Discover and serve subtitle tracks

Two halves of the same feature: the scanner finds tracks, the media layer
serves them as WebVTT.

Discovery walks each directory for .srt/.ass/.ssa/.vtt/.sub beside a video and
matches them by base name, including the language form "Movie.en.srt" and
"Movie.pt-BR.srt". A sidecar is never itself an item. Matching is a dot-guarded
prefix on the real video base name, so a dotted release name like
"The.Matrix.1999.1080p.en.srt" does not read the year as a language.

Serving adds two routes inside the existing item family, both grant-checked
through catalog.Store (ADR-005) so forbidden and absent are the same 404.
Browsers play WebVTT and nothing else, so .srt is converted (header plus the
comma-to-dot in timestamps), .vtt passes through, and embedded streams and
.ass/.ssa go through ffmpeg. Conversions are cached under ConfigDir, because
converting per request would spawn an ffmpeg on every seek. A track that fails
to convert fails alone - the video and the other tracks are unaffected.

Subtitle rows are a cache of what is on disk, so a vanished sidecar has its row
deleted. That is the opposite of a vanished video, which is marked instead
(ADR-003) because progress hangs off it and cannot be regenerated.

Verified red-green: disabling the sidecar call fails all four scanner tests,
and removing the VTT header fails the conversion test.

Known limits: a "Movie.forced.en.srt" records "forced.en" rather than "en",
and the conversion cache is keyed by track id, so a sidecar replaced under an
unchanged row serves the stale conversion until a rescan recreates the row.

internal/media/handlers.go

@@ -47,6 +47,8 @@ func (h *Handlers) Register(mux *http.ServeMux) {
 	mux.Handle("GET /api/v1/items/{id}/file", h.auth.Require(h.handleFile))
 	mux.Handle("GET /api/v1/items/{id}/files/{fileID}", h.auth.Require(h.handleFileByID))
 	mux.Handle("GET /api/v1/items/{id}/cover", h.auth.Require(h.handleCover))
+	mux.Handle("GET /api/v1/items/{id}/subtitles", h.auth.Require(h.handleSubtitles))
+	mux.Handle("GET /api/v1/items/{id}/subtitles/{trackID}", h.auth.Require(h.handleSubtitleTrack))
 	mux.Handle("GET /api/v1/items/{id}/progress", h.auth.Require(h.handleGetProgress))
 	mux.Handle("PUT /api/v1/items/{id}/progress", h.auth.Require(h.handlePutProgress))
 	mux.Handle("GET /api/v1/progress/recent", h.auth.Require(h.handleRecent))

