feat(backups): 主人和管理员可下载单个备份、导出停服世界，字节经一次性票据从导出 Job 流式转给浏览器，备份按 sha256 核对

internal/api/api.go

@@ -98,6 +98,12 @@ type API struct {
 	FileStage       *fileedit.Stage
 	InternalBaseURL string
 
+	// Exporter starts the Job behind a world or backup download (exports.go),
+	// which PUTs the archive to InternalBaseURL. Optional like Restorer: the
+	// export routes report 503 unless both are set, after the owner-or-admin,
+	// backup and stopped gates.
+	Exporter Exporter
+
 	// Schedules stores the servers' scheduled tasks (schedules.go), which
 	// RunSchedules fires. Optional: when nil the schedule routes report 503 and
 	// RunSchedules does nothing.
@@ -222,6 +228,9 @@ type API struct {
 	streamCapOnce sync.Once
 	streamCap     *streamLimiter
 
+	exportsOnce sync.Once
+	exports     *exportRegistry
+
 	authDoorOnce    sync.Once
 	authDoorBuckets *bucketSet
 	mailOnce        sync.Once
@@ -459,6 +468,13 @@ func (a *API) internalAPIRoutes() []apiRoute {
 		// because that Job holds no service token; the one-time bearer token minted
 		// with the upload is the check (handlers_files.go).
 		{Method: "GET", Pattern: "/api/v1/internal/file-uploads/{id}", Public: true, h: a.handleInternalFileUpload},
+
+		// An export Job's archive, held open until the owner's browser downloads
+		// it. Public for the same reason as file uploads: the Job holds no service
+		// token, and the one-time bearer token minted with the export is the check
+		// (exports.go). The body is read at the browser's pace, so the handler
+		// lifts the minimum-rate body deadline and applies its own stall bound.
+		{Method: "PUT", Pattern: "/api/v1/internal/exports/{id}", Public: true, h: a.handleInternalExportUpload},
 	}
 }
 
@@ -558,6 +574,14 @@ func (a *API) externalAPIRoutes() []apiRoute {
 		{Method: "GET", Pattern: "/api/v1/servers/{name}/jobs", h: a.handleServerJobs},
 		{Method: "POST", Pattern: "/api/v1/servers/{name}/restore-backup", h: a.handleRestoreBackup},
 		{Method: "POST", Pattern: "/api/v1/servers/{name}/backup", h: a.handleBackupNow},
+		// World export (exports.go): download a backup, or a stopped server's world
+		// as it is now, straight to the browser. App-tier with the restore gate
+		// inside each start route; the ticket routes answer only the user who
+		// started the export.
+		{Method: "POST", Pattern: "/api/v1/servers/{name}/backups/{id}/export", h: a.handleExportBackup},
+		{Method: "POST", Pattern: "/api/v1/servers/{name}/world/export", h: a.handleExportWorld},
+		{Method: "GET", Pattern: "/api/v1/exports/{ticket}", h: a.handleExportStatus},
+		{Method: "GET", Pattern: "/api/v1/exports/{ticket}/download", h: a.handleExportDownload},
 		// Server file manager: list, read, write, make a folder, delete, rename and
 		// upload in a STOPPED server's world volume (handlers_files.go). App-tier,
 		// exactly like the backup pair above and for the same reason — every route

internal/api/api_test.go

@@ -255,8 +255,9 @@ type fakeSession struct {
 // fakeBackup mirrors a world_backups row: the client-facing view plus the
 // server-side backup_ref the list queries never expose.
 type fakeBackup struct {
-	view BackupView
-	ref  string
+	view   BackupView
+	ref    string
+	sha256 string
 }
 
 // fakeLinkCode mirrors an account_link_codes row.
@@ -1051,7 +1052,7 @@ func (f *fakeRepo) BackupByID(_ context.Context, id string) (*BackupRecord, erro
 			return &BackupRecord{
 				ID: b.view.ID, ServerName: b.view.ServerName,
 				FormerOwner: b.view.FormerOwner, BackupRef: b.ref,
-				SizeBytes: b.view.SizeBytes, Corrupt: b.view.Corrupt,
+				SizeBytes: b.view.SizeBytes, Corrupt: b.view.Corrupt, SHA256: b.sha256,
 			}, nil
 		}
 	}

internal/api/exports.go

@@ -0,0 +1,677 @@
+package api
+
+import (
+	"context"
+	"crypto/rand"
+	"crypto/sha256"
+	"crypto/subtle"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"hash"
+	"io"
+	"log"
+	"mime"
+	"net/http"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+
+	"felis.lolicon.best/internal/apis/felis/v1alpha1"
+	"felis.lolicon.best/internal/maintenance"
+	"felis.lolicon.best/internal/naming"
+	"felis.lolicon.best/internal/worldexport"
+)
+
+// World export. The owner downloads a tar.gz of their world, either as it is
+// now (the server stopped) or as one of its backups, straight into the browser:
+//
+//  1. POST /servers/{name}/world/export or /servers/{name}/backups/{id}/export
+//     checks the caller and the server, admits the export against the limits
+//     below and starts a one-shot felis-export Job (internal/worldexport) that
+//     reads the world or the archive read-only. It answers 202 with a ticket:
+//     256 random bits, good for the caller who started it and nobody else.
+//  2. The Job PUTs the archive to the internal face (PUT
+//     /api/v1/internal/exports/{id} with the one-time token it was started
+//     with), and that request waits, body unread, for the browser.
+//  3. The panel polls GET /exports/{ticket} until it reads ready, then points a
+//     hidden <a download> at GET /exports/{ticket}/download, which claims the
+//     waiting PUT and copies its body into the response 32 KiB at a time. The
+//     archive passes through felis-api's memory once and never touches a disk
+//     it owns, and the Job moves at the browser's pace.
+//
+// A backup is checked against the sha256 recorded when it was written as it
+// streams, and the last read is held back until the digest is known: a mismatch
+// aborts the response, so the browser reports a failed download and never keeps
+// a complete-looking corrupt file, and the Job is told backup_corrupt.
+//
+// Tickets live in felis-api's memory. A restart forgets them, and a Job that
+// then PUTs finds nothing and fails, which the jobs list shows.
+
+// Exporter starts the Job that archives a world or a backup and hands it to the
+// internal upload route (internal/worldexport). Optional: when nil the export
+// routes answer 503.
+type Exporter interface {
+	Start(ctx context.Context, r worldexport.Request) (job string, err error)
+}
+
+// Limits on exports. Each one keeps a Job, a connection and a 64 KiB copy
+// buffer alive for as long as a download takes, and a world export also keeps
+// its server from starting.
+const (
+	exportMaxActive  = 2 // admitted and not yet over, install-wide
+	exportMaxPerUser = 1
+	exportPerHour    = 6 // started by one user in any hour
+	exportCopyBuffer = 32 << 10
+)
+
+// Timings. Vars only so a test can shrink them.
+var (
+	// exportClaimTTL is how long the Job's upload waits for the browser.
+	exportClaimTTL = 90 * time.Second
+	// exportPendingTTL is how long an export may take to reach ready: the Job
+	// is scheduled, pulls its image and connects well inside it.
+	exportPendingTTL = 10 * time.Minute
+	// exportStall is the longest either end may go without moving a byte.
+	exportStall = 2 * time.Minute
+	// exportKeepSpent is how long a finished ticket still answers 410
+	// export_expired before it reads as unknown.
+	exportKeepSpent = 10 * time.Minute
+)
+
+// States of an export.
+const (
+	exportPending   = "pending"   // the Job has not connected yet
+	exportReady     = "ready"     // its upload is waiting for the browser
+	exportStreaming = "streaming" // the browser is downloading
+	exportFailed    = "failed"    // the Job died before it connected
+	exportSpent     = "spent"     // downloaded, or given up
+)
+
+// exportTicketView answers both export routes.
+type exportTicketView struct {
+	Ticket   string `json:"ticket"`
+	State    string `json:"state"`
+	Filename string `json:"filename"`
+}
+
+// exportStatusView answers GET /exports/{ticket}.
+type exportStatusView struct {
+	State   string `json:"state"`
+	Message string `json:"message,omitempty"`
+}
+
+var errExportDigest = errors.New("the archive does not match the sha256 recorded when it was written")
+
+func errExportExpired() error {
+	return newError(http.StatusGone, "export_expired", "this export has expired or was already downloaded; start a new one")
+}
+
+func errNoExport() error {
+	return newError(http.StatusNotFound, "not_found", "no such export")
+}
+
+type exportEntry struct {
+	ticket    string
+	id        string
+	tokenHash [32]byte
+	userID    string
+	server    string
+	mode      string
+	sha256    string // what a backup must hash to; empty for a world
+	filename  string
+	job       string
+	state     string
+	message   string
+	at        time.Time // when it entered its state
+	upload    *exportUpload
+}
+
+// exportUpload is the Job's PUT, parked until a browser claims it.
