mangaBookmark/backend/internal/store/store.go

package store

import (
	"database/sql"
	"embed"
	"errors"
	"fmt"
	"io/fs"
	"path"
	"regexp"
	"slices"
	"strconv"
	"strings"

	_ "github.com/jackc/pgx/v5/stdlib"
)

// Bookmark is one tracked series, keyed "<site>:<series_id>" across both sites.
//
// LastChapter* is the user's read progress; LatestChapter* is the newest
// chapter the site has published, captured opportunistically by the userscript.
//
// Title, SeriesURL, Cover, Kind and LatestChapter* live on the shared Series
// row (ADR-0003) and are joined in on read; Bookmark carries only what differs
// between readers: progress, favourite, lifecycle bucket, updated_at. The wire
// format stays flat regardless — see ADR-0004.
type Bookmark struct {
	Key              string   `json:"key"`
	Site             string   `json:"site"`
	SeriesID         string   `json:"series_id"`
	Title            string   `json:"title"`
	SeriesURL        string   `json:"series_url"`
	Cover            string   `json:"cover"`
	LastChapter      string   `json:"last_chapter"`
	LastChapterNum   float64  `json:"last_chapter_num"`
	LastChapterURL   string   `json:"last_chapter_url"`
	Favorite         bool     `json:"favorite"`
	LatestChapter    string   `json:"latest_chapter"`
	LatestChapterNum *float64 `json:"latest_chapter_num"` // nil until first captured
	UpdatedAt        int64    `json:"updated_at"`         // unix ms; see Upsert
	// Status is the lifecycle bucket: reading, archived, or finished.
	// Archived series stay polled for new chapters; finished ones do not.
	// Empty on the way in means "no opinion" — see Upsert.
	Status string `json:"status"`
	// Kind is the library bucket: manga or novel. Empty on the way in means
	// "no opinion" — see Upsert.
	Kind string `json:"kind"`
}

// Series is one distinct work, shared by every bookmark that tracks it. It is
// keyed (site, series_id) — the pair a bookmark key decomposes into — and
// exists once no matter how many bookmarks point at it (ADR-0003).
//
// Title, SeriesURL and Cover are written once, at creation: a PUT naming an
// existing Series has them ignored, and only the backend's own Poll may change
// them. Kind and the latest-chapter fields are last-write-wins like the
// bookmark's own fields. Never serialized: the wire format is the flat
// Bookmark (ADR-0004).
type Series struct {
	Site             string
	SeriesID         string
	Title            string
	SeriesURL        string
	Cover            string
	Kind             string
	LatestChapter    string
	LatestChapterNum *float64 // nil until first captured
	LatestCheckedAt  int64    // unix ms; see MarkLatestChecked

	// readerCount is the number of bookmarks referencing this series, filled
	// only by the due-queue query that orders on it.
	readerCount int
}

// Key returns the canonical identity in bookmark-key form ("<site>:<series_id>"),
// used by the poller's logs and by tests asserting on the due queue.
func (s Series) Key() string { return s.Site + ":" + s.SeriesID }

// HasNewChapter reports whether the site has published past the read point.
// A nil LatestChapterNum means nothing has been captured yet, which is not the
// same as "nothing new".
func (b Bookmark) HasNewChapter() bool {
	return b.LatestChapterNum != nil && *b.LatestChapterNum > b.LastChapterNum
}

// chapterLeadIn matches the prefix the userscript and the poller both write
// ("Chapter 250"), so the UI can add exactly one "Ch " of its own instead of
// doubling it. A manual edit through the web UI stores a bare "250", which is
// the same string minus the lead-in.
var chapterLeadIn = regexp.MustCompile(`(?i)^\s*(?:chapter|ch\.?)\s*`)

func displayChapter(raw string, num float64) string {
	rest := strings.TrimSpace(chapterLeadIn.ReplaceAllString(raw, ""))
	if rest == "" {
		rest = strconv.FormatFloat(num, 'f', -1, 64)
	}
	// "Ch " only makes sense in front of a number; anything else is a label the
	// site gave us, so pass it through as written.
	if rest[0] < '0' || rest[0] > '9' {
		return rest
	}
	return "Ch " + rest
}