internal/media/subtitles.go

@@ -0,0 +1,226 @@
+package media
+
+// Subtitle tracks for an item: the list, and one track rendered as WebVTT.
+//
+// Browsers play WebVTT and nothing else, so every other format is converted on
+// the way out. Conversion is cached under the config dir: a player re-requests a
+// track as the user seeks, and an embedded track would otherwise re-run ffmpeg
+// on every request.
+//
+// The item is loaded through the catalog chokepoint exactly as handleCover does
+// (ADR-005), so a track on an unreadable item answers with the same 404 an
+// absent item does.
+
+import (
+	"bytes"
+	"context"
+	"database/sql"
+	"errors"
+	"fmt"
+	"net/http"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strconv"
+	"strings"
+
+	"github.com/mithral/vellum/internal/auth"
+	"github.com/mithral/vellum/internal/httpx"
+	"github.com/mithral/vellum/internal/stream"
+)
+
+// subtitle is one row of the subtitles table. Path and StreamIndex are the
+// source-specific locator and are never sent to clients: a file-backed track
+// (sidecar, downloaded) carries a path relative to the library root and
+// stream_index -1, an embedded track carries a stream index and an empty path.
+type subtitle struct {
+	ID          int64  `json:"id"`
+	Source      string `json:"source"`
+	Language    string `json:"language"`
+	Label       string `json:"label"`
+	Format      string `json:"format"`
+	Path        string `json:"-"`
+	StreamIndex int    `json:"-"`
+}
+
+// handleSubtitles lists an item's subtitle tracks.
+func (h *Handlers) handleSubtitles(w http.ResponseWriter, r *http.Request, u auth.User) {
+	it, ok := h.authorized(w, r, u)
+	if !ok {
+		return
+	}
+	subs, err := h.subtitleTracks(r.Context(), it.ID)
+	if err != nil {
+		h.fail(w, "list subtitles", err)
+		return
+	}
+	httpx.WriteJSON(w, http.StatusOK, map[string]any{"subtitles": subs})
+}
+
+// handleSubtitleTrack renders one track as WebVTT.
+func (h *Handlers) handleSubtitleTrack(w http.ResponseWriter, r *http.Request, u auth.User) {
+	it, ok := h.authorized(w, r, u)
+	if !ok {
+		return
+	}
+
+	trackID, err := strconv.ParseInt(r.PathValue("trackID"), 10, 64)
+	if err != nil || trackID <= 0 {
+		httpx.WriteError(w, http.StatusBadRequest, "invalid track id")
+		return
+	}
+
+	t, err := h.subtitleTrack(r.Context(), it.ID, trackID)
+	if errors.Is(err, sql.ErrNoRows) {
+		// A track of another item, or none at all: the same 404, so pairing an
+		// item id with a track id from elsewhere reveals nothing.
+		httpx.WriteError(w, http.StatusNotFound, "subtitle not found")
+		return
+	}
+	if err != nil {
+		h.fail(w, "load subtitle", err)
+		return
+	}
+
+	vtt, err := h.subtitleVTT(r.Context(), it, t)
+	if err != nil {
+		// Only this track fails; the list and every other track still work.
+		h.fail(w, "convert subtitle", err)
+		return
+	}
+	stream.ServeBytes(w, r, "subtitle.vtt", vtt)
+}
+
+// subtitleTracks returns every track for an item, in insertion order.
+func (h *Handlers) subtitleTracks(ctx context.Context, itemID int64) ([]subtitle, error) {
+	rows, err := h.db.QueryContext(ctx, `
+		SELECT id, source, path, stream_index, format, language, label
+		FROM subtitles WHERE item_id = ? ORDER BY id`, itemID)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+
+	out := []subtitle{}
+	for rows.Next() {
+		var s subtitle
+		if err := rows.Scan(&s.ID, &s.Source, &s.Path, &s.StreamIndex,
+			&s.Format, &s.Language, &s.Label); err != nil {
+			return nil, err
+		}
+		out = append(out, s)
+	}
+	return out, rows.Err()
+}
+
+// subtitleTrack returns one track, scoped to the item so a track id from