+type exportUpload struct {
+	body    io.Reader
+	size    int64 // -1 when the Job streams it chunked
+	claimed chan struct{}
+	done    chan error // how the download ended; buffered
+}
+
+func (e *exportEntry) active() bool {
+	return e.state == exportPending || e.state == exportReady || e.state == exportStreaming
+}
+
+type exportRegistry struct {
+	mu       sync.Mutex
+	byTicket map[string]*exportEntry
+	byID     map[string]*exportEntry
+	starts   map[string][]time.Time // per user, oldest first
+}
+
+func (a *API) exportTickets() *exportRegistry {
+	a.exportsOnce.Do(func() {
+		a.exports = &exportRegistry{
+			byTicket: map[string]*exportEntry{},
+			byID:     map[string]*exportEntry{},
+			starts:   map[string][]time.Time{},
+		}
+	})
+	return a.exports
+}
+
+// sweepLocked expires what has waited too long and forgets what ended long ago.
+func (g *exportRegistry) sweepLocked(now time.Time) {
+	for t, e := range g.byTicket {
+		switch {
+		case e.state == exportPending && now.Sub(e.at) >= exportPendingTTL:
+			e.state, e.at = exportSpent, now
+		case !e.active() && now.Sub(e.at) >= exportKeepSpent:
+			delete(g.byTicket, t)
+			delete(g.byID, e.id)
+		}
+	}
+	for u, ts := range g.starts {
+		for len(ts) > 0 && now.Sub(ts[0]) >= time.Hour {
+			ts = ts[1:]
+		}
+		if len(ts) == 0 {
+			delete(g.starts, u)
+		} else {
+			g.starts[u] = ts
+		}
+	}
+}
+
+func randomHex(n int) string {
+	b := make([]byte, n)
+	_, _ = rand.Read(b)
+	return hex.EncodeToString(b)
+}
+
+// admit reserves the export e describes (its user, server, mode, filename and
+// digest) and returns its upload token, or refuses it with export_busy.
+func (g *exportRegistry) admit(e *exportEntry, now time.Time) (string, error) {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	g.sweepLocked(now)
+	active, mine := 0, 0
+	for _, o := range g.byTicket {
+		if o.active() {
+			active++
+			if o.userID == e.userID {
+				mine++
+			}
+		}
+	}
+	switch starts := g.starts[e.userID]; {
+	case mine >= exportMaxPerUser:
+		return "", newError(http.StatusTooManyRequests, "export_busy",
+			"you already have an export in progress; download it or let it expire first").retryAfter(exportClaimTTL)
+	case active >= exportMaxActive:
+		return "", newError(http.StatusTooManyRequests, "export_busy",
+			"%d exports are already in progress; retry in a few minutes", exportMaxActive).retryAfter(time.Minute)
+	case len(starts) >= exportPerHour:
+		return "", newError(http.StatusTooManyRequests, "export_busy",
+			"you have started %d exports in the last hour; retry later", exportPerHour).retryAfter(starts[0].Add(time.Hour).Sub(now))
+	}
+	token := randomHex(32)
+	e.ticket, e.id, e.tokenHash = randomHex(32), randomHex(8), sha256.Sum256([]byte(token))
+	e.state, e.at = exportPending, now
+	g.byTicket[e.ticket] = e
+	g.byID[e.id] = e
+	g.starts[e.userID] = append(g.starts[e.userID], now)
+	return token, nil
+}
+
+// drop forgets an export that never got its Job, and gives its start back.
+func (g *exportRegistry) drop(e *exportEntry) {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	delete(g.byTicket, e.ticket)
+	delete(g.byID, e.id)
+	ts := g.starts[e.userID]
+	for i := len(ts) - 1; i >= 0; i-- {
+		if ts[i].Equal(e.at) {
+			g.starts[e.userID] = append(ts[:i:i], ts[i+1:]...)
+			break
+		}
+	}
+}
+
+func (g *exportRegistry) started(e *exportEntry, job string) {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	e.job = job
+}
+
+// lookup returns a copy of userID's export by its ticket. Another user's reads
+// as unknown, and a spent one as export_expired.
+func (g *exportRegistry) lookup(ticket, userID string, now time.Time) (exportEntry, error) {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	g.sweepLocked(now)
+	e, ok := g.byTicket[ticket]
+	switch {
+	case !ok || e.userID != userID:
+		return exportEntry{}, errNoExport()
+	case e.state == exportSpent || e.state == exportStreaming:
+		return exportEntry{}, errExportExpired()
+	}
+	return *e, nil
+}
+
+// fail records that the export's Job died before it connected, and reports
+// false when the export has moved on since.
+func (g *exportRegistry) fail(ticket, message string, now time.Time) bool {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	e, ok := g.byTicket[ticket]
+	if !ok || e.state != exportPending {
+		return false
+	}
+	e.state, e.message, e.at = exportFailed, message, now
+	return true
+}
+
+// arrive parks the Job's upload on the export the id and token open. Unknown
+// ids, wrong tokens and exports past pending all read the same.
+func (g *exportRegistry) arrive(id, token string, up *exportUpload, now time.Time) bool {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	g.sweepLocked(now)
+	e, ok := g.byID[id]
+	if !ok {
+		return false
+	}
+	sum := sha256.Sum256([]byte(token))
+	if subtle.ConstantTimeCompare(sum[:], e.tokenHash[:]) != 1 || e.state != exportPending {
+		return false
+	}
+	e.state, e.at, e.upload = exportReady, now, up
+	return true
+}
+
+// giveUp spends an export whose upload nobody claimed, and reports false when a
+// claim got there first.
+func (g *exportRegistry) giveUp(up *exportUpload, now time.Time) bool {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	for _, e := range g.byID {
+		if e.upload == up && e.state == exportReady {
+			e.state, e.at = exportSpent, now
+			return true
+		}
+	}
+	return false
+}
+
+// claim hands userID's ready export to its download: from here on the ticket is
+// spent whatever happens to the download.
+func (g *exportRegistry) claim(ticket, userID string, now time.Time) (exportEntry, error) {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	g.sweepLocked(now)
+	e, ok := g.byTicket[ticket]
+	switch {
+	case !ok || e.userID != userID:
+		return exportEntry{}, errNoExport()
+	case e.state == exportPending:
+		return exportEntry{}, newError(http.StatusConflict, "export_not_ready",
+			"the export is still being prepared; wait until its status reads ready")
+	case e.state != exportReady:
+		return exportEntry{}, errExportExpired()
+	}
+	e.state, e.at = exportStreaming, now
+	close(e.upload.claimed)
+	return *e, nil
+}
+
+func (g *exportRegistry) finish(ticket string, now time.Time) {
+	g.mu.Lock()
+	defer g.mu.Unlock()
+	if e, ok := g.byTicket[ticket]; ok {
+		e.state, e.at = exportSpent, now
+	}
+}
+
+// handleExportBackup starts the download of one backup (POST
+// /api/v1/servers/{name}/backups/{id}/export). The gate is restore's and
+// delete's together:
+//
+//	① name validation; ② an unknown server is 404; ③ owner-or-admin, else 403
+//	④ the backup must exist, and a non-admin must be its former owner: any
+//	   other id reads as unknown (404 no_backup), as it does in their list
+//	⑤ cross-server guard: the backup must be this server's (403)
+//	⑥ an archive that failed its read-back is 409 backup_corrupt
+//
+// It needs no stopped server and takes no world lock: the Job reads the backup
+// store only.