// DisplayChapter is the read-progress line: one canonical "Ch N" whatever
// format the write came in as.
func (b Bookmark) DisplayChapter() string {
	return displayChapter(b.LastChapter, b.LastChapterNum)
}

// DisplayLatest is the same for the newest published chapter, which arrives
// with the same "Chapter N" lead-in from both the userscript and the poller.
func (b Bookmark) DisplayLatest() string {
	var num float64
	if b.LatestChapterNum != nil {
		num = *b.LatestChapterNum
	}
	return displayChapter(b.LatestChapter, num)
}

// ContinueURL is where the Continue button points: the chapter last read, or
// the series page when no chapter URL was ever captured.
func (b Bookmark) ContinueURL() string {
	if b.LastChapterURL != "" {
		return b.LastChapterURL
	}
	return b.SeriesURL
}

// Initial is the monogram the web UI shows in place of a cover when the
// source site never gave us an og:image. First rune, uppercased; "?" when even
// the title is missing, so the slot is never empty.
func (b Bookmark) Initial() string {
	for _, r := range b.Title {
		return strings.ToUpper(string(r))
	}
	return "?"
}

// kaganeCoverRe matches the cover URL kagane's og:image carries, which is what
// the userscript stores for that site.
var kaganeCoverRe = regexp.MustCompile(`^https://kagane\.to/api/v2/image/([0-9a-f-]{36})/compressed$`)

// CoverURL is the src the web UI puts in an <img>. For every site but kagane
// that is Cover as stored. kagane serves its images behind a Cloudflare
// challenge *and* with `cross-origin-resource-policy: same-origin`, so no page
// on another origin can load one however it asks (verified 2026-08-08); those
// go through the backend's own proxy instead.
func (b Bookmark) CoverURL() string {
	if m := kaganeCoverRe.FindStringSubmatch(b.Cover); m != nil {
		return "/img/kagane/" + m[1]
	}
	return b.Cover
}

// Library buckets. A bookmark is in exactly one. This cannot be derived from
// Site: asurascans serves manga and novels from the same /comics/ path, so the
// userscript that recorded the page is the only party that knows which.
const (
	KindManga = "manga"
	KindNovel = "novel"
)

// Lifecycle buckets. A bookmark is in exactly one; favorite is orthogonal.
const (
	StatusReading  = "reading"
	StatusArchived = "archived"
	StatusFinished = "finished"
)

//go:embed migrations/*.sql
var migrations embed.FS

// bookmarkColumns is the only value ever concatenated into query text. It is a
// compile-time constant; every request value is bound as a parameter. The
// series-owned fields are joined in from the series table, in scanBookmark
// order, so the flat Bookmark reads back whole despite the split (ADR-0004).
const bookmarkColumns = `b.site, b.series_id, s.title, s.series_url, s.cover,
	b.last_chapter, b.last_chapter_num, b.last_chapter_url,
	b.favorite, s.latest_chapter, s.latest_chapter_num, b.updated_at, b.status, s.kind`

// seriesColumns is the series row in scanSeries order, used by the poller's
// due query. latest_checked_at lives only on series — see MarkLatestChecked
// for why it stays off every client-visible write.
const seriesColumns = `s.site, s.series_id, s.title, s.series_url, s.cover,
	s.kind, s.latest_chapter, s.latest_chapter_num, s.latest_checked_at`

// Owner is the person running the service: the first Reader, seeded at startup
// so a fresh deployment has a library before anyone logs in. The seed makes
// sure exactly one readers row matches their Discord ID, carrying the SHA-256
// of their epoch-0 userscript credential (derived by internal/token). Every
// other Reader is created by their own first login (EnsureReader).
type Owner struct {
	DiscordID string
	// TokenHash is the SHA-256 of the epoch-0 credential; the array shape
	// makes it a compile error to store anything that is not a hash.
	TokenHash [32]byte
}

// Store is the Postgres-backed bookmark store.
type Store struct {
	db *sql.DB
	// ownerID is the seeded owner Reader (issue #22) — the only Reader with
	// administrative reach (revoking another Reader's sessions). Every store
	// method takes a reader id explicitly, so ownership is never implicit.
	ownerID int64
}