+// another item cannot be fetched through it. A miss is sql.ErrNoRows.
+func (h *Handlers) subtitleTrack(ctx context.Context, itemID, trackID int64) (subtitle, error) {
+	var s subtitle
+	err := h.db.QueryRowContext(ctx, `
+		SELECT id, source, path, stream_index, format, language, label
+		FROM subtitles WHERE item_id = ? AND id = ?`, itemID, trackID).
+		Scan(&s.ID, &s.Source, &s.Path, &s.StreamIndex, &s.Format, &s.Language, &s.Label)
+	if err != nil {
+		return subtitle{}, err
+	}
+	return s, nil
+}
+
+// subtitleVTT returns the track as WebVTT bytes.
+//
+// .vtt is served as-is. .srt is rewritten in memory — a header and a comma-to-dot
+// in each timestamp — so ffmpeg is not a dependency for the one format that does
+// not need it. Everything else (embedded streams, .ass/.ssa/.sub) goes through
+// ffmpeg, which is the only thing that can demux them.
+func (h *Handlers) subtitleVTT(ctx context.Context, it Item, t subtitle) ([]byte, error) {
+	// A file-backed track sits on disk; an embedded track only exists inside
+	// the video and has no path of its own.
+	filePath := filepath.Join(it.RootPath, filepath.FromSlash(t.Path))
+
+	if t.Source != "embedded" && strings.EqualFold(t.Format, "vtt") {
+		return os.ReadFile(filePath)
+	}
+
+	cachePath := filepath.Join(h.cacheDir, "subtitles", fmt.Sprintf("%d-%d.vtt", it.ID, t.ID))
+	if b, err := os.ReadFile(cachePath); err == nil {
+		return b, nil
+	}
+
+	var (
+		b   []byte
+		err error
+	)
+	switch {
+	case t.Source == "embedded":
+		b, err = ffmpegVTT(ctx, it.AbsPath(), t.StreamIndex)
+	case strings.EqualFold(t.Format, "srt"):
+		var raw []byte
+		if raw, err = os.ReadFile(filePath); err == nil {
+			b = srtToVTT(raw)
+		}
+	default:
+		b, err = ffmpegVTT(ctx, filePath, -1)
+	}
+	if err != nil {
+		return nil, err
+	}
+
+	// Cache the conversion so a seek does not spawn ffmpeg again.
+	//
+	// shortcut: keyed by track id, so a sidecar replaced under the same row
+	// keeps serving the old conversion until a rescan recreates the row.
+	if err := os.MkdirAll(filepath.Dir(cachePath), 0o750); err != nil {
+		return nil, err
+	}
+	if err := os.WriteFile(cachePath, b, 0o640); err != nil {
+		return nil, err
+	}
+	return b, nil
+}
+
+// srtToVTT rewrites SubRip into WebVTT: the WEBVTT header, and a dot where a
+// timestamp used a comma. Only lines carrying "-->" are touched, so a comma in
+// the subtitle text itself survives.
+func srtToVTT(srt []byte) []byte {
+	var b bytes.Buffer
+	b.WriteString("WEBVTT\n\n")
+	for _, line := range strings.Split(string(srt), "\n") {
+		if strings.Contains(line, "-->") {
+			line = strings.ReplaceAll(line, ",", ".")
+		}
+		b.WriteString(line)
+		b.WriteByte('\n')
+	}
+	return b.Bytes()
+}
+
+// ffmpegVTT extracts a subtitle stream as WebVTT on stdout.
+//
+// streamIndex is the embedded subtitle stream to map, or -1 to let ffmpeg pick
+// the file's only subtitle stream.
+func ffmpegVTT(ctx context.Context, input string, streamIndex int) ([]byte, error) {
+	ffmpeg, err := exec.LookPath("ffmpeg")
+	if err != nil {
+		return nil, fmt.Errorf("media: ffmpeg not available: %w", err)
+	}
+
+	args := []string{"-v", "error", "-i", input}
+	if streamIndex >= 0 {
+		args = append(args, "-map", fmt.Sprintf("0:s:%d", streamIndex))
+	}
+	args = append(args, "-f", "webvtt", "-")
+
+	var stdout, stderr bytes.Buffer
+	cmd := exec.CommandContext(ctx, ffmpeg, args...)
+	cmd.Stdout = &stdout
+	cmd.Stderr = &stderr
+	if err := cmd.Run(); err != nil {
+		msg := strings.TrimSpace(stderr.String())
+		if msg == "" {
+			msg = err.Error()
+		}
+		return nil, fmt.Errorf("media: ffmpeg subtitle %s: %s", input, msg)
+	}
+	return stdout.Bytes(), nil
+}