+func (a *API) handleExportBackup(w http.ResponseWriter, r *http.Request) {
+	p := principalFromContext(r.Context())
+	name, rec, ok := a.exportGate(w, r)
+	if !ok {
+		return
+	}
+	backup, err := a.Repo.BackupByID(r.Context(), r.PathValue("id"))
+	if err == nil && !p.IsAdmin() && backup.FormerOwner != p.UserID {
+		err = ErrNotFound
+	}
+	if errors.Is(err, ErrNotFound) {
+		writeError(w, r, newError(http.StatusNotFound, "no_backup", "no matching backup exists"))
+		return
+	}
+	if err != nil {
+		writeError(w, r, err)
+		return
+	}
+	if backup.ServerName != name {
+		writeError(w, r, errForbidden)
+		return
+	}
+	if backup.Corrupt {
+		writeError(w, r, newError(http.StatusConflict, "backup_corrupt",
+			"this backup did not read back intact and cannot be exported; pick another"))
+		return
+	}
+	if a.Exporter == nil || a.InternalBaseURL == "" {
+		writeError(w, r, errExportUnavailable())
+		return
+	}
+	e := &exportEntry{userID: p.UserID, server: name, mode: worldexport.ModeBackup,
+		filename: fmt.Sprintf("%s-backup-%s.tar.gz", name, backup.ID), sha256: backup.SHA256}
+	token, err := a.exportTickets().admit(e, a.now())
+	if err != nil {
+		writeError(w, r, err)
+		return
+	}
+	if !a.startExport(w, r, e, token, backup.BackupRef) {
+		return
+	}
+	a.auditEntry(r, AuditEntry{Actor: auditActor(p), ActorUserID: p.UserID, Action: "backup.export", ServerName: rec.Name,
+		Payload: auditPayload(map[string]any{"backup_id": backup.ID, "size_bytes": backup.SizeBytes})})
+	writeJSON(w, http.StatusAccepted, exportTicketView{Ticket: e.ticket, State: exportPending, Filename: e.filename})
+}
+
+// handleExportWorld starts the download of a stopped server's world as it is
+// now (POST /api/v1/servers/{name}/world/export). The gate is a backup's: owner
+// or admin, the server fully stopped, a world volume to read, and the
+// world-volume lock, which the export Job then holds (as KindExport) until the
+// download ends, so the server cannot start under it and tear the archive.
+func (a *API) handleExportWorld(w http.ResponseWriter, r *http.Request) {
+	p := principalFromContext(r.Context())
+	name, rec, ok := a.exportGate(w, r)
+	if !ok {
+		return
+	}
+	const notStopped = "stop the server before exporting its world"
+	info, err := a.Cluster.GetServer(r.Context(), name)
+	if err != nil {
+		a.writeLookupError(w, r, err)
+		return
+	}
+	if info.Ready || info.DesiredState != string(v1alpha1.DesiredStopped) {
+		writeError(w, r, newError(http.StatusConflict, "not_stopped", notStopped))
+		return
+	}
+	if exists, err := a.Cluster.WorldVolumeExists(r.Context(), name); err != nil {
+		writeError(w, r, err)
+		return
+	} else if !exists {
+		writeError(w, r, errNoWorldVolume())
+		return
+	}
+	if a.Exporter == nil || a.InternalBaseURL == "" {
+		writeError(w, r, errExportUnavailable())
+		return
+	}
+	reg := a.exportTickets()
+	e := &exportEntry{userID: p.UserID, server: name, mode: worldexport.ModeWorld,
+		filename: fmt.Sprintf("%s-world-%s.tar.gz", name, a.now().UTC().Format("20060102-150405"))}
+	token, err := reg.admit(e, a.now())
+	if err != nil {
+		writeError(w, r, err)
+		return
+	}
+	release, ok := a.acquireWorld(w, r, name, maintenance.KindExport, notStopped)
+	if !ok {
+		reg.drop(e)
+		return
+	}
+	defer release()
+	if !a.startExport(w, r, e, token, "") {
+		return
+	}
+	a.auditEntry(r, AuditEntry{Actor: auditActor(p), ActorUserID: p.UserID, Action: "world.export", ServerName: rec.Name})
+	writeJSON(w, http.StatusAccepted, exportTicketView{Ticket: e.ticket, State: exportPending, Filename: e.filename})
+}
+
+func errExportUnavailable() error {
+	return newError(http.StatusServiceUnavailable, "export_unavailable", "world export is not configured")
+}
+
+// exportGate is the front half both export routes share: a valid name, a known
+// server, and a caller who owns it or is an admin. The ticket is bound to the
+// caller's account, so a principal without one cannot start an export.
+func (a *API) exportGate(w http.ResponseWriter, r *http.Request) (string, *ServerRecord, bool) {
+	p := principalFromContext(r.Context())
+	name := r.PathValue("name")
+	if err := naming.ValidateServerName(name); err != nil {
+		writeError(w, r, newError(http.StatusBadRequest, "bad_name", "invalid server name: %v", err))
+		return "", nil, false
+	}
+	rec, err := a.Repo.ServerByName(r.Context(), name)
+	if err != nil {
+		a.writeLookupError(w, r, err)
+		return "", nil, false
+	}
+	if p.UserID == "" || !a.isOwnerOrAdmin(p, rec) {
+		writeError(w, r, errForbidden)
+		return "", nil, false
+	}
+	return name, rec, true
+}
+
+// startExport creates the admitted export's Job, or forgets the export and
+// writes the error.
+func (a *API) startExport(w http.ResponseWriter, r *http.Request, e *exportEntry, token, backupRef string) bool {
+	job, err := a.Exporter.Start(r.Context(), worldexport.Request{
+		Server: e.server, Mode: e.mode, BackupRef: backupRef, ID: e.id, Token: token,
+		TargetURL: a.InternalBaseURL + "/api/v1/internal/exports/" + e.id,
+	})
+	if err != nil {
+		a.exportTickets().drop(e)
+		writeError(w, r, err)
+		return false
+	}
+	a.exportTickets().started(e, job)
+	return true
+}
+
+// handleExportStatus answers GET /api/v1/exports/{ticket}: pending while the
+// Job gets going, ready once its upload waits for the browser, failed (with the
+// Job's error) when it died first. A downloaded or expired ticket is 410
+// export_expired; an unknown one, or another user's, is 404.
+func (a *API) handleExportStatus(w http.ResponseWriter, r *http.Request) {
+	p := principalFromContext(r.Context())
+	reg := a.exportTickets()
+	e, err := reg.lookup(r.PathValue("ticket"), p.UserID, a.now())
+	if err != nil {
+		writeError(w, r, err)
+		return
+	}
+	// A Job list that fails leaves the export pending: the next poll asks again,
+	// and exportPendingTTL ends it either way.
+	if e.state == exportPending && e.job != "" && a.JobStatus != nil {
+		jobs, _ := a.JobStatus.LatestJobs(r.Context(), e.server)
+		for _, j := range jobs {
+			if j.Name == e.job && j.State == "failed" && reg.fail(e.ticket, j.Message, a.now()) {
+				e.state, e.message = exportFailed, j.Message
+			}
+		}
+	}
+	writeJSON(w, http.StatusOK, exportStatusView{State: e.state, Message: e.message})
+}
+
+// handleExportDownload streams a ready export to its browser (GET
+// /api/v1/exports/{ticket}/download). The claim spends the ticket before the
+// first byte moves, so a second tab, a retry or a HEAD probe never gets a
+// second copy; a HEAD is refused outright, since claiming on it would spend the
+// ticket on a response with no body.
+func (a *API) handleExportDownload(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodGet {
+		w.Header().Set("Allow", http.MethodGet)
+		writeError(w, r, newError(http.StatusMethodNotAllowed, "method_not_allowed",
+			"%s is not allowed here; use GET", r.Method))
+		return
+	}
+	p := principalFromContext(r.Context())
+	reg := a.exportTickets()
+	e, err := reg.claim(r.PathValue("ticket"), p.UserID, a.now())
+	if err != nil {
+		writeError(w, r, err)
+		return
+	}
+	defer reg.finish(e.ticket, a.now())
+
+	h := w.Header()
+	h.Set("Content-Type", "application/gzip")
+	if cd := mime.FormatMediaType("attachment", map[string]string{"filename": e.filename}); cd != "" {
+		h.Set("Content-Disposition", cd)
+	} else {
+		h.Set("Content-Disposition", "attachment")
+	}
+	h.Set("X-Content-Type-Options", "nosniff")
+	h.Set("Cache-Control", "no-store")
+	if e.upload.size >= 0 {
+		h.Set("Content-Length", strconv.FormatInt(e.upload.size, 10))
+	}
+	w.WriteHeader(http.StatusOK)
+
+	err = copyExport(w, e.upload.body, e.sha256)
+	e.upload.done <- err
+	if err != nil {
+		log.Printf("api: export %s of %s ended early: %v", e.id, e.server, err)
+		// Abort the response rather than end it: a download cut short must
+		// read as failed in the browser, never as a complete file.
+		panic(http.ErrAbortHandler)
+	}
+}
+
+// copyExport copies body into w through a fixed 32 KiB buffer, restarting w's
+// write deadline on every write. With want set, the last read is held back
+// until the whole body hashes to it.