// OwnerID returns the seeded owner Reader's id: the administrator, and the
// Reader every pre-registration bookmark belongs to.
func (s *Store) OwnerID() int64 { return s.ownerID }

// ReaderIDForTokenHash resolves the Reader whose stored credential hash
// matches, reporting absence with ok=false. The comparison is an equality on
// the 32-byte SHA-256 of the presented credential — never on the credential
// itself — and the indexed lookup reveals only whether some Reader matches,
// which the 401/200 split has to reveal anyway. An attacker's probe is the
// hash of their guess, so even the index's prefix comparisons leak nothing
// about the real credential.
func (s *Store) ReaderIDForTokenHash(hash [32]byte) (int64, bool, error) {
	var id int64
	err := s.db.QueryRow(
		`SELECT id FROM readers WHERE token_sha256 = $1`, hash[:]).Scan(&id)
	if errors.Is(err, sql.ErrNoRows) {
		return 0, false, nil
	}
	if err != nil {
		return 0, false, fmt.Errorf("reader by token hash: %w", err)
	}
	return id, true, nil
}

// ReaderTokenInfo returns the identity halves a Reader's credential is
// derived from (internal/token.Token): their Discord id and token epoch. The
// web UI needs these to rebuild the install URL — the only place a credential
// is ever produced in plaintext.
func (s *Store) ReaderTokenInfo(readerID int64) (string, int64, error) {
	var (
		discordID string
		epoch     int64
	)
	err := s.db.QueryRow(
		`SELECT discord_id, token_epoch FROM readers WHERE id = $1`, readerID).
		Scan(&discordID, &epoch)
	if err != nil {
		return "", 0, fmt.Errorf("reader %d token info: %w", readerID, err)
	}
	return discordID, epoch, nil
}

// RotateToken bumps a Reader's token epoch and rewrites the stored hash in
// one statement, so the new hash always matches the new epoch. expectedEpoch
// is the epoch the caller derived newHash for (ReaderTokenInfo + 1); a
// concurrent rotation — or an unknown reader — leaves the row untouched and
// is reported as an error rather than silently succeeding.
func (s *Store) RotateToken(readerID, expectedEpoch int64, newHash [32]byte) error {
	var epoch int64
	err := s.db.QueryRow(`
		UPDATE readers SET token_epoch = token_epoch + 1, token_sha256 = $3
		WHERE id = $1 AND token_epoch = $2
		RETURNING token_epoch`, readerID, expectedEpoch, newHash[:]).Scan(&epoch)
	if errors.Is(err, sql.ErrNoRows) {
		return fmt.Errorf("rotate token for reader %d: concurrent rotation or unknown reader", readerID)
	}
	if err != nil {
		return fmt.Errorf("rotate token for reader %d: %w", readerID, err)
	}
	return nil
}

// EnsureReader returns the Reader registered to discordID, creating the row on
// first sight. Registration is open to every guild member (issue #27), and the
// Discord identity is the only thing that decides which Reader a login is: one
// code path serves the first login and every later one, so a returning Reader
// can never end up with a second library.
//
// epochZeroHash is only used for a brand-new row. An existing row keeps its
// stored hash untouched, or a login would silently undo a rotation and revive
// the credential the Reader rotated away from.
func (s *Store) EnsureReader(discordID string, epochZeroHash [32]byte) (int64, error) {
	var id int64
	// DO UPDATE rather than DO NOTHING because only an updated row is
	// returned by RETURNING; assigning the column to itself is the no-op that
	// makes the existing id come back.
	err := s.db.QueryRow(`
		INSERT INTO readers (discord_id, token_sha256) VALUES ($1, $2)
		ON CONFLICT (discord_id) DO UPDATE SET discord_id = readers.discord_id
		RETURNING id`, discordID, epochZeroHash[:]).Scan(&id)
	if err != nil {
		return 0, fmt.Errorf("ensure reader: %w", err)
	}
	return id, nil
}

// ReaderSummary is one Reader as the owner's administration panel sees them:
// who they are and how many live sessions they hold. No credential material,
// hashed or otherwise, is exposed.
type ReaderSummary struct {
	ID        int64
	DiscordID string
	// Sessions counts unexpired session rows — what the owner revokes.
	Sessions int
}