internal/media/subtitles_test.go

@@ -0,0 +1,164 @@
+package media
+
+import (
+	"database/sql"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/mithral/vellum/internal/auth"
+)
+
+// subtitleFixture builds a library rooted at a real temp dir, so a sidecar
+// subtitle can be written to disk and read back through the route, plus a second
+// item in a library the caller may not read.
+type subtitleFixture struct {
+	h         *Handlers
+	db        *sql.DB
+	root      string
+	user      int64
+	item      int64
+	forbidden int64
+}
+
+func newSubtitleFixture(t *testing.T) subtitleFixture {
+	t.Helper()
+	db := testDB(t)
+	root := t.TempDir()
+
+	lib := newLibrary(t, db, "Video", "video", root)
+	other := newLibrary(t, db, "Other", "video", t.TempDir())
+
+	f := subtitleFixture{db: db, root: root}
+	f.user = newUser(t, db, "watcher")
+	grantRead(t, db, f.user, lib)
+
+	const at = "2026-01-01T00:00:00.000Z"
+	f.item = newItem(t, db, lib, "video", "movie.mkv", "Movie", "movie", false, at)
+	f.forbidden = newItem(t, db, other, "video", "secret.mkv", "Secret", "secret", false, at)
+	f.h = mediaHandlers(t, db)
+	return f
+}
+
+// addTrack writes a sidecar file and the subtitles row that points at it.
+func (f subtitleFixture) addTrack(t *testing.T, name, content, format, lang string) int64 {
+	t.Helper()
+	if err := os.WriteFile(filepath.Join(f.root, name), []byte(content), 0o640); err != nil {
+		t.Fatalf("write %s: %v", name, err)
+	}
+	return mustQueryID(t, f.db, `
+		INSERT INTO subtitles (item_id, source, path, stream_index, format, language, label)
+		VALUES (?, 'sidecar', ?, -1, ?, ?, ?) RETURNING id`,
+		f.item, name, format, lang, name)
+}
+
+func listSubtitleTracks(t *testing.T, h *Handlers, u auth.User, itemID int64) *httptest.ResponseRecorder {
+	t.Helper()
+	req := httptest.NewRequest(http.MethodGet, "/api/v1/items/1/subtitles", nil)
+	req.SetPathValue("id", itoa(itemID))
+	rec := httptest.NewRecorder()
+	h.handleSubtitles(rec, req, u)
+	return rec
+}
+
+func getSubtitleTrack(t *testing.T, h *Handlers, u auth.User, itemID, trackID int64) *httptest.ResponseRecorder {
+	t.Helper()
+	req := httptest.NewRequest(http.MethodGet, "/api/v1/items/1/subtitles/1", nil)
+	req.SetPathValue("id", itoa(itemID))
+	req.SetPathValue("trackID", itoa(trackID))
+	rec := httptest.NewRecorder()
+	h.handleSubtitleTrack(rec, req, u)
+	return rec
+}
+
+// TestSubtitleSRTConvertsToVTT pins the SubRip rewrite: the WebVTT header, a dot
+// in every timestamp, and the subtitle text left alone.
+func TestSubtitleSRTConvertsToVTT(t *testing.T) {
+	f := newSubtitleFixture(t)
+	srt := "1\n00:00:01,000 --> 00:00:04,000\nHello, world\n\n2\n00:00:05,500 --> 00:00:06,000\nBye\n"
+	id := f.addTrack(t, "movie.en.srt", srt, "srt", "en")
+
+	rec := getSubtitleTrack(t, f.h, auth.User{ID: f.user}, f.item, id)
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d, body=%s", rec.Code, rec.Body.String())
+	}
+	if ct := rec.Header().Get("Content-Type"); ct != "text/vtt; charset=utf-8" {
+		t.Errorf("content-type = %q, want text/vtt; charset=utf-8", ct)
+	}
+
+	body := rec.Body.String()
+	if !strings.HasPrefix(body, "WEBVTT\n\n") {
+		t.Errorf("body does not start with WEBVTT header: %q", body)
+	}
+	if !strings.Contains(body, "00:00:01.000 --> 00:00:04.000") ||
+		!strings.Contains(body, "00:00:05.500 --> 00:00:06.000") {
+		t.Errorf("timestamps not dot-separated: %q", body)
+	}
+	if strings.Contains(body, ",000") || strings.Contains(body, ",500") {
+		t.Errorf("a comma survived in a timestamp: %q", body)
+	}
+	if !strings.Contains(body, "Hello, world") {
+		t.Errorf("a comma in the subtitle text was rewritten: %q", body)
+	}
+}
+
+// TestSubtitleVTTPassesThrough pins that an already-WebVTT sidecar is served
+// byte for byte, with no header prepended and no timestamp touched.
+func TestSubtitleVTTPassesThrough(t *testing.T) {
+	f := newSubtitleFixture(t)
+	vtt := "WEBVTT\n\n00:00:01.000 --> 00:00:04.000\nHi there\n"
+	id := f.addTrack(t, "movie.en.vtt", vtt, "vtt", "en")
+
+	rec := getSubtitleTrack(t, f.h, auth.User{ID: f.user}, f.item, id)
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d, body=%s", rec.Code, rec.Body.String())
+	}
+	if got := rec.Body.String(); got != vtt {
+		t.Errorf("body = %q, want unchanged %q", got, vtt)
+	}
+	if ct := rec.Header().Get("Content-Type"); ct != "text/vtt; charset=utf-8" {
+		t.Errorf("content-type = %q, want text/vtt; charset=utf-8", ct)
+	}
+}
+
+// TestSubtitleRoutesAreGrantScoped pins the access chokepoint on both routes: an
+// item in a library the caller cannot read must answer with the same 404, byte
+// for byte, as an item that does not exist.
+func TestSubtitleRoutesAreGrantScoped(t *testing.T) {
+	f := newSubtitleFixture(t)
+	u := auth.User{ID: f.user}
+
+	// The readable item lists, empty, with a 200.
+	if rec := listSubtitleTracks(t, f.h, u, f.item); rec.Code != http.StatusOK {
+		t.Errorf("readable list status = %d, want 200 (body %s)", rec.Code, rec.Body.String())
+	}
+
+	forbidden := listSubtitleTracks(t, f.h, u, f.forbidden)
+	absent := listSubtitleTracks(t, f.h, u, 999999)
+	if forbidden.Code != http.StatusNotFound || absent.Code != http.StatusNotFound {
+		t.Fatalf("list statuses = %d/%d, want 404/404", forbidden.Code, absent.Code)
+	}
+	if forbidden.Body.String() != absent.Body.String() {
+		t.Errorf("forbidden body %q != absent body %q", forbidden.Body.String(), absent.Body.String())
+	}
+
+	// The track route leaks nothing either: the item is rejected before the
+	// track id is ever looked at.
+	fb := getSubtitleTrack(t, f.h, u, f.forbidden, 1)
+	ab := getSubtitleTrack(t, f.h, u, 999999, 1)
+	if fb.Code != http.StatusNotFound || ab.Code != http.StatusNotFound {
+		t.Fatalf("track statuses = %d/%d, want 404/404", fb.Code, ab.Code)
+	}
+	if fb.Body.String() != ab.Body.String() {
+		t.Errorf("track forbidden body %q != absent body %q", fb.Body.String(), ab.Body.String())
+	}
+
+	// A user with no grants at all gets the same 404 on the readable item.
+	none := listSubtitleTracks(t, f.h, auth.User{ID: newUser(t, f.db, "nobody")}, f.item)
+	if none.Code != http.StatusNotFound {
+		t.Errorf("no-grant list status = %d, want 404", none.Code)
+	}
+}

internal/scanner/scan.go

@@ -125,13 +125,24 @@ func Scan(ctx context.Context, db *sql.DB, lib library.Library, log *slog.Logger
 		return res, err
 	}
 