+func copyExport(w http.ResponseWriter, body io.Reader, want string) error {
+	out := &heldWriter{w: w, rc: http.NewResponseController(w)}
+	var sum hash.Hash
+	if want != "" {
+		sum = sha256.New()
+		body = io.TeeReader(body, sum)
+	}
+	if _, err := io.CopyBuffer(out, body, make([]byte, exportCopyBuffer)); err != nil {
+		return err
+	}
+	if sum != nil && !strings.EqualFold(hex.EncodeToString(sum.Sum(nil)), want) {
+		return errExportDigest
+	}
+	if err := out.flush(); err != nil {
+		return err
+	}
+	_ = out.rc.SetWriteDeadline(time.Time{}) // the connection may serve another request
+	return nil
+}
+
+// heldWriter passes each write on one behind, keeping the latest back until
+// flush, so the end of an archive reaches the browser only once it is checked.
+// It has no ReadFrom, so io.CopyBuffer uses the buffer it is given.
+type heldWriter struct {
+	w    io.Writer
+	rc   *http.ResponseController
+	held []byte
+}
+
+func (h *heldWriter) Write(p []byte) (int, error) {
+	if err := h.flush(); err != nil {
+		return 0, err
+	}
+	h.held = append(h.held[:0], p...)
+	return len(p), nil
+}
+
+func (h *heldWriter) flush() error {
+	if len(h.held) == 0 {
+		return nil
+	}
+	_ = h.rc.SetWriteDeadline(time.Now().Add(exportStall))
+	_, err := h.w.Write(h.held)
+	h.held = h.held[:0]
+	return err
+}
+
+// handleInternalExportUpload takes an export Job's archive (PUT
+// /api/v1/internal/exports/{id}). The one-time bearer token minted with the
+// export is the check, as for file uploads: an unknown id, a wrong token and a
+// token already used all read the same 404. The request then waits for the
+// browser, up to exportClaimTTL, and answers once the download has ended: 204
+// when it got the whole archive, 409 backup_corrupt when the archive did not
+// match its recorded digest, 410 export_expired when nobody came for it or the
+// browser left early.
+func (a *API) handleInternalExportUpload(w http.ResponseWriter, r *http.Request) {
+	token, ok := strings.CutPrefix(r.Header.Get("Authorization"), "Bearer ")
+	if !ok {
+		writeError(w, r, errNoExport())
+		return
+	}
+	rc := takeBodyDeadline(w, r)
+	up := &exportUpload{
+		body: &stallBody{r: r.Body, rc: rc}, size: r.ContentLength,
+		claimed: make(chan struct{}), done: make(chan error, 1),
+	}
+	reg := a.exportTickets()
+	if !reg.arrive(r.PathValue("id"), token, up, a.now()) {
+		writeError(w, r, errNoExport())
+		return
+	}
+	timer := time.NewTimer(exportClaimTTL)
+	defer timer.Stop()
+	select {
+	case <-up.claimed:
+	case <-timer.C:
+		if reg.giveUp(up, a.now()) {
+			writeError(w, r, errExportExpired())
+			return
+		}
+	case <-r.Context().Done():
+		if reg.giveUp(up, a.now()) {
+			return
+		}
+	}
+	switch err := <-up.done; {
+	case err == nil:
+		w.WriteHeader(http.StatusNoContent)
+	case errors.Is(err, errExportDigest):
+		writeError(w, r, newError(http.StatusConflict, "backup_corrupt", "%v", err))
+	default:
+		writeError(w, r, newError(http.StatusGone, "export_expired", "the download ended before the archive did: %v", err))
+	}
+}
+
+// stallBody restarts the connection's read deadline on every read, so reading
+// fails only once no byte has come for exportStall, and clears it when the body
+// ends, as deadlineBody does.
+type stallBody struct {
+	r    io.Reader
+	rc   *http.ResponseController
+	done bool
+}
+
+func (b *stallBody) Read(p []byte) (int, error) {
+	if b.done {
+		return b.r.Read(p)
+	}
+	_ = b.rc.SetReadDeadline(time.Now().Add(exportStall))
+	n, err := b.r.Read(p)
+	if err != nil {
+		b.done = true
+		_ = b.rc.SetReadDeadline(time.Time{})
+	}
+	return n, err
+}

internal/api/exports_test.go

@@ -0,0 +1,921 @@
+package api
+
+import (
+	"bytes"
+	"context"
+	"crypto/rand"
+	"crypto/sha256"
+	"encoding/hex"
+	"encoding/json"
+	"errors"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"regexp"
+	"strconv"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"felis.lolicon.best/internal/apis/felis/v1alpha1"
+	"felis.lolicon.best/internal/maintenance"
+	"felis.lolicon.best/internal/worldexport"
+	batchv1 "k8s.io/api/batch/v1"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/runtime"
+	clientgoscheme "k8s.io/client-go/kubernetes/scheme"
+	"sigs.k8s.io/controller-runtime/pkg/client/fake"
+)
+
+type fakeExporter struct {
+	reqs []worldexport.Request
+	err  error
+}
+
+func (f *fakeExporter) Start(_ context.Context, r worldexport.Request) (string, error) {
+	f.reqs = append(f.reqs, r)
+	if f.err != nil {
+		return "", f.err
+	}
+	return worldexport.JobName(r.Server, r.ID), nil
+}
+
+// asUser authenticates each request as the principal its X-Test-User header
+// names, so one handler serves several users.
+type asUser map[string]*Principal
+
+func (u asUser) Authenticate(r *http.Request) (*Principal, error) {
+	if p, ok := u[r.Header.Get("X-Test-User")]; ok {
+		return p, nil
+	}
+	return nil, errUnauthorized
+}
+
+var exportUsers = asUser{
+	"owner1":   {UserID: "owner1", Email: "[email protected]", Role: "user"},
+	"owner3":   {UserID: "owner3", Email: "[email protected]", Role: "user"},
+	"newowner": {UserID: "newowner", Email: "[email protected]", Role: "user"},
+	"stranger": {UserID: "stranger", Email: "[email protected]", Role: "user"},
+	"admin1":   {UserID: "admin1", Email: "[email protected]", Role: "admin", ViaAdminAccess: true},
+	"nouser":   {Email: "[email protected]", Role: "admin", ViaAdminAccess: true},
+}
+
+func as(user string) map[string]string { return map[string]string{"X-Test-User": user} }
+
+const (
+	exportBase  = "http://felis-api-internal.felis.svc:8081"
+	worldPath   = "/api/v1/servers/survival/world/export"
+	backupPath  = "/api/v1/servers/survival/backups/bk1/export"
+	exportStamp = "20231114-221320" // newTestAPI's clock, UTC
+)
+
+var (
+	hex64 = regexp.MustCompile(`^[0-9a-f]{64}$`)
+	hex16 = regexp.MustCompile(`^[0-9a-f]{16}$`)
+)
+
+// exportFixture: owner1 owns survival (stopped) and creative, owner3 owns gamma
+// (stopped); bk1 is survival's backup, bk2 creative's, both formerly owner1's.
+func exportFixture() (*API, *fakeRepo, *fakeCluster, *fakeExporter) {
+	repo := newFakeRepo()
+	for name, owner := range map[string]string{"survival": "owner1", "creative": "owner1", "gamma": "owner3"} {
+		repo.byName[name] = &ServerRecord{Name: name, OwnerID: owner}
+	}
+	repo.backups = []fakeBackup{
+		{view: BackupView{ID: "bk1", ServerName: "survival", FormerOwner: "owner1", Status: "present", SizeBytes: 1024},
+			ref: "/backups/survival-bk1.tar.gz"},
+		{view: BackupView{ID: "bk2", ServerName: "creative", FormerOwner: "owner1", Status: "present", SizeBytes: 2048},
+			ref: "/backups/creative-bk2.tar.gz"},
+	}
+	cl := newFakeCluster()
+	for _, name := range []string{"survival", "gamma"} {
+		cl.byName[name] = &ServerInfo{Name: name, Phase: "Stopped", DesiredState: string(v1alpha1.DesiredStopped)}
+	}
+	ex := &fakeExporter{}
+	a := newTestAPI(repo, cl)
+	a.External = exportUsers
+	a.Exporter, a.InternalBaseURL = ex, exportBase
+	return a, repo, cl, ex
+}
+
+func beginExport(t *testing.T, h http.Handler, path, user string) exportTicketView {
+	t.Helper()
+	w := do(h, "POST", path, "", as(user))
+	if w.Code != http.StatusAccepted {
+		t.Fatalf("POST %s as %s = %d, want 202 (%s)", path, user, w.Code, w.Body.String())
+	}
+	var v exportTicketView
+	if err := json.Unmarshal(w.Body.Bytes(), &v); err != nil {
+		t.Fatalf("ticket not JSON: %v (%s)", err, w.Body.String())
+	}
+	return v
+}
+
+func exportState(t *testing.T, h http.Handler, ticket, user string) (int, exportStatusView) {
+	t.Helper()
+	w := do(h, "GET", "/api/v1/exports/"+ticket, "", as(user))
+	var v exportStatusView
+	if w.Code == http.StatusOK {
+		if err := json.Unmarshal(w.Body.Bytes(), &v); err != nil {
+			t.Fatalf("status not JSON: %v (%s)", err, w.Body.String())
+		}
+	} else {
+		v.State = decodeErr(t, w)
+	}
+	return w.Code, v
+}
+
+func waitExportReady(t *testing.T, h http.Handler, ticket, user string) {
+	t.Helper()
+	for deadline := time.Now().Add(5 * time.Second); ; {
+		if code, v := exportState(t, h, ticket, user); code == http.StatusOK && v.State == exportReady {
+			return
+		}
+		if time.Now().After(deadline) {
+			t.Fatal("the export never read ready")
+		}
+		time.Sleep(2 * time.Millisecond)
+	}
+}
+
+// uploadExport serves the Job's PUT on the internal face in the background.
