diff --git a/internal/services/podcast_http.go b/internal/services/podcast_http.go
new file mode 100644
index 00000000..df7d4b89
--- /dev/null
+++ b/internal/services/podcast_http.go
@@ -0,0 +1,190 @@
+// SSRF-guarded HTTP fetching for the podcast fetcher (plan §7).
+//
+// This is the code path that turns USER-SUPPLIED URLs into server
+// fetches — treat it as a security boundary, not a convenience wrapper:
+//
+//   - https-only schemes (feed URL, enclosure, artwork, EVERY redirect hop)
+//   - private/loopback/link-local/multicast/unspecified IP ranges blocked
+//     at DIAL time via Dialer.Control — the check sees the RESOLVED
+//     address, so a DNS name that rebinding-resolves to 127.0.0.1 (or a
+//     redirect into the internal network) is refused at the socket, not
+//     merely at URL-parse time
+//   - hard response size caps on every body (feed XML, artwork, enclosures)
+//
+// Go's redirect follower re-dials through the same dialer per hop, so
+// the IP guard re-applies automatically; the scheme guard re-applies
+// via CheckRedirect.
+package services
+
+import (
+	"context"
+	"crypto/tls"
+	"errors"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"net/url"
+	"syscall"
+	"time"
+)
+
+// Response size caps (bytes). Feeds are XML; artwork is an image;
+// enclosures are audio episodes — the enclosure cap is deliberately the
+// large one (audiobook chapters run long) but still bounded.
+const (
+	maxFeedSizeBytes     = 25 << 20 // 25 MiB
+	maxArtworkSizeBytes  = 20 << 20 // 20 MiB
+	maxEnclosureCapBytes = 1 << 30  // 1 GiB
+)
+
+// ErrBlockedPrivateAddress marks SSRF-guard rejections (dial-time IP or
+// scheme checks). Handlers surface it as a validation failure, and the
+// download worker records it as the download_error text.
+var ErrBlockedPrivateAddress = errors.New("address blocked: private, loopback, or otherwise non-public")
+
+// podcastHTTPClient builds the shared guard client. Timeout is generous
+// (podcast servers are slow) but bounded; enclosures stream through a
+// LimitReader so a lying Content-Length cannot exhaust disk.
+func podcastHTTPClient() *http.Client {
+	dialer := &net.Dialer{
+		Timeout: 15 * time.Second,
+		Control: func(network, address string, _ syscall.RawConn) error {
+			return guardDialAddress(address)
+		},
+	}
+	transport := &http.Transport{
+		DialContext:           dialer.DialContext,
+		TLSHandshakeTimeout:   15 * time.Second,
+		ResponseHeaderTimeout: 60 * time.Second,
+		// Enclosure downloads are long; idle conn reuse still applies.
+		IdleConnTimeout: 90 * time.Second,
+		TLSClientConfig: &tls.Config{MinVersion: tls.VersionTLS12},
+		Proxy:           nil, // never route through a user-controllable proxy
+	}
+	return &http.Client{
+		Transport: transport,
+		Timeout:   15 * time.Minute, // outer bound (biggest enclosures on slow lines)
+		CheckRedirect: func(req *http.Request, via []*http.Request) error {
+			if len(via) >= 10 {
+				return fmt.Errorf("stopped after 10 redirects")
+			}
+			if err := guardURLScheme(req.URL); err != nil {
+				return err
+			}
+			return nil
+		},
+	}
+}
+
+// guardDialAddress validates the RESOLVED host:port at dial time.
+func guardDialAddress(address string) error {
+	host, _, err := net.SplitHostPort(address)
+	if err != nil {
+		return fmt.Errorf("dial guard: %w", err)
+	}
+	ip := net.ParseIP(host)
+	if ip == nil {
+		return fmt.Errorf("dial guard: unparseable address %q", host)
+	}
+	if !isPublicIP(ip) {
+		return ErrBlockedPrivateAddress
+	}
+	return nil
+}
+
+// guardURLScheme enforces https on any URL we are about to touch.
+func guardURLScheme(u *url.URL) error {
+	if u == nil || u.Scheme != "https" {
+		return fmt.Errorf("only https feed/enclosure URLs are supported")
+	}
+	return nil
+}
+
+// isPublicIP reports whether ip is a routable public unicast address.
+// Everything loopback/private/link-local/multicast/unspecified —
+// including IPv4-mapped IPv6 spellings of those — is rejected.
+func isPublicIP(ip net.IP) bool {
+	if ip == nil {
+		return false
+	}
+	// Normalize IPv4-mapped IPv6 (::ffff:10.0.0.1) to its v4 form so
+	// the v4 range checks below actually fire.
+	if ip4 := ip.To4(); ip4 != nil {
+		ip = ip4
+	}
+	if ip.IsLoopback() || ip.IsPrivate() || ip.IsLinkLocalUnicast() ||
+		ip.IsLinkLocalMulticast() || ip.IsInterfaceLocalMulticast() ||
+		ip.IsMulticast() || ip.IsUnspecified() {
+		return false
+	}
+	return true
+}
+
+// guardFetch performs a GET with the scheme/IP guards, conditional-GET
+// validators, and a hard body cap. Returns the body bytes, the FINAL
+// URL after redirects (new-feed-url / migration detection), validators
+// to persist, and whether the server answered 304 Not Modified.
+type fetchResult struct {
+	Body       []byte
+	FinalURL   string
+	ETag       string
+	LastModified string
+	NotModified bool
+	StatusCode int
+}
+
+func guardFetch(ctx context.Context, rawURL, etag, lastModified string, maxSize int64) (*fetchResult, error) {
+	u, err := url.Parse(rawURL)
+	if err != nil {
+		return nil, fmt.Errorf("invalid URL: %w", err)
+	}
+	if err := guardURLScheme(u); err != nil {
+		return nil, err
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, rawURL, nil)
+	if err != nil {
+		return nil, err
+	}
+	// A recognizable UA; some podcast hosts 403 empty/default agents.