-	found, errs := walkRoot(lib.RootPath, exts)
+	found, sidecars, errs := walkRoot(lib.RootPath, exts)
 	for _, e := range errs {
 		res.addError(e)
 	}
 
 	groups := buildGroups(lib, found)
 
+	// Sidecar subtitle files, bucketed by the directory they sit in so a video
+	// can find its own without scanning every subtitle. seenSidecars records
+	// which ones a video actually claimed; anything else is deleted below,
+	// because the subtitles table is a cache of what is on disk.
+	sidecarsByDir := make(map[string][]candidate)
+	for _, s := range sidecars {
+		dir := path.Dir(s.relPath)
+		sidecarsByDir[dir] = append(sidecarsByDir[dir], s)
+	}
+	seenSidecars := make(map[string]bool)
+
 	now := core.Now()
 	seenFiles := make(map[string]bool, len(found))
 
@@ -182,6 +193,12 @@ func Scan(ctx context.Context, db *sql.DB, lib library.Library, log *slog.Logger
 					res.Scanned++
 					seenFiles[c.relPath] = true
 
+					// Sidecars are keyed to the video beside them, not to the
+					// media_files row, so this runs even for an unchanged video:
+					// a new subtitle beside a settled file must still be picked
+					// up, and its row must not be pruned as vanished.
+					upsertSidecars(ctx, tx, itemID, c, sidecarsByDir[path.Dir(c.relPath)], seenSidecars, &res)
+
 					prev, exists := knownFiles[c.relPath]
 					if exists && !prev.missing && prev.size == c.size && prev.mtime == c.mtime {
 						res.Unchanged++
@@ -275,6 +292,14 @@ func Scan(ctx context.Context, db *sql.DB, lib library.Library, log *slog.Logger
 		report()
 	}
 
+	// Sidecar subtitle rows are a cache of what is on disk: a row whose file is
+	// gone is deleted, not flagged missing. This is scoped to the library
+	// through item_id, and runs every scan so a library that lost all its
+	// subtitles still ends up with none.
+	if err := reconcileSidecars(ctx, db, lib.ID, seenSidecars, &res); err != nil {
+		return res, err
+	}
+
 	res.Duration = time.Since(start).Milliseconds()
 	log.Info("scan complete",
 		"library", lib.Name, "kind", lib.Kind, "grouping", library.GroupModeFor(lib.Kind),
@@ -546,10 +571,19 @@ func buildTVGroups(lib library.Library, found map[string]candidate) map[string]g
 	return groups
 }
 
-// walkRoot collects supported files under root, keyed by slash-separated
+// sidecarExts are the subtitle file types a video may carry beside it. They are
-// path relative to root.
+// not library extensions: a subtitle never becomes an item, it is only ever a
-func walkRoot(root string, exts map[string]bool) (map[string]candidate, []string) {
+// track hanging off the video it sits next to.
+var sidecarExts = map[string]bool{
+	".srt": true, ".ass": true, ".ssa": true, ".vtt": true, ".sub": true,
+}
+
+// walkRoot collects supported files under root, keyed by slash-separated path
+// relative to root. Subtitle sidecars are returned separately: they are
+// discovered here, with everything else, but are not items.
+func walkRoot(root string, exts map[string]bool) (map[string]candidate, map[string]candidate, []string) {
 	out := make(map[string]candidate)
+	subs := make(map[string]candidate)
 	var errs []string
 
 	err := filepath.WalkDir(root, func(p string, d fs.DirEntry, err error) error {
@@ -573,7 +607,8 @@ func walkRoot(root string, exts map[string]bool) (map[string]candidate, []string
 		}
 
 		ext := strings.ToLower(filepath.Ext(name))
-		if !exts[ext] {
+		isMedia := exts[ext]
+		if !isMedia && !sidecarExts[ext] {
 			return nil
 		}
 