// Readers lists every Reader with their live session count, oldest first, so
// the owner row (always the oldest) heads the list.
func (s *Store) Readers() ([]ReaderSummary, error) {
	rows, err := s.db.Query(`
		SELECT r.id, r.discord_id,
		       count(sess.id) FILTER (WHERE sess.expires_at > now()) AS sessions
		FROM readers r
		LEFT JOIN sessions sess ON sess.reader_id = r.id
		GROUP BY r.id, r.discord_id
		ORDER BY r.id`)
	if err != nil {
		return nil, fmt.Errorf("query readers: %w", err)
	}
	defer rows.Close()

	out := []ReaderSummary{}
	for rows.Next() {
		var r ReaderSummary
		if err := rows.Scan(&r.ID, &r.DiscordID, &r.Sessions); err != nil {
			return nil, fmt.Errorf("scan reader: %w", err)
		}
		out = append(out, r)
	}
	return out, rows.Err()
}

// readersMigration is the version that creates the readers table. The owner
// seed runs between two migrate passes, so that the run-once migration which
// attaches existing bookmarks (0004) finds the owner row.
const readersMigration = 3

// allMigrations is the migrate() cap that applies every pending version.
const allMigrations = 0

// Open connects to Postgres at url — a libpq connection URL such as
// "postgres://user:pass@host:5432/bookmarks?sslmode=disable" — brings its
// schema up to date, and seeds the owner Reader.
func Open(url string, owner Owner) (*Store, error) {
	db, err := sql.Open("pgx", url)
	if err != nil {
		return nil, fmt.Errorf("open postgres: %w", err)
	}
	// Schema runs in two passes with the seed between: 0003 creates the
	// readers table, the owner row must exist before 0004 attaches the
	// existing bookmarks to it. Anything past 0004 is applied by the second
	// pass.
	if err := migrate(db, readersMigration); err != nil {
		db.Close()
		return nil, fmt.Errorf("migrate schema: %w", err)
	}
	// The owner row must exist before 0004 attaches the existing bookmarks to
	// it. The hash refresh is a separate statement after all migrations: the
	// token_epoch column 0006 adds does not exist yet at this point, and the
	// refresh only ever concerns rows that have never been rotated.
	if err := seedOwner(db, owner); err != nil {
		db.Close()
		return nil, fmt.Errorf("seed owner: %w", err)
	}
	if err := migrate(db, allMigrations); err != nil {
		db.Close()
		return nil, fmt.Errorf("migrate: %w", err)
	}
	if err := refreshOwnerToken(db, owner); err != nil {
		db.Close()
		return nil, fmt.Errorf("refresh owner token: %w", err)
	}
	var ownerID int64
	if err := db.QueryRow(
		`SELECT id FROM readers WHERE discord_id = $1`, owner.DiscordID).Scan(&ownerID); err != nil {
		db.Close()
		return nil, fmt.Errorf("resolve owner: %w", err)
	}
	return &Store{db: db, ownerID: ownerID}, nil
}

// seedOwner makes sure the configured owner exists as exactly one readers row.
// The hash is only ever written here for a brand-new row; existing rows keep
// what they have until refreshOwnerToken decides otherwise, so the seed can
// never clobber a rotation.
func seedOwner(db *sql.DB, o Owner) error {
	if _, err := db.Exec(`
		INSERT INTO readers (discord_id, token_sha256) VALUES ($1, $2)
		ON CONFLICT (discord_id) DO NOTHING`,
		o.DiscordID, o.TokenHash[:]); err != nil {
		return fmt.Errorf("seed owner: %w", err)
	}
	return nil
}

// refreshOwnerToken brings a never-rotated owner row's hash current with the
// configured credential. That is the cutover path: a database seeded under
// the retired global token still carries its hash at epoch 0, and the
// epoch-0 derivation is the caller's TokenHash. A rotated row (epoch > 0) is
// left alone — a restart must not resurrect the old credential by
// overwriting the hash a rotation wrote.
func refreshOwnerToken(db *sql.DB, o Owner) error {
	if _, err := db.Exec(`
		UPDATE readers SET token_sha256 = $2
		WHERE discord_id = $1 AND token_epoch = 0`,
		o.DiscordID, o.TokenHash[:]); err != nil {
		return fmt.Errorf("refresh owner token: %w", err)
	}
	return nil
}