+func uploadExport(h http.Handler, id, token string, body io.Reader) <-chan *httptest.ResponseRecorder {
+	out := make(chan *httptest.ResponseRecorder, 1)
+	go func() {
+		r := httptest.NewRequest("PUT", "/api/v1/internal/exports/"+id, body)
+		r.Header.Set("Authorization", "Bearer "+token)
+		w := httptest.NewRecorder()
+		h.ServeHTTP(w, r)
+		out <- w
+	}()
+	return out
+}
+
+func awaitUpload(t *testing.T, up <-chan *httptest.ResponseRecorder) *httptest.ResponseRecorder {
+	t.Helper()
+	select {
+	case w := <-up:
+		return w
+	case <-time.After(5 * time.Second):
+		t.Fatal("the upload never answered")
+		return nil
+	}
+}
+
+// doSoon is do for a request that must be refused outright. A download that
+// claimed the export by mistake would block on the parked body, and an upload
+// let in by mistake would wait for a browser, so either fails the test after a
+// bound instead of hanging it.
+func doSoon(t *testing.T, h http.Handler, method, path, body string, headers map[string]string) *httptest.ResponseRecorder {
+	t.Helper()
+	out := make(chan *httptest.ResponseRecorder, 1)
+	go func() { out <- do(h, method, path, body, headers) }()
+	select {
+	case w := <-out:
+		return w
+	case <-time.After(2 * time.Second):
+		t.Fatalf("%s %s never answered: it claimed the export", method, path)
+		return nil
+	}
+}
+
+func randomBytes(n int) []byte {
+	b := make([]byte, n)
+	_, _ = rand.Read(b)
+	return b
+}
+
+func TestExportBackupGate(t *testing.T) {
+	t.Run("former owner starts a backup export", func(t *testing.T) {
+		a, repo, cl, ex := exportFixture()
+		v := beginExport(t, a.ExternalHandler(), backupPath, "owner1")
+		if !hex64.MatchString(v.Ticket) || v.State != "pending" || v.Filename != "survival-backup-bk1.tar.gz" {
+			t.Fatalf("ticket = %+v", v)
+		}
+		if len(ex.reqs) != 1 {
+			t.Fatalf("exporter started %d Jobs, want 1", len(ex.reqs))
+		}
+		r := ex.reqs[0]
+		if r.Server != "survival" || r.Mode != worldexport.ModeBackup || r.BackupRef != "/backups/survival-bk1.tar.gz" ||
+			!hex16.MatchString(r.ID) || !hex64.MatchString(r.Token) || r.Token == v.Ticket ||
+			r.TargetURL != exportBase+"/api/v1/internal/exports/"+r.ID {
+			t.Fatalf("export request = %+v", r)
+		}
+		if len(cl.acquired) != 0 {
+			t.Fatalf("a backup export took the world lock: %v", cl.acquired)
+		}
+		if len(repo.audits) != 1 || repo.audits[0].Action != "backup.export" || repo.audits[0].ServerName != "survival" ||
+			repo.audits[0].ActorUserID != "owner1" || string(repo.audits[0].Payload) != `{"backup_id":"bk1","size_bytes":1024}` {
+			t.Fatalf("audit = %+v", repo.audits)
+		}
+	})
+
+	for _, tc := range []struct {
+		name, user, path string
+		edit             func(*API, *fakeRepo)
+		code             int
+		errCode          string
+	}{
+		{name: "admin exports another's world", user: "admin1", path: backupPath,
+			edit: func(_ *API, r *fakeRepo) { r.backups[0].view.FormerOwner = "someone" }, code: http.StatusAccepted},
+		{name: "stranger", user: "stranger", path: backupPath, code: http.StatusForbidden, errCode: "forbidden"},
+		{name: "principal without an account", user: "nouser", path: backupPath, code: http.StatusForbidden, errCode: "forbidden"},
+		{name: "current owner who is not the former owner", user: "newowner", path: backupPath,
+			edit: func(_ *API, r *fakeRepo) { r.byName["survival"].OwnerID = "newowner" },
+			code: http.StatusNotFound, errCode: "no_backup"},
+		{name: "unknown backup", user: "owner1", path: "/api/v1/servers/survival/backups/nope/export",
+			code: http.StatusNotFound, errCode: "no_backup"},
+		{name: "another server's backup", user: "owner1", path: "/api/v1/servers/survival/backups/bk2/export",
+			code: http.StatusForbidden, errCode: "forbidden"},
+		{name: "corrupt backup", user: "owner1", path: backupPath,
+			edit: func(_ *API, r *fakeRepo) { r.backups[0].view.Corrupt = true }, code: http.StatusConflict, errCode: "backup_corrupt"},
+		{name: "bad name", user: "owner1", path: "/api/v1/servers/Bad_Name/backups/bk1/export",
+			code: http.StatusBadRequest, errCode: "bad_name"},
+		{name: "unknown server", user: "owner1", path: "/api/v1/servers/nosuch/backups/bk1/export",
+			code: http.StatusNotFound, errCode: "not_found"},
+		{name: "no exporter", user: "owner1", path: backupPath,
+			edit: func(a *API, _ *fakeRepo) { a.Exporter = nil }, code: http.StatusServiceUnavailable, errCode: "export_unavailable"},
+		{name: "no internal URL", user: "owner1", path: backupPath,
+			edit: func(a *API, _ *fakeRepo) { a.InternalBaseURL = "" }, code: http.StatusServiceUnavailable, errCode: "export_unavailable"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			a, repo, _, ex := exportFixture()
+			if tc.edit != nil {
+				tc.edit(a, repo)
+			}
+			w := do(a.ExternalHandler(), "POST", tc.path, "", as(tc.user))
+			if w.Code != tc.code {
+				t.Fatalf("code = %d, want %d (%s)", w.Code, tc.code, w.Body.String())
+			}
+			if tc.code == http.StatusAccepted {
+				if len(ex.reqs) != 1 {
+					t.Fatalf("exporter started %d Jobs, want 1", len(ex.reqs))
+				}
+				return
+			}
+			if got := decodeErr(t, w); got != tc.errCode {
+				t.Errorf("error code = %q, want %q", got, tc.errCode)
+			}
+			if len(ex.reqs) != 0 || len(repo.audits) != 0 {
+				t.Errorf("a refused export started %d Jobs and wrote %d audits", len(ex.reqs), len(repo.audits))
+			}
+		})
+	}
+}
+
+func TestExportWorldGate(t *testing.T) {
+	t.Run("owner exports a stopped world under the world lock", func(t *testing.T) {
+		a, repo, cl, ex := exportFixture()
+		v := beginExport(t, a.ExternalHandler(), worldPath, "owner1")
+		if v.Filename != "survival-world-"+exportStamp+".tar.gz" || v.State != "pending" {
+			t.Fatalf("ticket = %+v", v)
+		}
+		if len(ex.reqs) != 1 || ex.reqs[0].Mode != worldexport.ModeWorld || ex.reqs[0].BackupRef != "" || ex.reqs[0].Server != "survival" {
+			t.Fatalf("export requests = %+v", ex.reqs)
+		}
+		if strings.Join(cl.acquired, ",") != "survival:"+maintenance.KindExport || strings.Join(cl.released, ",") != "survival" {
+			t.Fatalf("lock acquired %v, released %v", cl.acquired, cl.released)
+		}
+		if len(repo.audits) != 1 || repo.audits[0].Action != "world.export" || repo.audits[0].ServerName != "survival" {
+			t.Fatalf("audit = %+v", repo.audits)
+		}
+	})
+
+	for _, tc := range []struct {
+		name, user string
+		edit       func(*API, *fakeCluster)
+		code       int
+		errCode    string
+	}{