@@ -601,13 +636,19 @@ func walkRoot(root string, exts map[string]bool) (map[string]candidate, []string
 		if st, ok := info.Sys().(*syscall.Stat_t); ok {
 			c.inode = int64(st.Ino)
 		}
+		// A media extension wins: if a kind ever owns one of these, it is a
+		// media file first and a subtitle only by coincidence.
+		if isMedia {
 			out[c.relPath] = c
+		} else {
+			subs[c.relPath] = c
+		}
 		return nil
 	})
 	if err != nil {
 		errs = append(errs, fmt.Sprintf("%s: %v", root, err))
 	}
-	return out, errs
+	return out, subs, errs
 }
 
 func loadKnownFiles(ctx context.Context, db *sql.DB, libraryID int64) (map[string]knownFile, error) {
@@ -777,6 +818,116 @@ func upsertFile(ctx context.Context, tx *sql.Tx, itemID int64, c candidate, info
 	return nil
 }
 
+// upsertSidecars matches the subtitle files in one directory to the media file
+// they sit beside, and writes a row for each. A sidecar belongs to a video when
+// its name is the video's base name with an optional language segment between:
+// "Movie.srt" and "Movie.en.srt" both belong to "Movie.mkv", the latter with
+// language "en". The language is the whole middle segment, so "pt-BR" survives
+// intact.
+//
+// The match is a prefix guarded by the dot, not a split, because a release name
+// may itself contain dots: for "The.Matrix.1999.1080p.mkv" the sidecar is
+// "The.Matrix.1999.1080p.en.srt" and only the trailing segment is a language.
+func upsertSidecars(ctx context.Context, tx *sql.Tx, itemID int64, media candidate, sidecars []candidate, seen map[string]bool, res *ScanResult) {
+	if len(sidecars) == 0 {
+		return
+	}
+	base := strings.TrimSuffix(path.Base(media.relPath), path.Ext(media.relPath))
+
+	for _, s := range sidecars {
+		stem := strings.TrimSuffix(path.Base(s.relPath), path.Ext(s.relPath))
+		var lang string
+		switch {
+		case stem == base:
+			lang = ""
+		case strings.HasPrefix(stem, base+"."):
+			lang = stem[len(base)+1:]
+		default:
+			continue
+		}
+		seen[s.relPath] = true
+		if err := upsertSidecar(ctx, tx, itemID, s, lang); err != nil {
+			res.addError(fmt.Sprintf("%s: %v", s.relPath, err))
+		}
+	}
+}
+
+// upsertSidecar writes one sidecar track. The conflict target is the table's
+// UNIQUE (item_id, source, path, stream_index), so a rescan refreshes the row
+// instead of adding a second one. stream_index is -1, the file-backed value.
+func upsertSidecar(ctx context.Context, tx *sql.Tx, itemID int64, s candidate, lang string) error {
+	_, err := tx.ExecContext(ctx, `
+		INSERT INTO subtitles (item_id, source, path, stream_index, format, language, label)
+		VALUES (?, 'sidecar', ?, -1, ?, ?, ?)
+		ON CONFLICT (item_id, source, path, stream_index)
+		DO UPDATE SET format = excluded.format, language = excluded.language,
+		              label = excluded.label`,
+		itemID, s.relPath, s.ext, lang, path.Base(s.relPath))
+	if err != nil {
+		return fmt.Errorf("upsert sidecar: %w", err)
+	}
+	return nil
+}
+
+// reconcileSidecars deletes sidecar rows whose file is no longer on disk. It
+// reads the library's existing rows and removes only the ones no video claimed
+// this pass, so the subtitles table stays exactly a cache of what is beside the
+// media. It never touches embedded or downloaded tracks.
+func reconcileSidecars(ctx context.Context, db *sql.DB, libraryID int64, seen map[string]bool, res *ScanResult) error {
+	rows, err := db.QueryContext(ctx, `
+		SELECT s.path FROM subtitles s
+		JOIN media_items i ON i.id = s.item_id
+		WHERE i.library_id = ? AND s.source = 'sidecar'`, libraryID)
+	if err != nil {
+		return fmt.Errorf("load sidecars: %w", err)
+	}
+	var stale []string
+	for rows.Next() {
+		var p string
+		if err := rows.Scan(&p); err != nil {
+			rows.Close()
+			return err
+		}
+		if !seen[p] {
+			stale = append(stale, p)
+		}
+	}
+	if err := rows.Close(); err != nil {
+		return err
+	}
+	if err := rows.Err(); err != nil {
+		return err
+	}
+	if len(stale) == 0 {
+		return nil
+	}
+	sort.Strings(stale)
+
+	for start := 0; start < len(stale); start += scanBatchSize {
+		end := start + scanBatchSize
+		if end > len(stale) {
+			end = len(stale)
+		}
+		batch := stale[start:end]
+		err := commitBatch(ctx, db, func(tx *sql.Tx) error {
+			for _, p := range batch {
+				if _, err := tx.ExecContext(ctx, `
+					DELETE FROM subtitles
+					WHERE source = 'sidecar' AND path = ?
+					  AND item_id IN (SELECT id FROM media_items WHERE library_id = ?)`,
+					p, libraryID); err != nil {
+					res.addError(fmt.Sprintf("%s: %v", p, err))
+				}
+			}
+			return nil
+		})
+		if err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
 // probeBatch probes every new or changed video/audio file in one batch and
 // returns the facts keyed by relative path.
 //