+	req.Header.Set("User-Agent", "Bookhoard/1.0 (podcast fetcher; +https://github.com/bookhoard)")
+	req.Header.Set("Accept", "*/*")
+	if etag != "" {
+		req.Header.Set("If-None-Match", etag)
+	}
+	if lastModified != "" {
+		req.Header.Set("If-Modified-Since", lastModified)
+	}
+
+	resp, err := podcastHTTPClient().Do(req)
+	if err != nil {
+		// Unwrap redirect/scheme guard errors so callers can classify.
+		return nil, err
+	}
+	defer func() { _ = resp.Body.Close() }()
+
+	res := &fetchResult{
+		FinalURL:   resp.Request.URL.String(),
+		ETag:       resp.Header.Get("ETag"),
+		LastModified: resp.Header.Get("Last-Modified"),
+		StatusCode: resp.StatusCode,
+	}
+	if resp.StatusCode == http.StatusNotModified {
+		res.NotModified = true
+		return res, nil
+	}
+	if resp.StatusCode != http.StatusOK {
+		return res, fmt.Errorf("fetch %s: HTTP %d", rawURL, resp.StatusCode)
+	}
+
+	body, err := io.ReadAll(io.LimitReader(resp.Body, maxSize+1))
+	if err != nil {
+		return res, fmt.Errorf("fetch %s: %w", rawURL, err)
+	}
+	if int64(len(body)) > maxSize {
+		return res, fmt.Errorf("fetch %s: response exceeds %d byte cap", rawURL, maxSize)
+	}
+	res.Body = body
+	return res, nil
+}
diff --git a/internal/services/podcast_rss.go b/internal/services/podcast_rss.go
new file mode 100644
index 00000000..9f2907d1
--- /dev/null
+++ b/internal/services/podcast_rss.go
@@ -0,0 +1,257 @@
+// RSS 2.0 + iTunes + Podcasting-2.0 feed parsing for the podcast
+// fetcher (plan podcast-integration.md §5/§7). Stdlib encoding/xml only
+// — no new deps.
+//
+// Namespace strategy: iTunes fields MUST be namespace-qualified in the
+// struct tags (the itunes URI) because their local names collide with
+// core RSS ("title", "subtitle", "author", "summary", "explicit",
+// "image", "duration"). Podcasting-2.0 fields (podcast:guid,
+// podcast:transcript, ...) have collision-free local names, so they
+// match on local name — encoding/xml matches Name.Local when the tag
+// carries no namespace, which also tolerates both declared namespace
+// URIs the ecosystem uses.
+//
+// Out of scope at parse time (plan §5): podcast:value, liveItem,
+// soundbite, alternateEnclosure, socialInteract, remoteItem/podroll,
+// locked/block, txt, chat. lastBuildDate/ttl are poll hints only.
+package services
+
+import (
+	"encoding/xml"
+	"fmt"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// itunesNS is the canonical iTunes podcast namespace URI.
+const itunesNS = "http://www.itunes.com/dtds/podcast-1.0.dtd"
+
+// PodcastFeedXML is one parsed RSS document.
+type PodcastFeedXML struct {
+	XMLName xml.Name         `xml:"rss"`
+	Channel PodcastChannelXML `xml:"channel"`
+}
+
+type PodcastChannelXML struct {
+	Title       string `xml:"title"`
+	Link        string `xml:"link"`
+	Language    string `xml:"language"`
+	Description string `xml:"description"`
+
+	// iTunes channel namespace.
+	ITunesAuthor  string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd author"`
+	ITunesOwner   struct {
+		Name  string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd name"`
+		Email string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd email"`
+	} `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd owner"`
+	ITunesImage struct {
+		Href string `xml:"href,attr"`
+	} `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd image"`
+	ITunesExplicit string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd explicit"`
+	ITunesType     string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd type"`
+	ITunesCategory struct {
+		Text       string `xml:"text,attr"`
+		Subcategory struct {
+			Text string `xml:"text,attr"`
+		} `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd category"`
+	} `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd category"`
+	ITunesNewFeedURL string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd new-feed-url"`
+	ITunesComplete   string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd complete"`
+
+	// Podcasting 2.0 (matched on local name; see file comment).
+	PodcastGUID string `xml:"guid"`
+	PodcastMedium string `xml:"medium"`
+
+	LastBuildDate string `xml:"lastBuildDate"`
+
+	Items []PodcastItemXML `xml:"item"`
+}
+
+type PodcastItemXML struct {
+	GUID        string `xml:"guid"`
+	Title       string `xml:"title"`
+	ITunesTitle string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd title"`
+	Subtitle    string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd subtitle"`
+	// Description resolution order (plan §5): description >
+	// itunes:summary > content:encoded.
+	Description  string `xml:"description"`
+	ITunesSummary string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd summary"`
+	ContentEncoded string `xml:"encoded"`
+
+	// Item author: itunes:author preferred, then RSS author, then dc:creator.
+	ITunesAuthor string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd author"`
+	Author       string `xml:"author"`
+	Creator      string `xml:"http://purl.org/dc/elements/1.1/ creator"`
+
+	ITunesImage struct {
+		Href string `xml:"href,attr"`
+	} `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd image"`
+	ITunesExplicit string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd explicit"`
+	ITunesDuration string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd duration"`
+	ITunesEpisode  string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd episode"`
+	ITunesSeason   string `xml:"http://www.itunes.com/dtds/podcast-1.0.dtd season"`
+
+	PubDate string `xml:"pubDate"`
+
+	Enclosure struct {
+		URL    string `xml:"url,attr"`
+		Length int64  `xml:"length,attr"`
+		Type   string `xml:"type,attr"`
+	} `xml:"enclosure"`
+
+	// Podcasting 2.0: store-now, feature-later URLs.
+	Transcript struct {
+		URL string `xml:"url,attr"`
+	} `xml:"transcript"`
+	Chapters struct {
+		URL string `xml:"url,attr"`
+	} `xml:"chapters"`
+	Persons []struct {
+		Href string `xml:"href,attr"`
+		Role string `xml:"role,attr"`
+	} `xml:"person"`
+}
+
+// ParsePodcastFeed decodes a feed document. Returns an error when the
+// XML is unparseable or the document is not an RSS channel.