// migrate applies every embedded migration this database has not recorded, in
// filename order, each in its own transaction. upto caps the highest version
// applied; 0 means all. Files are named "<version>_<name>.sql" and are
// append-only: editing an applied file changes nothing, because
// schema_migrations is how a database remembers what it ran. Runs on every
// start and is a no-op once current.
func migrate(db *sql.DB, upto int64) error {
	if _, err := db.Exec(`CREATE TABLE IF NOT EXISTS schema_migrations (
		version    bigint PRIMARY KEY,
		applied_at timestamptz NOT NULL DEFAULT now())`); err != nil {
		return fmt.Errorf("create version table: %w", err)
	}

	names, err := fs.Glob(migrations, "migrations/*.sql")
	if err != nil {
		return err
	}
	slices.Sort(names)

	for _, name := range names {
		version, err := strconv.ParseInt(strings.SplitN(path.Base(name), "_", 2)[0], 10, 64)
		if err != nil {
			return fmt.Errorf("migration %q: filename must start with a version number", name)
		}
		if upto > 0 && version > upto {
			continue
		}
		body, err := migrations.ReadFile(name)
		if err != nil {
			return err
		}
		if err := applyMigration(db, version, string(body)); err != nil {
			return fmt.Errorf("migration %q: %w", name, err)
		}
	}
	return nil
}

// applyMigration runs one migration and records its version in the same
// transaction, so an interrupted start leaves neither half behind.
func applyMigration(db *sql.DB, version int64, body string) error {
	tx, err := db.Begin()
	if err != nil {
		return err
	}
	defer tx.Rollback()

	var applied bool
	if err := tx.QueryRow(
		`SELECT EXISTS (SELECT 1 FROM schema_migrations WHERE version = $1)`,
		version).Scan(&applied); err != nil {
		return err
	}
	if applied {
		return nil
	}
	// No parameters, so this goes over the simple protocol and a migration may
	// hold more than one statement.
	if _, err := tx.Exec(body); err != nil {
		return err
	}
	if _, err := tx.Exec(`INSERT INTO schema_migrations (version) VALUES ($1)`, version); err != nil {
		return err
	}
	return tx.Commit()
}

// scanBookmark reads one row in bookmarkColumns order. Every column is NOT
// NULL except latest_chapter_num, where NULL means "never captured" — a
// distinct state from chapter zero, and the reason for the pointer.
func scanBookmark(scan func(...any) error) (Bookmark, error) {
	var (
		b                Bookmark
		latestChapterNum sql.NullFloat64
	)
	if err := scan(
		&b.Site, &b.SeriesID, &b.Title, &b.SeriesURL, &b.Cover,
		&b.LastChapter, &b.LastChapterNum, &b.LastChapterURL,
		&b.Favorite, &b.LatestChapter, &latestChapterNum, &b.UpdatedAt, &b.Status, &b.Kind,
	); err != nil {
		return Bookmark{}, err
	}
	if latestChapterNum.Valid {
		b.LatestChapterNum = &latestChapterNum.Float64
	}
	// The wire identity is derived: there is no stored key column, the
	// bookmark is keyed (reader_id, site, series_id) (issue #22).
	b.Key = b.Site + ":" + b.SeriesID
	// An unrecognised bucket (a hand-edited row) would leave the row in no list
	// at all, so anything outside the three known buckets reads as the default
	// rather than being passed through.
	if b.Status != StatusReading && b.Status != StatusArchived && b.Status != StatusFinished {
		b.Status = StatusReading
	}
	return b, nil
}