internal/scanner/scan_test.go

@@ -794,7 +794,199 @@ func TestParseEpisodePatterns(t *testing.T) {
 	}
 }
 
-// The same tree must behave differently depending on the library's declared
+// subtitle is one subtitles row as the tests want to see it.
+type subtitle struct {
+	source string
+	path   string
+	format string
+	lang   string
+	label  string
+	stream int
+}
+
+// scanSubtitles returns every subtitle track in a library, ordered by path.
+func scanSubtitles(t *testing.T, db *sql.DB, libraryID int64) []subtitle {
+	t.Helper()
+	rows, err := db.Query(`
+		SELECT s.source, s.path, s.format, s.language, s.label, s.stream_index
+		FROM subtitles s JOIN media_items i ON i.id = s.item_id
+		WHERE i.library_id = ? ORDER BY s.path`, libraryID)
+	if err != nil {
+		t.Fatalf("query subtitles: %v", err)
+	}
+	defer rows.Close()
+
+	var out []subtitle
+	for rows.Next() {
+		var s subtitle
+		if err := rows.Scan(&s.source, &s.path, &s.format, &s.lang, &s.label, &s.stream); err != nil {
+			t.Fatalf("scan subtitle: %v", err)
+		}
+		out = append(out, s)
+	}
+	if err := rows.Err(); err != nil {
+		t.Fatalf("rows: %v", err)
+	}
+	return out
+}
+
+// A .srt beside a video is discovered as a sidecar track, with the fields the
+// table defines: the path relative to the root, the format from the extension,
+// no language, and stream_index -1.
+func TestScanSidecarSubtitle(t *testing.T) {
+	db := testDB(t)
+	ctx := context.Background()
+	root := t.TempDir()
+
+	writeFile(t, filepath.Join(root, "Movie.mkv"), "video")
+	writeFile(t, filepath.Join(root, "Movie.srt"), "subs")
+
+	lib, err := library.Create(ctx, db, "Movies", "movies", root)
+	if err != nil {
+		t.Fatalf("Create: %v", err)
+	}
+	if _, err := Scan(ctx, db, lib, testLogger()); err != nil {
+		t.Fatalf("Scan: %v", err)
+	}
+
+	got := scanSubtitles(t, db, lib.ID)
+	if len(got) != 1 {
+		t.Fatalf("got %d subtitle rows, want 1: %+v", len(got), got)
+	}
+	want := subtitle{source: "sidecar", path: "Movie.srt", format: "srt", lang: "", label: "Movie.srt", stream: -1}
+	if got[0] != want {
+		t.Errorf("subtitle = %+v, want %+v", got[0], want)
+	}
+}
+
+// The middle segment of "Movie.en.srt" / "Movie.pt-BR.srt" is the language, and
+// a release name full of dots must not be mistaken for one.
+func TestScanSidecarSubtitleLanguage(t *testing.T) {
+	db := testDB(t)
+	ctx := context.Background()
+	root := t.TempDir()
+
+	writeFile(t, filepath.Join(root, "Movie.mkv"), "video")
+	writeFile(t, filepath.Join(root, "Movie.en.srt"), "en")
+	writeFile(t, filepath.Join(root, "Movie.pt-BR.srt"), "pt")
+	writeFile(t, filepath.Join(root, "The.Matrix.1999.1080p.mkv"), "video2")
+	writeFile(t, filepath.Join(root, "The.Matrix.1999.1080p.fr.ass"), "fr")
+
+	lib, err := library.Create(ctx, db, "Movies", "movies", root)
+	if err != nil {
+		t.Fatalf("Create: %v", err)
+	}
+	if _, err := Scan(ctx, db, lib, testLogger()); err != nil {
+		t.Fatalf("Scan: %v", err)
+	}
+
+	want := []subtitle{
+		{source: "sidecar", path: "Movie.en.srt", format: "srt", lang: "en", label: "Movie.en.srt", stream: -1},
+		{source: "sidecar", path: "Movie.pt-BR.srt", format: "srt", lang: "pt-BR", label: "Movie.pt-BR.srt", stream: -1},
+		{source: "sidecar", path: "The.Matrix.1999.1080p.fr.ass", format: "ass", lang: "fr", label: "The.Matrix.1999.1080p.fr.ass", stream: -1},
+	}
+	got := scanSubtitles(t, db, lib.ID)
+	if len(got) != len(want) {
+		t.Fatalf("got %d subtitle rows, want %d: %+v", len(got), len(want), got)
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Errorf("subtitle %d = %+v, want %+v", i, got[i], want[i])