+func ParsePodcastFeed(data []byte) (*PodcastFeedXML, error) {
+	var feed PodcastFeedXML
+	if err := xml.Unmarshal(data, &feed); err != nil {
+		return nil, fmt.Errorf("feed parse: %w", err)
+	}
+	if feed.Channel.Title == "" && len(feed.Channel.Items) == 0 {
+		return nil, fmt.Errorf("feed parse: not an RSS channel (no title, no items)")
+	}
+	return &feed, nil
+}
+
+// EpisodeTitle prefers itunes:title over the raw RSS title (plan §5).
+func (it *PodcastItemXML) EpisodeTitle() string {
+	if t := strings.TrimSpace(it.ITunesTitle); t != "" {
+		return t
+	}
+	return strings.TrimSpace(it.Title)
+}
+
+// EpisodeDescription resolves the description fallback chain.
+func (it *PodcastItemXML) EpisodeDescription() string {
+	for _, c := range []string{it.Description, it.ITunesSummary, it.ContentEncoded} {
+		if t := strings.TrimSpace(c); t != "" {
+			return t
+		}
+	}
+	return ""
+}
+
+// EpisodeAuthor resolves the author fallback chain.
+func (it *PodcastItemXML) EpisodeAuthor() string {
+	for _, c := range []string{it.ITunesAuthor, it.Author, it.Creator} {
+		if t := strings.TrimSpace(c); t != "" {
+			return t
+		}
+	}
+	return ""
+}
+
+// EpisodeExplicit maps the itunes:explicit spelling variants to bool.
+func (it *PodcastItemXML) EpisodeExplicit() bool {
+	return parseExplicit(it.ITunesExplicit)
+}
+
+func parseExplicit(v string) bool {
+	switch strings.ToLower(strings.TrimSpace(v)) {
+	case "true", "yes", "explicit", "1":
+		return true
+	}
+	return false
+}
+
+// HostsAndGuests splits podcast:person entries by role (plan §5).
+func (it *PodcastItemXML) HostsAndGuests() (hosts []string, guests []string) {
+	for _, p := range it.Persons {
+		name := p.Href
+		if name == "" {
+			continue
+		}
+		switch strings.ToLower(p.Role) {
+		case "guest":
+			guests = append(guests, name)
+		default: // host and unlabeled roles surface as hosts
+			hosts = append(hosts, name)
+		}
+	}
+	return hosts, guests
+}
+
+// parseITunesDuration accepts the three spellings the spec allows —
+// plain seconds, HH:MM:SS, MM:SS — plus a best-effort prose fallback
+// ("58 minutes"). Returns 0 for anything unparseable (the column stays
+// NULL; ffprobe reconciles after download anyway).
+func parseITunesDuration(v string) int {
+	v = strings.TrimSpace(v)
+	if v == "" {
+		return 0
+	}
+	if secs, err := strconv.Atoi(v); err == nil {
+		return secs
+	}
+	parts := strings.Split(v, ":")
+	if len(parts) <= 3 {
+		total := 0
+		ok := true
+		for _, p := range parts {
+			n, err := strconv.Atoi(strings.TrimSpace(p))
+			if err != nil {
+				ok = false
+				break
+			}
+			total = total*60 + n
+		}
+		if ok {
+			return total
+		}
+	}
+	// Prose best-effort: "<N> minutes/hours/seconds".
+	lower := strings.ToLower(v)
+	fields := strings.Fields(lower)
+	for i, f := range fields {
+		if n, err := strconv.Atoi(f); err == nil && i+1 < len(fields) {
+			switch {
+			case strings.HasPrefix(fields[i+1], "hour"):
+				return n * 3600
+			case strings.HasPrefix(fields[i+1], "minute"):
+				return n * 60
+			case strings.HasPrefix(fields[i+1], "second"):
+				return n
+			}
+		}
+	}
+	return 0
+}
+
+// parseFeedTime parses the RFC-822/1123 variants feeds actually emit.
+func parseFeedTime(v string) (t time.Time, ok bool) {
+	v = strings.TrimSpace(v)
+	if v == "" {
+		return time.Time{}, false
+	}
+	layouts := []string{
+		time.RFC1123Z,
+		time.RFC1123,
+		time.RFC822Z,
+		time.RFC822,
+		"Mon, 2 Jan 2006 15:04:05 -0700",
+		"Mon, 2 Jan 2006 15:04:05 MST",
+		"2 Jan 2006 15:04:05 -0700",
+		"2 Jan 2006 15:04:05 MST",
+		time.ANSIC,
+		time.UnixDate,
+	}
+	for _, layout := range layouts {
+		if t, err := time.Parse(layout, v); err == nil {
+			return t, true
+		}
+	}
+	return time.Time{}, false
+}
diff --git a/internal/services/podcast_service.go b/internal/services/podcast_service.go
new file mode 100644
index 00000000..d55086bf
--- /dev/null
+++ b/internal/services/podcast_service.go
@@ -0,0 +1,762 @@
+// Podcast fetcher (plan podcast-integration.md §7): fetch → parse →
+// upsert-without-delete. The fetcher is the SOLE writer of feed/episode
+// metadata (D9) — user-facing podcast writes are limited to
+// podcast_subscriptions rows and the reader_settings order override.
+//
+// RLS identity (invariant 12): fetches run detached from any request,
+// so every governed statement runs under the FIRST-ADMIN identity
+// (scanner fallback precedent). With the D20 admin escape that identity
+// sees and updates every feed/episode row. NEVER a bypass flag, never
+// the bootstrap superuser.
+//
+// Upsert semantics: episodes are created/updated METADATA-ONLY, keyed
+// on (source_feed_id, episode_guid); rows for items that vanish from
+// the XML are NEVER deleted (history accumulates — D14 search covers
+// gone episodes). Download state is owned by the download worker; the
+// upsert cannot clobber it (guarded in the query itself).
+package services
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"log"
+	"net/url"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+
+	"bookhoard/internal/database"
+	"bookhoard/internal/rls"
+
+	"github.com/jackc/pgx/v5/pgtype"
+)
+
+// PodcastError is a status-carrying service error: handlers map it to
+// the HTTP response directly (services stay echo-free).