// scanSeries reads one row in seriesColumns order, plus the due query's
// reader_count column. latest_chapter_num is NULL until the first capture,
// same as on the bookmark read path.
func scanSeries(scan func(...any) error) (Series, error) {
	var (
		sr               Series
		latestChapterNum sql.NullFloat64
	)
	if err := scan(
		&sr.Site, &sr.SeriesID, &sr.Title, &sr.SeriesURL, &sr.Cover,
		&sr.Kind, &sr.LatestChapter, &latestChapterNum, &sr.LatestCheckedAt,
		&sr.readerCount,
	); err != nil {
		return Series{}, err
	}
	if latestChapterNum.Valid {
		sr.LatestChapterNum = &latestChapterNum.Float64
	}
	return sr, nil
}

// Close releases the underlying database handle.
func (s *Store) Close() error { return s.db.Close() }

// List returns every bookmark of one reader, newest activity first.
// Series-owned fields are joined in, so each Bookmark reads back whole and
// flat (ADR-0004).
func (s *Store) List(readerID int64) ([]Bookmark, error) {
	rows, err := s.db.Query(`SELECT `+bookmarkColumns+`
		FROM bookmarks b
		JOIN series s ON s.site = b.site AND s.series_id = b.series_id
		WHERE b.reader_id = $1
		ORDER BY b.updated_at DESC`, readerID)
	if err != nil {
		return nil, fmt.Errorf("query bookmarks: %w", err)
	}
	defer rows.Close()

	out := []Bookmark{}
	for rows.Next() {
		b, err := scanBookmark(rows.Scan)
		if err != nil {
			return nil, fmt.Errorf("scan bookmark: %w", err)
		}
		out = append(out, b)
	}
	return out, rows.Err()
}

// Get returns one bookmark of one reader by key. A missing key is not an
// error: ok is false and err is nil. UI mutations read-modify-write through
// this so they preserve the fields they do not touch.
func (s *Store) Get(readerID int64, key string) (Bookmark, bool, error) {
	site, seriesID, ok := strings.Cut(key, ":")
	if !ok {
		return Bookmark{}, false, nil
	}
	b, err := scanBookmark(s.db.QueryRow(
		`SELECT `+bookmarkColumns+` FROM bookmarks b
		JOIN series s ON s.site = b.site AND s.series_id = b.series_id
		WHERE b.reader_id = $1 AND b.site = $2 AND b.series_id = $3`,
		readerID, site, seriesID).Scan)
	if errors.Is(err, sql.ErrNoRows) {
		return Bookmark{}, false, nil
	}
	if err != nil {
		return Bookmark{}, false, fmt.Errorf("get %q: %w", key, err)
	}
	return b, true, nil
}

// Upsert inserts or replaces one reader's bookmark by key (last-write-wins)
// and returns the row as actually stored — one flat object with the
// series-owned fields joined in, exactly as GET reports it (ADR-0004). A
// bookmark is keyed (reader_id, site, series_id), so the same key upserts two
// independent rows for two readers.
//
// The flat body is decomposed across two tables in one transaction. The series
// row is written first (the bookmarks FK requires it to exist), then the
// bookmark row. On the series side, title/series_url/cover are applied only
// when the row is brand new: once a series exists, client-supplied values are
// ignored, because the row is shared and the values are scraped page content —
// see ADR-0003. Kind and the latest-chapter fields are last-write-wins.
//
// b.UpdatedAt is only a candidate: it is applied when the row is new or when
// last_chapter_num changes, and otherwise the stored value is kept. Clients
// order their list by updated_at, so favoriting a series or recording a newly
// published chapter must not disturb that order — only real reading progress
// does. Callers must therefore use the returned bookmark, not the argument.
func (s *Store) Upsert(readerID int64, b Bookmark) (Bookmark, error) {
	tx, err := s.db.Begin()
	if err != nil {
		return Bookmark{}, fmt.Errorf("begin %q: %w", b.Key, err)
	}
	defer tx.Rollback()

	var latestNum any
	if b.LatestChapterNum != nil {
		latestNum = *b.LatestChapterNum
	}