+		}
+	}
+}
+
+// A rescan must refresh the sidecar row, not add a second one for the same
+// file: the table's UNIQUE constraint is the dedupe, not a lookup.
+func TestScanSidecarRescanDoesNotDuplicate(t *testing.T) {
+	db := testDB(t)
+	ctx := context.Background()
+	root := t.TempDir()
+
+	writeFile(t, filepath.Join(root, "Movie.mkv"), "video")
+	writeFile(t, filepath.Join(root, "Movie.en.srt"), "subs")
+
+	lib, err := library.Create(ctx, db, "Movies", "movies", root)
+	if err != nil {
+		t.Fatalf("Create: %v", err)
+	}
+	for i := 0; i < 3; i++ {
+		if _, err := Scan(ctx, db, lib, testLogger()); err != nil {
+			t.Fatalf("Scan %d: %v", i, err)
+		}
+	}
+
+	if got := scanSubtitles(t, db, lib.ID); len(got) != 1 {
+		t.Fatalf("got %d subtitle rows after rescans, want 1: %+v", len(got), got)
+	}
+}
+
+// Subtitles are a cache of what is on disk: a sidecar that disappears has its
+// row deleted, while the video beside it is untouched and still not missing.
+func TestScanSidecarRemovalDeletesRow(t *testing.T) {
+	db := testDB(t)
+	ctx := context.Background()
+	root := t.TempDir()
+
+	writeFile(t, filepath.Join(root, "Movie.mkv"), "video")
+	writeFile(t, filepath.Join(root, "Movie.srt"), "subs")
+
+	lib, err := library.Create(ctx, db, "Movies", "movies", root)
+	if err != nil {
+		t.Fatalf("Create: %v", err)
+	}
+	if _, err := Scan(ctx, db, lib, testLogger()); err != nil {
+		t.Fatalf("first Scan: %v", err)
+	}
+	if got := scanSubtitles(t, db, lib.ID); len(got) != 1 {
+		t.Fatalf("first scan produced %d subtitle rows, want 1", len(got))
+	}
+
+	if err := os.Remove(filepath.Join(root, "Movie.srt")); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := Scan(ctx, db, lib, testLogger()); err != nil {
+		t.Fatalf("second Scan: %v", err)
+	}
+
+	if got := scanSubtitles(t, db, lib.ID); len(got) != 0 {
+		t.Errorf("vanished sidecar left %d rows, want 0: %+v", len(got), got)
+	}
+
+	// The video is still there and was never flagged missing.
+	var missing int
+	if err := db.QueryRowContext(ctx,
+		`SELECT missing FROM media_items WHERE library_id = ? AND path = ?`, lib.ID, "Movie.mkv").Scan(&missing); err != nil {
+		t.Fatalf("video item is gone: %v", err)
+	}
+	if missing != 0 {
+		t.Errorf("video missing = %d, want 0", missing)
+	}
+}
+
+// A sidecar never becomes an item of its own, and an unsupported file beside a
+// video is ignored entirely.
+func TestScanSidecarIsNotAnItem(t *testing.T) {
+	db := testDB(t)
+	ctx := context.Background()
+	root := t.TempDir()
+
+	writeFile(t, filepath.Join(root, "Movie.mkv"), "video")
+	writeFile(t, filepath.Join(root, "Movie.srt"), "subs")
+
+	lib, err := library.Create(ctx, db, "Movies", "movies", root)
+	if err != nil {
+		t.Fatalf("Create: %v", err)
+	}
+	if _, err := Scan(ctx, db, lib, testLogger()); err != nil {
+		t.Fatalf("Scan: %v", err)
+	}
+
+	if n, _ := library.CountItems(ctx, db, lib.ID); n != 1 {
+		t.Errorf("item count = %d, want 1 (the video only)", n)
+	}
+}
+
 // grouping. That is the whole point of separating the video kinds: a film shelf
 // must not try to parse episode numbers, and a TV library must not scatter its
 // episodes across the root.

internal/stream/stream.go

@@ -49,6 +49,9 @@ var contentTypes = map[string]string{
 	".webm": "video/webm",
 	".mov":  "video/quicktime",
 	".avi":  "video/x-msvideo",
+	// Subtitles. The server converts every other format to this one, because
+	// it is the only subtitle format browsers play.
+	".vtt": "text/vtt; charset=utf-8",
 	// Images (comic pages, covers, photos).
 	".jpg":  "image/jpeg",
 	".jpeg": "image/jpeg",