+type PodcastError struct {
+	Status  int
+	Message string
+}
+
+func (e *PodcastError) Error() string { return e.Message }
+
+func errNotFound(msg string) error    { return &PodcastError{Status: 404, Message: msg} }
+func errForbidden(msg string) error   { return &PodcastError{Status: 403, Message: msg} }
+func errBadRequest(msg string) error  { return &PodcastError{Status: 400, Message: msg} }
+
+// PodcastService carries the fetcher/worker dependencies. One instance
+// per server (handlers + poller share it so the per-feed locks dedupe).
+type PodcastService struct {
+	db *database.Queries
+
+	adminMu   sync.Mutex
+	adminID   pgtype.UUID
+	adminOnce bool
+
+	// Per-feed mutexes: a poll and a download for the same feed must
+	// not race (e.g. both writing the episode row). Keyed by feed UUID.
+	feedLocksMu sync.Mutex
+	feedLocks   map[string]*sync.Mutex
+}
+
+func NewPodcastService(db *database.Queries) *PodcastService {
+	return &PodcastService{
+		db:        db,
+		feedLocks: make(map[string]*sync.Mutex),
+	}
+}
+
+// feedLock returns the per-feed mutex (creates on first use).
+func (s *PodcastService) feedLock(feedID string) *sync.Mutex {
+	s.feedLocksMu.Lock()
+	defer s.feedLocksMu.Unlock()
+	if m, ok := s.feedLocks[feedID]; ok {
+		return m
+	}
+	m := &sync.Mutex{}
+	s.feedLocks[feedID] = m
+	return m
+}
+
+// adminIdentity resolves the acting identity for detached podcast work:
+// the first admin, resolved once per service lifetime. Errors surface
+// to the caller (no admin → no fetch — same contract as the scanner).
+func (s *PodcastService) adminIdentity() (pgtype.UUID, error) {
+	s.adminMu.Lock()
+	defer s.adminMu.Unlock()
+	if !s.adminOnce && !s.adminID.Valid {
+		id, err := s.db.GetFirstAdmin(context.Background())
+		if err != nil {
+			return pgtype.UUID{}, fmt.Errorf("podcasts: no admin identity available for detached work: %w", err)
+		}
+		s.adminID = id
+		s.adminOnce = true
+	}
+	if !s.adminID.Valid {
+		return pgtype.UUID{}, fmt.Errorf("podcasts: no admin identity available for detached work")
+	}
+	return s.adminID, nil
+}
+
+// adminCtx wraps ctx with the first-admin identity (invariant 12).
+func (s *PodcastService) adminCtx(ctx context.Context) (context.Context, error) {
+	id, err := s.adminIdentity()
+	if err != nil {
+		return nil, err
+	}
+	return rls.WithUser(ctx, id), nil
+}
+
+// AddFeedResult carries what the add-feed flow did (the handler turns
+// it into the response / serial feed hint).
+type AddFeedResult struct {
+	Feed *database.PodcastFeeds
+	// Created is true when this call created the feed row; false when
+	// an existing row was reused (URL dedupe or guid migration match).
+	Created bool
+	// Subscribed is true when the requesting user's subscription was
+	// created by this call (false = already subscribed).
+	Subscribed bool
+	// EpisodeCount is the number of episode rows the fetch upserted.
+	EpisodeCount int
+	// FetchError, when set, means the feed row exists but the fetch
+	// failed — the subscription stands, the error surfaces to the user.
+	FetchError string
+}
+
+// AddFeed implements POST /api/podcasts/feeds (plan §8): fetch+parse the
+// user-supplied URL, create-or-reuse the feed row, upsert metadata-only
+// episodes, subscribe the creator. Authorization on explicit arguments
+// (D20 scoping model): the target library must be visible to the actor
+// and, for non-admins, allow member feeds.
+func (s *PodcastService) AddFeed(ctx context.Context, userID pgtype.UUID, isAdmin bool, libraryID pgtype.UUID, rawFeedURL string) (*AddFeedResult, error) {
+	// ---- explicit-argument authorization (the ONLY Go-level gate) ----
+	lib, err := s.db.GetLibrary(ctx, libraryID)
+	if err != nil {
+		// Invisible-or-nonexistent: 404 semantics (existence leak).
+		return nil, errNotFound("library not found")
+	}
+	if !isAdmin {
+		if lib.TypeName != LibraryTypePodcasts {
+			return nil, errNotFound("library not found")
+		}
+		if !lib.AllowMemberFeeds {
+			return nil, errForbidden("this library does not allow members to add feeds")
+		}
+	}
+
+	feedURL := strings.TrimSpace(rawFeedURL)
+	if feedURL == "" {
+		return nil, errBadRequest("feed URL is required")
+	}
+
+	// The fetch/upsert runs under the first-admin identity (detached-
+	// equivalent); the request context never carries special powers into
+	// feed writes.
+	actx, err := s.adminCtx(ctx)
+	if err != nil {
+		return nil, err
+	}
+
+	res := &AddFeedResult{}
+
+	// ---- fetch + parse (SSRF-guarded; plan §7) ----
+	fetched, err := guardFetch(ctx, feedURL, "", "", maxFeedSizeBytes)
+	if err != nil {
+		return nil, fmt.Errorf("could not fetch feed: %w", classifyFetchError(err))
+	}
+	parsed, err := ParsePodcastFeed(fetched.Body)
+	if err != nil {
+		return nil, fmt.Errorf("could not parse feed: %w", err)
+	}
+
+	// Follow itunes:new-feed-url once at add time so the row is born
+	// pointing at the publisher's current home (migration semantics).
+	finalURL := fetched.FinalURL
+	if neu := strings.TrimSpace(parsed.Channel.ITunesNewFeedURL); neu != "" && neu != finalURL {
+		if retried, err := guardFetch(ctx, neu, "", "", maxFeedSizeBytes); err == nil {
+			if reparsed, err := ParsePodcastFeed(retried.Body); err == nil {
+				parsed = reparsed
+				finalURL = retried.FinalURL
+			}
+		}
+	}
+
+	title := parsed.Channel.Title
+	if title == "" {
+		title = "Untitled feed"
+	}
+
+	// ---- create-or-reuse (UNIQUE(library_id, feed_url) + guid dedupe) ----
+	feed, err := s.db.CreatePodcastFeed(actx, database.CreatePodcastFeedParams{
+		LibraryID:        libraryID,
+		FeedUrl:          finalURL,
+		Title:            truncate(title, 500),
+		CreatedByAdminID: pgtype.UUID{}, // fetcher owns feed metadata; creator attribution is not user metadata
+	})
+	if err != nil {
+		return nil, fmt.Errorf("could not create feed: %w", err)
+	}
+	res.Feed = &feed
+	res.Created = true
+
+	// Guid migration dedupe: a row with the same canonical podcast_guid
+	// under a different URL reuses THAT row (do not fork the feed).