	// The kind column resolves on the VALUES side, not in the conflict clause:
	// excluded.* is the row *after* these expressions are evaluated, so a
	// default applied there would look identical to a real 'manga' and would
	// overwrite a novel series on every PUT from a client that knows nothing
	// about the column. Resolved once here, an empty incoming kind means "keep
	// what is stored", and only a brand-new row falls through to the literal
	// default. The subquery runs inside this transaction, so it sees the row
	// this statement is about to conflict with. Same pattern as the status
	// COALESCE on the bookmark insert below.
	//
	// The ::text casts are load-bearing: inside COALESCE/NULLIF there is no
	// target column to infer the parameter type from, and Postgres rejects the
	// statement rather than guessing.
	if _, err := tx.Exec(`
		INSERT INTO series (site, series_id, title, series_url, cover, kind,
		                    latest_chapter, latest_chapter_num)
		VALUES ($1, $2, $3, $4, $5,
		        COALESCE(NULLIF($6::text, ''), (SELECT kind FROM series WHERE site = $1 AND series_id = $2), 'manga'),
		        $7, $8)
		ON CONFLICT (site, series_id) DO UPDATE SET
		  kind=excluded.kind,
		  latest_chapter=excluded.latest_chapter,
		  latest_chapter_num=excluded.latest_chapter_num`,
		b.Site, b.SeriesID, b.Title, b.SeriesURL, b.Cover, b.Kind,
		b.LatestChapter, latestNum); err != nil {
		return Bookmark{}, fmt.Errorf("upsert series for %q: %w", b.Key, err)
	}

	// IS DISTINCT FROM is Postgres's null-safe comparison, and it is what
	// implements the ordering rule. Within DO UPDATE, a bare column is the
	// stored row and excluded.* is the incoming one; a brand-new key never
	// reaches this clause, so it keeps the fresh timestamp from VALUES.
	if _, err := tx.Exec(`
		INSERT INTO bookmarks (reader_id, site, series_id, last_chapter, last_chapter_num,
		                       last_chapter_url, favorite, status, updated_at)
		VALUES ($1, $2, $3, $4, $5, $6, $7,
		        COALESCE(NULLIF($8::text, ''), (SELECT status FROM bookmarks WHERE reader_id = $1 AND site = $2 AND series_id = $3), 'reading'),
		        $9)
		ON CONFLICT (reader_id, site, series_id) DO UPDATE SET
		  last_chapter=excluded.last_chapter, last_chapter_num=excluded.last_chapter_num,
		  last_chapter_url=excluded.last_chapter_url,
		  favorite=excluded.favorite,
		  status=excluded.status,
		  updated_at=CASE
		    WHEN bookmarks.last_chapter_num IS DISTINCT FROM excluded.last_chapter_num
		    THEN excluded.updated_at
		    ELSE bookmarks.updated_at
		  END`,
		readerID, b.Site, b.SeriesID,
		b.LastChapter, b.LastChapterNum, b.LastChapterURL,
		b.Favorite, b.Status, b.UpdatedAt); err != nil {
		return Bookmark{}, fmt.Errorf("upsert %q: %w", b.Key, err)
	}

	stored, err := scanBookmark(tx.QueryRow(
		`SELECT `+bookmarkColumns+` FROM bookmarks b
		JOIN series s ON s.site = b.site AND s.series_id = b.series_id
		WHERE b.reader_id = $1 AND b.site = $2 AND b.series_id = $3`,
		readerID, b.Site, b.SeriesID).Scan)
	if err != nil {
		return Bookmark{}, fmt.Errorf("read back %q: %w", b.Key, err)
	}
	if err := tx.Commit(); err != nil {
		return Bookmark{}, fmt.Errorf("commit %q: %w", b.Key, err)
	}
	return stored, nil
}

// Delete removes one reader's bookmark by key. Deleting a missing key is not
// an error.
func (s *Store) Delete(readerID int64, key string) error {
	site, seriesID, ok := strings.Cut(key, ":")
	if !ok {
		return nil
	}
	if _, err := s.db.Exec(
		`DELETE FROM bookmarks WHERE reader_id = $1 AND site = $2 AND series_id = $3`,
		readerID, site, seriesID); err != nil {
		return fmt.Errorf("delete %q: %w", key, err)
	}
	return nil
}