+		{name: "admin", user: "admin1", code: http.StatusAccepted},
+		{name: "stranger", user: "stranger", code: http.StatusForbidden, errCode: "forbidden"},
+		{name: "running", user: "owner1", edit: func(_ *API, c *fakeCluster) { c.byName["survival"].Ready = true },
+			code: http.StatusConflict, errCode: "not_stopped"},
+		{name: "coming up", user: "owner1",
+			edit: func(_ *API, c *fakeCluster) { c.byName["survival"].DesiredState = string(v1alpha1.DesiredRunning) },
+			code: http.StatusConflict, errCode: "not_stopped"},
+		{name: "no world volume", user: "owner1", edit: func(_ *API, c *fakeCluster) { c.noWorld["survival"] = true },
+			code: http.StatusConflict, errCode: "no_world_volume"},
+		{name: "world busy", user: "owner1",
+			edit: func(_ *API, c *fakeCluster) {
+				c.maintErr["survival"] = &MaintenanceBusyError{Kind: maintenance.KindRestore}
+			},
+			code: http.StatusConflict, errCode: "maintenance_in_progress"},
+		{name: "no exporter", user: "owner1", edit: func(a *API, _ *fakeCluster) { a.Exporter = nil },
+			code: http.StatusServiceUnavailable, errCode: "export_unavailable"},
+		{name: "no internal URL", user: "owner1", edit: func(a *API, _ *fakeCluster) { a.InternalBaseURL = "" },
+			code: http.StatusServiceUnavailable, errCode: "export_unavailable"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			a, repo, cl, ex := exportFixture()
+			if tc.edit != nil {
+				tc.edit(a, cl)
+			}
+			w := do(a.ExternalHandler(), "POST", worldPath, "", as(tc.user))
+			if w.Code != tc.code {
+				t.Fatalf("code = %d, want %d (%s)", w.Code, tc.code, w.Body.String())
+			}
+			if tc.code == http.StatusAccepted {
+				if len(ex.reqs) != 1 {
+					t.Fatalf("exporter started %d Jobs, want 1", len(ex.reqs))
+				}
+				return
+			}
+			if got := decodeErr(t, w); got != tc.errCode {
+				t.Errorf("error code = %q, want %q", got, tc.errCode)
+			}
+			if len(ex.reqs) != 0 || len(repo.audits) != 0 || len(cl.released) != 0 {
+				t.Errorf("a refused export started %d Jobs, wrote %d audits, released %v", len(ex.reqs), len(repo.audits), cl.released)
+			}
+		})
+	}
+
+	// What anyone refused by a running world export hears.
+	t.Run("a world export holding the lock is named", func(t *testing.T) {
+		a, _, cl, _ := exportFixture()
+		cl.maintErr["survival"] = &MaintenanceBusyError{Kind: maintenance.KindExport}
+		w := do(a.ExternalHandler(), "POST", worldPath, "", as("admin1"))
+		var raw map[string]map[string]string
+		_ = json.Unmarshal(w.Body.Bytes(), &raw)
+		if want := "a world export is running on this server's world; retry once it finishes"; w.Code != http.StatusConflict ||
+			raw["error"]["code"] != "maintenance_in_progress" || raw["error"]["message"] != want {
+			t.Fatalf("busy = %d %s, want 409 %q", w.Code, w.Body.String(), want)
+		}
+	})
+
+	// A refusal after admission gives the export back: the owner is not left
+	// blocked by an export that never got a Job.
+	t.Run("lock conflict and a failed Job are refunded", func(t *testing.T) {
+		a, _, cl, ex := exportFixture()
+		h := a.ExternalHandler()
+		cl.maintErr["survival"] = &MaintenanceBusyError{Kind: maintenance.KindBackup}
+		if w := do(h, "POST", worldPath, "", as("owner1")); w.Code != http.StatusConflict {
+			t.Fatalf("busy world: code = %d, want 409", w.Code)
+		}
+		delete(cl.maintErr, "survival")
+		ex.err = errors.New("jobs is forbidden")
+		if w := do(h, "POST", worldPath, "", as("owner1")); w.Code != http.StatusInternalServerError {
+			t.Fatalf("failed Job: code = %d, want 500 (%s)", w.Code, w.Body.String())
+		}
+		if strings.Join(cl.released, ",") != "survival" {
+			t.Fatalf("a failed Job left the lock held: released %v", cl.released)
+		}
+		ex.err = nil
+		beginExport(t, h, worldPath, "owner1")
+		if n := len(a.exportTickets().starts["owner1"]); n != 1 {
+			t.Fatalf("hourly starts = %d, want only the export that got a Job", n)
+		}
+	})
+}
+
+func TestExportLimits(t *testing.T) {
+	t.Run("one per user", func(t *testing.T) {
+		a, _, _, ex := exportFixture()
+		h := a.ExternalHandler()
+		beginExport(t, h, worldPath, "owner1")
+		w := do(h, "POST", backupPath, "", as("owner1"))
+		if w.Code != http.StatusTooManyRequests || decodeErr(t, w) != "export_busy" || w.Header().Get("Retry-After") != "90" {
+			t.Fatalf("second export: %d %s Retry-After %q", w.Code, w.Body.String(), w.Header().Get("Retry-After"))
+		}
+		if len(ex.reqs) != 1 {
+			t.Fatalf("exporter started %d Jobs, want 1", len(ex.reqs))
+		}
+	})
+
+	t.Run("two across the install", func(t *testing.T) {
+		a, _, cl, ex := exportFixture()
+		h := a.ExternalHandler()
+		beginExport(t, h, worldPath, "owner1")
+		beginExport(t, h, "/api/v1/servers/creative/backups/bk2/export", "admin1")
+		w := do(h, "POST", "/api/v1/servers/gamma/world/export", "", as("owner3"))
+		if w.Code != http.StatusTooManyRequests || decodeErr(t, w) != "export_busy" || w.Header().Get("Retry-After") != "60" {
+			t.Fatalf("third export: %d %s Retry-After %q", w.Code, w.Body.String(), w.Header().Get("Retry-After"))
+		}
+		if len(ex.reqs) != 2 || strings.Join(cl.acquired, ",") != "survival:export" {
+			t.Fatalf("a refused export started a Job or took gamma's lock: %d Jobs, acquired %v", len(ex.reqs), cl.acquired)
+		}
+	})
+
+	t.Run("six per user per hour", func(t *testing.T) {
+		defer func(old time.Duration) { exportPendingTTL = old }(exportPendingTTL)
+		exportPendingTTL = time.Minute
+		a, _, _, _ := exportFixture()
+		var clock atomic.Int64
+		clock.Store(1_700_000_000)
+		a.Now = func() time.Time { return time.Unix(clock.Load(), 0) }
+		h := a.ExternalHandler()
+		for i := range exportPerHour {
+			clock.Store(1_700_000_000 + int64(i)*120) // each start outlives the last one's pending TTL
+			beginExport(t, h, worldPath, "owner1")
+		}
+		clock.Store(1_700_000_000 + 12*60)
+		w := do(h, "POST", worldPath, "", as("owner1"))
+		if w.Code != http.StatusTooManyRequests || decodeErr(t, w) != "export_busy" || w.Header().Get("Retry-After") != "2880" {
+			t.Fatalf("seventh export: %d %s Retry-After %q", w.Code, w.Body.String(), w.Header().Get("Retry-After"))
+		}
+		clock.Store(1_700_000_000 + 3600)
+		beginExport(t, h, worldPath, "owner1")
+	})
+}
+
+// TestExportRendezvous walks one world export end to end through the handlers:
+// the Job's PUT parks until the owner's download claims it, the archive streams
+// through untouched, the ticket works once and for its user only.