+	if guid := strings.TrimSpace(parsed.Channel.PodcastGUID); guid != "" {
+		if feed.PodcastGuid.String == "" || feed.PodcastGuid.String != guid {
+			if existing, err := s.db.GetPodcastFeedByGuid(actx, database.GetPodcastFeedByGuidParams{
+				LibraryID:   feed.LibraryID,
+				PodcastGuid: pgtypeText(guid),
+			}); err == nil && existing.ID != feed.ID {
+				// Attach the subscriber to the canonical row; drop the
+				// freshly created stub (nothing subscribed to it yet).
+				_, _ = s.db.DeletePodcastFeed(actx, feed.ID)
+				feed = existing
+				res.Feed = &existing
+				// Re-point the canonical row at the fetched URL if it moved.
+				if feedUrlForComparison(existing.FeedUrl) != feedUrlForComparison(finalURL) {
+					_ = s.db.UpdatePodcastFeedRedirect(actx, database.UpdatePodcastFeedRedirectParams{
+						ID:      existing.ID,
+						FeedUrl: finalURL,
+					})
+				}
+			}
+		}
+	}
+
+	if _, err := s.db.SubscribePodcastFeed(actx, database.SubscribePodcastFeedParams{
+		UserID: userID,
+		FeedID: feed.ID,
+	}); err == nil {
+		res.Subscribed = true
+	}
+
+	// ---- metadata + episodes ----
+	if err := s.applyFetch(actx, &feed, parsed, fetched); err != nil {
+		// Feed exists + subscribed, but the metadata pass failed: not
+		// fatal for the request; the poller reconciles later.
+		res.FetchError = err.Error()
+		return res, nil
+	}
+
+	updated, err := s.db.GetPodcastFeed(actx, feed.ID)
+	if err == nil {
+		res.Feed = &updated
+	}
+	return res, nil
+}
+
+// RefreshFeed implements POST /api/podcasts/feeds/:id/refresh: a manual
+// poll. The per-feed lock serializes against poller ticks and downloads.
+func (s *PodcastService) RefreshFeed(ctx context.Context, feedID pgtype.UUID) error {
+	actx, err := s.adminCtx(ctx)
+	if err != nil {
+		return err
+	}
+	feed, err := s.db.GetPodcastFeed(actx, feedID)
+	if err != nil {
+		return errNotFound("feed not found")
+	}
+	return s.PollFeed(actx, &feed)
+}
+
+// PollFeed fetches + reconciles ONE feed (caller supplies the acting
+// identity in ctx). Safe to call concurrently per feed — the per-feed
+// lock serializes poll/download.
+func (s *PodcastService) PollFeed(ctx context.Context, feed *database.PodcastFeeds) error {
+	lock := s.feedLock(feed.ID.String())
+	lock.Lock()
+	defer lock.Unlock()
+
+	// Complete feeds back off: they are fetched only on explicit
+	// refresh (the poller skips them entirely — this path is manual).
+	fetched, err := guardFetch(ctx, feed.FeedUrl, feed.Etag.String, feed.LastModified.String, maxFeedSizeBytes)
+	if err != nil {
+		if feed.LastFetchedAt.Valid {
+			// A previously-good feed with a transient failure: keep the
+			// old state; the poller retries next tick.
+			return fmt.Errorf("fetch %s: %w", feed.Title, classifyFetchError(err))
+		}
+		return fmt.Errorf("fetch %s: %w", feed.Title, classifyFetchError(err))
+	}
+	if fetched.NotModified {
+		_, _ = s.db.UpdatePodcastFeedFetch(ctx, database.UpdatePodcastFeedFetchParams{
+			ID:            feed.ID,
+			Etag:          pgtypeText(feed.Etag.String),
+			LastModified:  pgtypeText(feed.LastModified.String),
+			LastBuildDate: feed.LastBuildDate,
+		})
+		return nil
+	}
+
+	parsed, err := ParsePodcastFeed(fetched.Body)
+	if err != nil {
+		return fmt.Errorf("parse %s: %w", feed.Title, err)
+	}
+
+	// itunes:new-feed-url: the publisher MOVED the feed. Rewrite the
+	// stored URL (row, subscriptions and episodes stay put) and refetch
+	// the new location once so this poll applies its content.
+	if neu := strings.TrimSpace(parsed.Channel.ITunesNewFeedURL); neu != "" && neu != feed.FeedUrl {
+		_ = s.db.UpdatePodcastFeedRedirect(ctx, database.UpdatePodcastFeedRedirectParams{ID: feed.ID, FeedUrl: neu})
+		if retried, err := guardFetch(ctx, neu, "", "", maxFeedSizeBytes); err == nil {
+			if reparsed, err := ParsePodcastFeed(retried.Body); err == nil {
+				parsed = reparsed
+				fetched = retried
+			}
+		}
+	} else if fetched.FinalURL != "" && feedUrlForComparison(fetched.FinalURL) != feedUrlForComparison(feed.FeedUrl) {
+		// A permanent HTTP redirect chain moved the feed; persist the
+		// final URL so conditional validators stop bouncing.
+		_ = s.db.UpdatePodcastFeedRedirect(ctx, database.UpdatePodcastFeedRedirectParams{ID: feed.ID, FeedUrl: fetched.FinalURL})
+	}
+
+	if err := s.applyFetch(ctx, feed, parsed, fetched); err != nil {
+		return err
+	}
+	return nil
+}
+
+// applyFetch writes channel metadata, upserts episodes metadata-only,
+// touches validators, and downloads artwork. Write-after-success rule
+// (invariant 7): failures here never clear existing rows.
