + user-configured categories by scraping Audible's curated HTML storefronts (`/adblbestsellers`, `/newreleases`, `/search?node=<id>&sort=popularity-rank`).
 2. Downloads and caches cover thumbnails locally.
 3. Stores metadata in `audible_cache`, ranked entries in `audible_cache_categories` with reserved IDs (`__popular__`, `__new_releases__`) and user category IDs.
 4. Cleans up unused thumbnails after sync.
 5. API routes query `AudibleCacheCategory` by categoryId → join with `AudibleCache` metadata → apply real-time matching → return enriched results.
-6. Homepage loads instantly (no Audible API hits).
+6. Homepage loads instantly (no Audible HTTP hits at request time).
 
 ## Thumbnail Caching
 
@@ -228,12 +277,25 @@ interface AuthorBooksResult {
 
 ## Tech Stack
 
-- `axios` (HTTP, two clients: `apiClient` for JSON catalog, `htmlClient` for series-page scraping only)
+- `axios` (HTTP, two clients: `apiClient` for JSON catalog API, `htmlClient` for HTML refresh + series scraping)
+- `cheerio` (HTML parsing for refresh job and `audible-series.ts`)
 - Audnexus API (per-ASIN details, primary)
 - PostgreSQL (`audible_cache`, `audible_cache_categories`)
 
 ## Fixed Issues
 
+**Series-page duplicates not collapsing across user views (2026-05-14)**
+- **Problem:** Two re-listings of the same audiobook (same title, same narrator set, same duration, different ASINs) showed as two cards on series detail pages, even after the works table had already linked them via search-page dedup.
+- **Root cause (two-part):** (1) HTML scrapers used `$el.find('a[href*="searchNarrator="]').first()` for multi-narrator productions, capturing only the first co-narrator. So two listings of the same recording landed in `deduplicateAndCollectGroups` with mismatched single-narrator strings and never merged. (2) `deduplicateAndCollectGroups` was stateless — it wrote to the works table but never read it back, so even when one path (e.g. search) successfully merged two ASINs and persisted the Work, every other path (series, author books) re-derived the dedup decision from scratch and split them again.
+- **Fix:** (1) New `extractAllNarrators()` helper (`src/lib/utils/extract-narrator.ts`) captures every `searchNarrator=` anchor and joins them; all three HTML scrapers route through it. (2) New `collapseByExistingWorks()` consults the works table after the local pass and collapses any remaining books sharing a `workId`. Wired into the three user-facing discovery routes (search / author books / series detail). Skipped for watched-list background jobs — those feed `request-creator.service.ts` which already does sibling-aware dedup.
+- **Location:** `src/lib/utils/extract-narrator.ts` (new); `src/lib/integrations/audible-series.ts` (parseSeriesBooks); `src/lib/integrations/audible.service.ts` (parseProductListItems + parseSearchResultItems); `src/lib/utils/deduplicate-audiobooks.ts` (`metadataScore` exported); `src/lib/services/works.service.ts` (`collapseByExistingWorks` added); three API routes updated.
+
+**Discovery refresh reverted to curated HTML scraping (2026-05-14)**
+- **Problem:** After switching all catalog ops to the JSON catalog API in `f564d0a`, the nightly discovery refresh (Popular / New Releases / user-configured Categories) started serving junk: New Releases became 100% preorders out to 2027, and Popular was dominated by launch-day no-name shovelware.
+- **Root cause:** `products_sort_by=BestSellers` is a right-now sales velocity rank that spikes on launch promos and preorder windows; `-ReleaseDate` returns all catalog items in date order with no released-only filter. The catalog API exposes no server-side filter to exclude preorders or sort by established popularity (verified by exhaustively testing `release_time`, `availability_status`, `customer_rights`, `Reviewed`/`MostListened`/`SalesRank` sorts — all silently ignored or rejected). Doing the curation client-side would have made RMAB the editorial curator, which Audible's storefront pages already do well.
+- **Fix:** Hybrid architecture — the three refresh-only methods (`getPopularAudiobooks`, `getNewReleases`, `getCategoryBooks`) went back to scraping Audible's curated HTML storefronts (`/adblbestsellers`, `/newreleases`, `/search?node=<id>&sort=popularity-rank`). All user-facing real-time paths (search, author books, categories listing, per-ASIN details) stayed on the JSON catalog API. To keep the higher-503-risk HTML traffic resilient on the unattended nightly job, `fetchWithRetry()` accepts an optional `maxBackoffMs` cap and HTML callers use `HTML_MAX_RETRIES=12` + `HTML_MAX_BACKOFF_MS=180_000` (3-min cap). Healthy users finish quickly; 503-blocked users grind through patiently.
+- **Location:** `src/lib/integrations/audible.service.ts` (three methods + two private parsers `parseProductListItems` / `parseSearchResultItems`); `src/lib/utils/scrape-resilience.ts` (`jitteredBackoff` cap parameter).
+
 **Audiobookshelf metadata matching not respecting configured region (2026-01-28)**
 - **Problem:** `triggerABSItemMatch()` hardcoded `'audible'` provider (audible.com) instead of respecting user's configured Audible region.
 - **Impact:** Users with non-US regions (CA, UK, AU, IN) had incorrect metadata matching in Audiobookshelf, causing wrong ASINs.