+func TestExportRendezvous(t *testing.T) {
+	a, repo, _, ex := exportFixture()
+	var clock atomic.Int64
+	clock.Store(1_700_000_000)
+	a.Now = func() time.Time { return time.Unix(clock.Load(), 0) }
+	ext, in := a.ExternalHandler(), a.InternalHandler()
+	v := beginExport(t, ext, worldPath, "owner1")
+	job := ex.reqs[0]
+	download := "/api/v1/exports/" + v.Ticket + "/download"
+
+	if code, s := exportState(t, ext, v.Ticket, "owner1"); code != http.StatusOK || s != (exportStatusView{State: "pending"}) {
+		t.Fatalf("fresh status = %d %+v", code, s)
+	}
+	for _, user := range []string{"stranger", "admin1"} {
+		if code, s := exportState(t, ext, v.Ticket, user); code != http.StatusNotFound || s.State != "not_found" {
+			t.Fatalf("%s reads the owner's ticket: %d %+v", user, code, s)
+		}
+	}
+	if w := do(ext, "GET", download, "", as("owner1")); w.Code != http.StatusConflict || decodeErr(t, w) != "export_not_ready" {
+		t.Fatalf("download while pending = %d %s", w.Code, w.Body.String())
+	}
+	for name, hdr := range map[string]map[string]string{
+		"no token":    nil,
+		"wrong token": {"Authorization": "Bearer " + strings.Repeat("0", 64)},
+		"the ticket":  {"Authorization": "Bearer " + v.Ticket},
+	} {
+		if w := doSoon(t, in, "PUT", "/api/v1/internal/exports/"+job.ID, "x", hdr); w.Code != http.StatusNotFound || decodeErr(t, w) != "not_found" {
+			t.Fatalf("PUT with %s = %d %s", name, w.Code, w.Body.String())
+		}
+	}
+	if w := doSoon(t, in, "PUT", "/api/v1/internal/exports/ffffffffffffffff", "x", map[string]string{"Authorization": "Bearer " + job.Token}); w.Code != http.StatusNotFound {
+		t.Fatalf("PUT to another id = %d", w.Code)
+	}
+
+	archive := randomBytes(3*exportCopyBuffer + 4321)
+	pr, pw := io.Pipe()
+	up := uploadExport(in, job.ID, job.Token, pr)
+	waitExportReady(t, ext, v.Ticket, "owner1")
+	if w := doSoon(t, in, "PUT", "/api/v1/internal/exports/"+job.ID, "x", map[string]string{"Authorization": "Bearer " + job.Token}); w.Code != http.StatusNotFound {
+		t.Fatalf("a second PUT with the spent token = %d, want 404", w.Code)
+	}
+
+	// Neither a HEAD nor another user's GET claims it.
+	if w := doSoon(t, ext, "HEAD", download, "", as("owner1")); w.Code != http.StatusMethodNotAllowed || w.Header().Get("Allow") != "GET" {
+		t.Fatalf("HEAD = %d, Allow %q", w.Code, w.Header().Get("Allow"))
+	}
+	if w := doSoon(t, ext, "GET", download, "", as("stranger")); w.Code != http.StatusNotFound {
+		t.Fatalf("stranger's download = %d", w.Code)
+	}
+	if code, s := exportState(t, ext, v.Ticket, "owner1"); code != http.StatusOK || s.State != exportReady {
+		t.Fatalf("after HEAD and a stranger: %d %+v, want still ready", code, s)
+	}
+
+	got := make(chan *httptest.ResponseRecorder, 1)
+	go func() { got <- do(ext, "GET", download, "", as("owner1")) }()
+	_, _ = pw.Write(archive[:10_000]) // returns once the download has read it
+	if code, s := exportState(t, ext, v.Ticket, "owner1"); code != http.StatusGone || s.State != "export_expired" {
+		t.Fatalf("status mid-download = %d %+v, want 410 export_expired", code, s)
+	}
+	if w := doSoon(t, ext, "GET", download, "", as("owner1")); w.Code != http.StatusGone || decodeErr(t, w) != "export_expired" {
+		t.Fatalf("a second download mid-stream = %d %s", w.Code, w.Body.String())
+	}
+	go func() {
+		for rest := archive[10_000:]; len(rest) > 0; {
+			n := min(len(rest), 10_000)
+			_, _ = pw.Write(rest[:n])
+			rest = rest[n:]
+		}
+		pw.Close()
+	}()
+	w := <-got
+	if w.Code != http.StatusOK || !bytes.Equal(w.Body.Bytes(), archive) {
+		t.Fatalf("download = %d, %d bytes; want 200 and the %d archive bytes", w.Code, w.Body.Len(), len(archive))
+	}
+	h := w.Header()
+	if h.Get("Content-Type") != "application/gzip" || h.Get("X-Content-Type-Options") != "nosniff" || h.Get("Cache-Control") != "no-store" ||
+		h.Get("Content-Disposition") != "attachment; filename=survival-world-"+exportStamp+".tar.gz" || h.Get("Content-Length") != "" {
+		t.Fatalf("download headers = %v", h)
+	}
+	if w := awaitUpload(t, up); w.Code != http.StatusNoContent {
+		t.Fatalf("upload answered %d %s, want 204", w.Code, w.Body.String())
+	}
+
+	if w := do(ext, "GET", download, "", as("owner1")); w.Code != http.StatusGone || decodeErr(t, w) != "export_expired" {
+		t.Fatalf("second download = %d %s", w.Code, w.Body.String())
+	}
+	if code, s := exportState(t, ext, v.Ticket, "owner1"); code != http.StatusGone || s.State != "export_expired" {
+		t.Fatalf("status after download = %d %+v", code, s)
+	}
+	clock.Add(int64(exportKeepSpent / time.Second))
+	if code, _ := exportState(t, ext, v.Ticket, "owner1"); code != http.StatusNotFound {
+		t.Fatalf("status long after = %d, want 404", code)
+	}
+	if len(repo.audits) != 1 {
+		t.Fatalf("audits = %+v, want the start only", repo.audits)
+	}
+	beginExport(t, ext, worldPath, "owner1") // the finished export no longer counts
+}
+
+func TestExportExpiry(t *testing.T) {
+	t.Run("a Job that never connects", func(t *testing.T) {
+		a, _, _, ex := exportFixture()
+		var clock atomic.Int64
+		clock.Store(1_700_000_000)
+		a.Now = func() time.Time { return time.Unix(clock.Load(), 0) }
+		ext := a.ExternalHandler()
+		v := beginExport(t, ext, worldPath, "owner1")
+		clock.Add(int64(exportPendingTTL/time.Second) - 1)
+		if code, s := exportState(t, ext, v.Ticket, "owner1"); code != http.StatusOK || s.State != "pending" {
+			t.Fatalf("just inside the pending TTL: %d %+v", code, s)
+		}
+		clock.Add(1)
+		if code, s := exportState(t, ext, v.Ticket, "owner1"); code != http.StatusGone || s.State != "export_expired" {
+			t.Fatalf("past the pending TTL: %d %+v", code, s)
+		}
+		job := ex.reqs[0]
+		if w := doSoon(t, a.InternalHandler(), "PUT", "/api/v1/internal/exports/"+job.ID, "x", map[string]string{"Authorization": "Bearer " + job.Token}); w.Code != http.StatusNotFound {
+			t.Fatalf("a late Job's PUT = %d, want 404", w.Code)
+		}
+	})
+
+	t.Run("a browser that never comes", func(t *testing.T) {
+		defer func(old time.Duration) { exportClaimTTL = old }(exportClaimTTL)
+		exportClaimTTL = 30 * time.Millisecond
+		a, _, _, ex := exportFixture()
+		ext := a.ExternalHandler()
+		v := beginExport(t, ext, backupPath, "owner1")
+		job := ex.reqs[0]
+		w := awaitUpload(t, uploadExport(a.InternalHandler(), job.ID, job.Token, strings.NewReader("archive")))
+		if w.Code != http.StatusGone || decodeErr(t, w) != "export_expired" {
+			t.Fatalf("unclaimed upload = %d %s, want 410 export_expired", w.Code, w.Body.String())
+		}
+		if w := do(ext, "GET", "/api/v1/exports/"+v.Ticket+"/download", "", as("owner1")); w.Code != http.StatusGone {
+			t.Fatalf("download after the claim TTL = %d, want 410", w.Code)
+		}
+	})
+}
+
+// TestExportStatusReportsFailedJob: a Job that dies before it connects turns
+// the export failed with the Job's own error, and frees the owner's slot.