+func (s *PodcastService) applyFetch(ctx context.Context, feed *database.PodcastFeeds, parsed *PodcastFeedXML, fetched *fetchResult) error {
+	// --- channel metadata (fetcher-owned) ---
+	title := parsed.Channel.Title
+	if title == "" {
+		title = feed.Title // never blank out a title on a broken doc
+	}
+	desc := stripHTML(parsed.Channel.Description)
+
+	host := parsed.Channel.ITunesAuthor
+	publisher := parsed.Channel.ITunesOwner.Name
+
+	tagSet := map[string]bool{}
+	if cat := strings.TrimSpace(parsed.Channel.ITunesCategory.Text); cat != "" {
+		tagSet[cat] = true
+	}
+	if sub := strings.TrimSpace(parsed.Channel.ITunesCategory.Subcategory.Text); sub != "" {
+		tagSet[sub] = true
+	}
+	tags := sortedStringSet(tagSet)
+
+	feedType := strings.TrimSpace(parsed.Channel.ITunesType)
+	if feedType != "serial" && feedType != "episodic" {
+		feedType = "" // unknown → column stays NULL
+	}
+	medium := strings.TrimSpace(parsed.Channel.PodcastMedium)
+
+	if _, err := s.db.UpdatePodcastFeedMetadata(ctx, database.UpdatePodcastFeedMetadataParams{
+		ID:          feed.ID,
+		Title:       truncate(title, 500),
+		Description: pgtypeText(desc),
+		Host:        pgtypeText(truncate(host, 500)),
+		Publisher:   pgtypeText(truncate(publisher, 500)),
+		OwnerName:   pgtypeText(truncate(parsed.Channel.ITunesOwner.Name, 500)),
+		OwnerEmail:  pgtypeText(truncate(parsed.Channel.ITunesOwner.Email, 500)),
+		Language:    pgtypeText(truncate(parsed.Channel.Language, 50)),
+		Tags:        tags,
+		Explicit:    parseExplicit(parsed.Channel.ITunesExplicit),
+		FeedType:    pgtypeText(feedType),
+		Medium:      pgtypeText(truncate(medium, 20)),
+		IsComplete:  parseExplicit(parsed.Channel.ITunesComplete),
+		PodcastGuid: pgtypeText(strings.TrimSpace(parsed.Channel.PodcastGUID)),
+	}); err != nil {
+		return fmt.Errorf("feed metadata update: %w", err)
+	}
+	// itunes:complete flips the poller's back-off flag (complete feeds
+	// poll rarely); explicit refresh still works.
+
+	// --- episodes (metadata-only upsert, positions from THIS parse) ---
+	root, err := s.downloadRootAbs(ctx, feed)
+	if err != nil {
+		return fmt.Errorf("download root: %w", err)
+	}
+	feedSlug := feedSlug(title, feed.ID)
+
+	count := 0
+	for i := range parsed.Channel.Items {
+		if err := s.upsertEpisode(ctx, feed, root, feedSlug, &parsed.Channel.Items[i], i+1); err != nil {
+			// One bad item must not sink the batch.
+			log.Printf("[PODCAST] episode upsert failed for feed %s item %d: %v", feed.ID, i+1, err)
+			continue
+		}
+		count++
+	}
+
+	// --- validators + fetch timestamp ---
+	var lastBuildDate pgtype.Timestamptz
+	if t, ok := parseFeedTime(parsed.Channel.LastBuildDate); ok {
+		lastBuildDate = pgtypeTimestamptz(t)
+	}
+	if _, err := s.db.UpdatePodcastFeedFetch(ctx, database.UpdatePodcastFeedFetchParams{
+		ID:            feed.ID,
+		Etag:          pgtypeText(fetched.ETag),
+		LastModified:  pgtypeText(fetched.LastModified),
+		LastBuildDate: lastBuildDate,
+	}); err != nil {
+		return fmt.Errorf("fetch touch: %w", err)
+	}
+
+	// --- artwork (best-effort; feed renders fine without it) ---
+	s.refreshArtwork(ctx, feed, root, parsed)
+
+	return nil
+}
+
+// upsertEpisode maps one XML item onto a media_items row (plan §5 field
+// inventory). Positions are 1-based, renumbered per fetch; file_path is
+// the PLANNED final layout — the upsert only applies it to rows that
+// are not yet downloaded (query-side guard).
+func (s *PodcastService) upsertEpisode(ctx context.Context, feed *database.PodcastFeeds, downloadRoot, feedSlug string, item *PodcastItemXML, position int) error {
+	title := item.EpisodeTitle()
+	if title == "" {
+		title = fmt.Sprintf("Episode %d", position)
+	}
+	guid := strings.TrimSpace(item.GUID)
+	if guid == "" {
+		// GUID-less feeds exist; fall back to a stable identity from the
+		// enclosure URL (the next-best immutable per-episode key).