// DueForLatestCheck returns series whose server-side latest-chapter check has
// aged past cutoffMs, ordered by how many bookmarks reference them (descending)
// then least-recently-checked first, at most limit of them.
//
// The reader_count ordering is the point of the split (ADR-0003): a series
// shared by several readers is fetched once per due cycle, and the popular
// ones stay freshest while the long tail absorbs any shortfall. Within one
// reader count, oldest-first keeps the poll fair when the backlog outgrows its
// throughput: the most neglected series is always next, so a large collection
// refreshes uniformly slower rather than leaving a tail that never refreshes at
// all. The userscript sorts its own queue the same way (L453).
//
// Series with no series_url are skipped — there is nothing to fetch, which is
// the same filter the userscript applies at L452. Series whose only bookmarks
// are finished are skipped too: nothing more is coming, so fetching them only
// burns requests. Archived bookmarks still count — knowing what a shelved
// series is up to is the whole reason for archiving instead of deleting.
// A series with no bookmarks at all never appears: the join excludes it.
func (s *Store) DueForLatestCheck(cutoffMs int64, limit int) ([]Series, error) {
	rows, err := s.db.Query(`SELECT `+seriesColumns+`, COUNT(*) AS reader_count
		FROM series s
		JOIN bookmarks b ON b.site = s.site AND b.series_id = s.series_id
		WHERE s.series_url <> ''
		  AND s.latest_checked_at <= $1
		GROUP BY s.site, s.series_id, s.title, s.series_url, s.cover,
		         s.kind, s.latest_chapter, s.latest_chapter_num, s.latest_checked_at
		HAVING COUNT(*) FILTER (WHERE b.status <> 'finished') > 0
		ORDER BY reader_count DESC, s.latest_checked_at ASC
		LIMIT $2`, cutoffMs, limit)
	if err != nil {
		return nil, fmt.Errorf("query due series: %w", err)
	}
	defer rows.Close()

	out := []Series{}
	for rows.Next() {
		sr, err := scanSeries(rows.Scan)
		if err != nil {
			return nil, fmt.Errorf("scan due series: %w", err)
		}
		out = append(out, sr)
	}
	return out, rows.Err()
}

// MarkLatestChecked records that the server looked at a series at ts, whatever
// the look turned up. Marking a missing series is not an error: the row may
// have been orphaned while a fetch was in flight.
//
// This is the one write that does not go through Upsert, and the column is kept
// out of the client-visible read path on purpose. PUT /bookmarks/{key} decodes
// a whole Bookmark from the client and Upsert writes every series column it
// knows about, so a userscript PUT — which has no idea this field exists —
// would write a zero and reset the cooldown, making the poller re-fetch that
// series every tick for as long as the user kept reading it.
func (s *Store) MarkLatestChecked(site, seriesID string, ts int64) error {
	if _, err := s.db.Exec(
		`UPDATE series SET latest_checked_at = $1 WHERE site = $2 AND series_id = $3`,
		ts, site, seriesID); err != nil {
		return fmt.Errorf("mark checked %s:%s: %w", site, seriesID, err)
	}
	return nil
}

// LatestCheckedAt reads the column MarkLatestChecked writes. It exists for
// tests outside this package (the poller's own tests assert on cooldown
// bookkeeping) — see MarkLatestChecked for why the field stays off the
// client-visible row.
func (s *Store) LatestCheckedAt(site, seriesID string) (int64, error) {
	var ts int64
	if err := s.db.QueryRow(
		`SELECT latest_checked_at FROM series WHERE site = $1 AND series_id = $2`,
		site, seriesID).Scan(&ts); err != nil {
		return 0, fmt.Errorf("latest checked at %s:%s: %w", site, seriesID, err)
	}
	return ts, nil
}

// SetLatestChapter records the newest chapter the poll found on a series page.
// The poller walks Series rather than Bookmarks, so this is a series-level
// write: the row is shared, and updating it once refreshes every bookmark that
// joins to it. Touching a missing series is not an error.
func (s *Store) SetLatestChapter(site, seriesID, label string, num float64) error {
	if _, err := s.db.Exec(
		`UPDATE series SET latest_chapter = $3, latest_chapter_num = $4
		 WHERE site = $1 AND series_id = $2`,
		site, seriesID, label, num); err != nil {
		return fmt.Errorf("set latest chapter %s:%s: %w", site, seriesID, err)
	}
	return nil
}