+		if item.Enclosure.URL != "" {
+			guid = "enc:" + item.Enclosure.URL
+		} else {
+			return fmt.Errorf("item %q has neither guid nor enclosure; skipping", title)
+		}
+	}
+
+	ext := enclosureExtension(item.Enclosure.URL, item.Enclosure.Type)
+	relPath := episodeRelPath(feedSlug, position, title, ext)
+
+	var published pgtype.Date
+	if t, ok := parseFeedTime(item.PubDate); ok {
+		published = pgtypeDate(t)
+	}
+
+	authors := []string{}
+	if a := item.EpisodeAuthor(); a != "" {
+		authors = append(authors, truncate(a, 255))
+	}
+	hosts, guests := item.HostsAndGuests()
+
+	_, err := s.db.UpsertPodcastEpisode(ctx, database.UpsertPodcastEpisodeParams{
+		LibraryID:     feed.LibraryID,
+		Title:         truncate(title, 255),
+		Subtitle:      pgtypeText(truncate(item.Subtitle, 500)),
+		Description:   pgtypeText(stripHTML(item.EpisodeDescription())),
+		Authors:       authors,
+		AuthorsSearch: lowerSearchSlice(authors),
+		FilePath:      relPath,
+		FileSize:      pgtypeInt8(item.Enclosure.Length),
+		MimeType:      pgtypeText(truncate(item.Enclosure.Type, 100)),
+		CoverImagePath: pgtypeText(artworkRelPath(feedSlug, artworkExtFromURL(item.ITunesImage.Href))),
+		Explicit:      item.EpisodeExplicit(),
+		DatePublished: published,
+		Language:      pgtypeText(truncate(feed.Language.String, 10)),
+		Tags:          feed.Tags,
+		Genre:         pgtypeText(strings.Join(feed.Tags, ", ")),
+		SourceFeedID:  feed.ID,
+		EpisodeGuid:   pgtypeText(truncate(guid, 500)),
+		FeedPosition:  pgtypeInt4(int32(position)),
+		SeasonNumber:  pgtypeInt4FromStr(item.ITunesSeason),
+		EpisodeNumber: pgtypeInt4FromStr(item.ITunesEpisode),
+		EnclosureUrl:  pgtypeText(item.Enclosure.URL),
+		DeclaredDurationSeconds: pgtypeInt4(int32(parseITunesDuration(item.ITunesDuration))),
+		TranscriptUrl: pgtypeText(item.Transcript.URL),
+		ChaptersUrl:   pgtypeText(item.Chapters.URL),
+		Hosts:         truncateSlice(hosts, 255),
+		Guests:        truncateSlice(guests, 255),
+	})
+	_ = downloadRoot // layout is relative; abs root resolves at download time
+	return err
+}
+
+// refreshArtwork downloads the channel image once per changed URL.
+// Artwork is feed-level: episodes point at the same relative path.
+func (s *PodcastService) refreshArtwork(ctx context.Context, feed *database.PodcastFeeds, downloadRoot string, parsed *PodcastFeedXML) {
+	imgURL := strings.TrimSpace(parsed.Channel.ITunesImage.Href)
+	if imgURL == "" {
+		return
+	}
+	target := filepath.Join(downloadRoot, artworkRelPath(feedSlug(feed.Title, feed.ID), artworkExtFromURL(imgURL)))
+	if feed.ArtworkPath.Valid {
+		if existing := filepath.Join(downloadRoot, feed.ArtworkPath.String); existing == target {
+			if _, err := os.Stat(existing); err == nil {
+				return // unchanged
+			}
+		}
+	}
+	if err := os.MkdirAll(filepath.Dir(target), 0o755); err != nil {
+		return
+	}
+	res, err := guardFetch(ctx, imgURL, "", "", maxArtworkSizeBytes)
+	if err != nil || len(res.Body) == 0 {
+		log.Printf("[PODCAST] artwork fetch failed for feed %s: %v", feed.ID, err)
+		return
+	}
+	if err := os.WriteFile(target, res.Body, 0o644); err != nil {
+		log.Printf("[PODCAST] artwork write failed for feed %s: %v", feed.ID, err)
+		return
+	}
+	rel, err := filepath.Rel(downloadRoot, target)
+	if err != nil {
+		return
+	}
+	_ = s.db.UpdatePodcastFeedArtwork(ctx, database.UpdatePodcastFeedArtworkParams{
+		ID:          feed.ID,
+		ArtworkPath: pgtypeText(rel),
+	})
+}
+
+// DeleteFeedFiles removes a feed's episode files + artwork from the
+// download root (admin feed-removal path; ?files=false skips it).
+// Best-effort per file — the row cascade proceeds regardless.
+func (s *PodcastService) DeleteFeedFiles(ctx context.Context, feed *database.PodcastFeeds, episodes []database.MediaItems) {
+	root, err := s.downloadRootAbs(ctx, feed)
+	if err != nil {
+		return
+	}
+	for _, ep := range episodes {
+		if ep.FilePath != "" {
+			_ = os.Remove(filepath.Join(root, ep.FilePath))
+			// Embedded-cover sibling (extractEmbeddedAudioCover convention).
+			base := filepath.Join(root, strings.TrimSuffix(ep.FilePath, filepath.Ext(ep.FilePath)))
+			_ = os.Remove(base + ".cover.jpg")
+		}
+	}
+	if feed.ArtworkPath.Valid {
+		_ = os.Remove(filepath.Join(root, feed.ArtworkPath.String))
+	}
+	// The feed directory itself: remove when empty (other feeds never
+	// share a slug — the feed-id suffix guarantees that).
+	_ = os.Remove(filepath.Join(root, feedSlug(feed.Title, feed.ID)))
+}
+
+// downloadRootAbs resolves the podcasts library's single download root
+// (purpose='podcast') as an absolute filesystem path.
+func (s *PodcastService) downloadRootAbs(ctx context.Context, feed *database.PodcastFeeds) (string, error) {
+	folders, err := s.db.GetPodcastLibraryFolder(ctx, feed.LibraryID)
+	if err != nil || len(folders) == 0 {
+		return "", fmt.Errorf("library %s has no download root registered", feed.LibraryID)
+	}
+	return folders[0].FolderPath, nil
+}
+
+// episodeRelPath is the pinned episode layout (plan §6):
+// <feed-slug>/<feed_position:04d>-<slug>.<ext> — relative to the
+// download root. getFullFilePath/ServeFile/Range streaming resolve it
+// unchanged.
+func episodeRelPath(feedSlug string, position int, title, ext string) string {
+	return fmt.Sprintf("%s/%04d-%s%s", feedSlug, position, slugify(title, 80), ext)
+}
+
+// artworkRelPath is the feed-level artwork location (feed-level, so
+// every episode row can point at the same file).
+func artworkRelPath(feedSlug, ext string) string {
+	return fmt.Sprintf("%s/artwork%s", feedSlug, ext)
+}
+
+// feedSlug derives the stable per-feed directory name: slugged title +
+// short feed-id suffix so two identically-titled feeds never collide.
+func feedSlug(title string, feedID pgtype.UUID) string {
+	base := slugify(title, 60)
+	suffix := "feed"
+	if feedID.Valid {
+		s := feedID.String()
+		if len(s) >= 8 {
+			suffix = s[:8]
+		}
+	}
+	if base == "" {
+		base = "feed"
+	}
+	return base + "-" + suffix
+}
+
+var slugStrip = regexp.MustCompile(`[^a-z0-9]+`)
+
+// slugify lowercases, strips non-alphanumerics to dashes, and clamps
+// the length (file-name-safe; unicode letters drop — file layout is
+// an internal key, never display text).
+func slugify(s string, maxLen int) string {
+	s = strings.ToLower(strings.TrimSpace(s))
+	s = slugStrip.ReplaceAllString(s, "-")
+	s = strings.Trim(s, "-")
+	if len(s) > maxLen {
+		s = strings.TrimRight(s[:maxLen], "-")
+	}
+	return s
+}
+
+// enclosureExtension picks a file extension from the URL, then the
+// MIME type, then a safe default.
+func enclosureExtension(rawURL, mimeType string) string {
+	if u, err := urlParse(rawURL); err == nil {
+		if ext := strings.ToLower(filepath.Ext(u.Path)); ext == ".mp3" || ext == ".m4a" || ext == ".m4b" ||
+			ext == ".flac" || ext == ".ogg" || ext == ".oga" || ext == ".opus" || ext == ".wav" || ext == ".aac" {
+			return ext
+		}
+	}
+	switch {
+	case strings.Contains(mimeType, "mp4") || strings.Contains(mimeType, "m4a"):
+		return ".m4a"
+	case strings.Contains(mimeType, "mpeg") || strings.Contains(mimeType, "mp3"):
+		return ".mp3"
+	case strings.Contains(mimeType, "flac"):
+		return ".flac"
+	case strings.Contains(mimeType, "ogg"):
+		return ".ogg"
+	case strings.Contains(mimeType, "wav"):
+		return ".wav"
+	}
+	return ".mp3"
+}
+
+// artworkExtFromURL keeps the source extension for artwork (jpg/png/webp).
+func artworkExtFromURL(rawURL string) string {
+	if u, err := urlParse(rawURL); err == nil {
+		switch strings.ToLower(filepath.Ext(u.Path)) {
+		case ".png":
+			return ".png"
+		case ".webp":
+			return ".webp"
+		case ".gif":
+			return ".gif"
+		}
+	}
+	return ".jpg"
+}
+
+// stripHTML removes tags and collapses whitespace — feed descriptions
+// are HTML; storage wants text (the TUI-ish surfaces render it plain).
+func stripHTML(s string) string {
+	if s == "" {
+		return ""
+	}
+	var b strings.Builder
+	inTag := false
+	for _, r := range s {
+		switch {
+		case r == '<':
+			inTag = true
+		case r == '>':
+			inTag = false
+		case !inTag:
+			b.WriteRune(r)
+		}
+	}
+	out := b.String()
+	// Collapse runs of whitespace.
+	out = strings.Join(strings.Fields(out), " ")
+	return strings.TrimSpace(out)
+}
+
+// truncate clamps a string to max runes (VARCHAR safety at the DB edge).
+func truncate(s string, max int) string {
+	s = strings.TrimSpace(s)
+	runes := []rune(s)
+	if len(runes) <= max {
+		return s
+	}
+	return string(runes[:max])
+}
+
+// truncateSlice clamps each element and drops empties.
+func truncateSlice(in []string, max int) []string {
+	out := make([]string, 0, len(in))
+	for _, s := range in {
+		if t := truncate(s, max); t != "" {
+			out = append(out, t)
+		}
+	}
+	return out
+}
+
+// lowerSearchSlice mirrors authors into authors_search (scanner convention).
+func lowerSearchSlice(in []string) []string {
+	out := make([]string, 0, len(in))
+	for _, s := range in {
+		out = append(out, strings.ToLower(s))
+	}
+	return out
+}
+
+// sortedStringSet → deterministic TEXT[] writes.
+func sortedStringSet(set map[string]bool) []string {
+	out := make([]string, 0, len(set))
+	for k := range set {
+		if strings.TrimSpace(k) == "" {
+			continue
+		}
+		out = append(out, strings.TrimSpace(k))
+	}
+	// tiny insertion sort — tag sets are 1–5 entries
+	for i := 1; i < len(out); i++ {
+		for j := i; j > 0 && out[j] < out[j-1]; j-- {
+			out[j], out[j-1] = out[j-1], out[j]
+		}
+	}
+	return out
+}
+
+// classifyFetchError maps guard/HTTP errors to user-facing text (the
+// add-feed dialog and the download_error column both render it).
+func classifyFetchError(err error) error {
+	if errors.Is(err, ErrBlockedPrivateAddress) {
+		return fmt.Errorf("blocked: private/loopback addresses are not fetchable")
+	}
+	return err
+}
+
+// urlParse local helper so podcast files don't each import net/url.
+func urlParse(raw string) (*url.URL, error) {
+	return url.Parse(raw)
+}
+
+// ---- pgtype helpers shared by the podcast files ----
+
+func pgtypeText(s string) pgtype.Text {
+	if strings.TrimSpace(s) == "" {
+		return pgtype.Text{}
+	}
+	return pgtype.Text{String: s, Valid: true}
+}
+
+func pgtypeInt4(v int32) pgtype.Int4 {
+	if v == 0 {
+		return pgtype.Int4{}
+	}
+	return pgtype.Int4{Int32: v, Valid: true}
+}
+
+func pgtypeInt4FromStr(s string) pgtype.Int4 {
+	n, err := strconv.Atoi(strings.TrimSpace(s))
+	if err != nil || n <= 0 {
+		return pgtype.Int4{}
+	}
+	return pgtype.Int4{Int32: int32(n), Valid: true}
+}
+
+func pgtypeInt8(v int64) pgtype.Int8 {
+	if v <= 0 {
+		return pgtype.Int8{}
+	}
+	return pgtype.Int8{Int64: v, Valid: true}
+}
+
+func pgtypeTimestamptz(t time.Time) pgtype.Timestamptz {
+	return pgtype.Timestamptz{Time: t, Valid: true}
+}
+
+func pgtypeDate(t time.Time) pgtype.Date {
+	return pgtype.Date{Time: t.UTC().Truncate(24 * time.Hour), Valid: true}
+}
+
+// feedUrlForComparison normalizes trailing slashes for the redirect check.
+func feedUrlForComparison(u string) string {
+	return strings.TrimRight(strings.TrimSpace(u), "/")
+}