+func TestExportStatusReportsFailedJob(t *testing.T) {
+	a, _, _, ex := exportFixture()
+	js := &fakeJobStatus{}
+	a.JobStatus = js
+	ext := a.ExternalHandler()
+	v := beginExport(t, ext, worldPath, "owner1")
+	job := ex.reqs[0]
+	name := worldexport.JobName("survival", job.ID)
+
+	js.jobs = []AsyncJob{
+		{Name: "export-survival-0000000000000000", Kind: "export_world", State: "failed", Message: "someone else's"},
+		{Name: name, Kind: "export_world", State: "running"},
+	}
+	if code, s := exportState(t, ext, v.Ticket, "owner1"); code != http.StatusOK || s != (exportStatusView{State: "pending"}) {
+		t.Fatalf("running Job: %d %+v", code, s)
+	}
+	js.jobs[1] = AsyncJob{Name: name, Kind: "export_world", State: "failed", Message: "felis export: open /world: permission denied"}
+	want := exportStatusView{State: "failed", Message: "felis export: open /world: permission denied"}
+	if code, s := exportState(t, ext, v.Ticket, "owner1"); code != http.StatusOK || s != want || js.got != "survival" {
+		t.Fatalf("failed Job: %d %+v (asked about %q)", code, s, js.got)
+	}
+	js.jobs = nil
+	if code, s := exportState(t, ext, v.Ticket, "owner1"); code != http.StatusOK || s != want {
+		t.Fatalf("failed export once the Job is gone: %d %+v", code, s)
+	}
+	if w := doSoon(t, a.InternalHandler(), "PUT", "/api/v1/internal/exports/"+job.ID, "x", map[string]string{"Authorization": "Bearer " + job.Token}); w.Code != http.StatusNotFound {
+		t.Fatalf("PUT on a failed export = %d, want 404", w.Code)
+	}
+	beginExport(t, ext, backupPath, "owner1")
+}
+
+// TestExportRegistryRaces: two transitions the handlers reach only in a race.
+// A Job reported failed just as its upload arrived must not fail the ready
+// export, and the claim timer firing as the browser claims must not spend the
+// export under its download.
+func TestExportRegistryRaces(t *testing.T) {
+	now := time.Unix(1_700_000_000, 0)
+	reg := (&API{}).exportTickets()
+	e := &exportEntry{userID: "owner1", server: "survival", mode: worldexport.ModeWorld, filename: "survival.tar.gz"}
+	token, err := reg.admit(e, now)
+	if err != nil {
+		t.Fatal(err)
+	}
+	up := &exportUpload{body: strings.NewReader("x"), size: 1, claimed: make(chan struct{}), done: make(chan error, 1)}
+	if !reg.arrive(e.id, token, up, now) {
+		t.Fatal("the upload did not arrive")
+	}
+
+	if reg.fail(e.ticket, "felis export: killed", now) {
+		t.Fatal("a late Job failure failed a ready export")
+	}
+	if got, err := reg.lookup(e.ticket, "owner1", now); err != nil || got.state != exportReady || got.message != "" {
+		t.Fatalf("after a late Job failure: state %q, message %q, err %v; want ready", got.state, got.message, err)
+	}
+
+	if _, err := reg.claim(e.ticket, "owner1", now); err != nil {
+		t.Fatalf("claim: %v", err)
+	}
+	if reg.giveUp(up, now) {
+		t.Fatal("the claim timer spent an export being downloaded")
+	}
+	if e.state != exportStreaming {
+		t.Fatalf("state after a late giveUp = %q, want streaming", e.state)
+	}
+}
+
+// exportServers runs both faces for real, so deadlines apply and an aborted
+// download reaches the client as an error.
+func exportServers(t *testing.T, a *API) (ext, in *httptest.Server) {
+	ext, in = httptest.NewServer(a.ExternalHandler()), httptest.NewServer(a.InternalHandler())
+	t.Cleanup(func() { ext.Close(); in.Close() })
+	return ext, in
+}
+
+func realUpload(t *testing.T, in *httptest.Server, job worldexport.Request, body io.Reader, size int64) <-chan *http.Response {
+	req, err := http.NewRequest("PUT", in.URL+"/api/v1/internal/exports/"+job.ID, body)
+	if err != nil {
+		t.Fatal(err)
+	}
+	req.ContentLength = size
+	req.Header.Set("Authorization", "Bearer "+job.Token)
+	out := make(chan *http.Response, 1)
+	go func() {
+		resp, err := http.DefaultClient.Do(req)
+		if err != nil {
+			out <- &http.Response{StatusCode: -1, Status: err.Error(), Body: io.NopCloser(strings.NewReader(""))}
+			return
+		}
+		out <- resp
+	}()
+	return out
+}
+
+// realDownload fetches the owner's download: the response (nil when the request
+// failed outright), the bytes read and the error that ended the read.
+func realDownload(ext *httptest.Server, ticket string) (*http.Response, []byte, error) {
+	req, _ := http.NewRequest("GET", ext.URL+"/api/v1/exports/"+ticket+"/download", nil)
+	req.Header.Set("X-Test-User", "owner1")
+	resp, err := http.DefaultClient.Do(req)
+	if err != nil {
+		return nil, nil, err
+	}
+	defer resp.Body.Close()
+	got, err := io.ReadAll(resp.Body)
+	return resp, got, err
+}
+
+func awaitResponse(t *testing.T, ch <-chan *http.Response) *http.Response {
+	t.Helper()
+	select {
+	case resp := <-ch:
+		t.Cleanup(func() { resp.Body.Close() })
+		return resp
+	case <-time.After(5 * time.Second):
+		t.Fatal("the upload never answered")
+		return nil
+	}
+}
+
+// TestExportBackupDigest: a backup streams with its length and is checked
+// against the sha256 recorded when it was written. A match downloads whole; a
+// mismatch withholds the tail and aborts, so the browser never holds a
+// complete-looking corrupt file, and the Job hears backup_corrupt.
+func TestExportBackupDigest(t *testing.T) {
+	archive := randomBytes(3*exportCopyBuffer + 4321)
+	sum := sha256.Sum256(archive)
+	for _, tc := range []struct {
+		name   string
+		digest string
+	}{
+		{"recorded digest matches", hex.EncodeToString(sum[:])},
+		{"no digest recorded", ""},
+		{"digest mismatch", strings.Repeat("ab", 32)},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			a, repo, _, ex := exportFixture()
+			repo.backups[0].sha256 = tc.digest
+			ext, in := exportServers(t, a)
+			v := beginExport(t, a.ExternalHandler(), backupPath, "owner1")
+			up := realUpload(t, in, ex.reqs[0], bytes.NewReader(archive), int64(len(archive)))
+			waitExportReady(t, a.ExternalHandler(), v.Ticket, "owner1")
+			resp, got, err := realDownload(ext, v.Ticket)
+			if resp == nil {
+				t.Fatalf("download: %v", err)
+			}
+			if resp.StatusCode != http.StatusOK || resp.Header.Get("Content-Length") != strconv.Itoa(len(archive)) {
+				t.Fatalf("download = %d, Content-Length %q", resp.StatusCode, resp.Header.Get("Content-Length"))
+			}
+			upResp := awaitResponse(t, up)
+			if tc.digest == strings.Repeat("ab", 32) {
+				if err == nil || len(got) >= len(archive) || !bytes.Equal(got, archive[:len(got)]) {
+					t.Fatalf("mismatch: read %d of %d bytes, err %v; want an error short of the end", len(got), len(archive), err)
+				}
+				if upResp.StatusCode != http.StatusConflict || errCode(mustRead(t, upResp.Body)) != "backup_corrupt" {
+					t.Fatalf("upload answered %d, want 409 backup_corrupt", upResp.StatusCode)
+				}
+				return
+			}
+			if err != nil || !bytes.Equal(got, archive) {
+				t.Fatalf("read %d of %d bytes, err %v", len(got), len(archive), err)
+			}
+			if upResp.StatusCode != http.StatusNoContent {
+				t.Fatalf("upload answered %d, want 204", upResp.StatusCode)
+			}
+		})
+	}
+
+	// The tail is withheld from the response writer itself, not only from
+	// whatever the connection had yet to send: with every write captured, a
+	// mismatch still ends short of the archive.
+	t.Run("the tail waits for the digest", func(t *testing.T) {
+		a, repo, _, ex := exportFixture()
+		repo.backups[0].sha256 = strings.Repeat("ab", 32)
+		ext := a.ExternalHandler()
+		v := beginExport(t, ext, backupPath, "owner1")
+		up := uploadExport(a.InternalHandler(), ex.reqs[0].ID, ex.reqs[0].Token, bytes.NewReader(archive))
+		waitExportReady(t, ext, v.Ticket, "owner1")
+		w := httptest.NewRecorder()
+		func() {
+			defer func() {
+				if p := recover(); p != http.ErrAbortHandler {
+					t.Errorf("the download ended with %v, want the abort", p)
+				}
+			}()
+			r := httptest.NewRequest("GET", "/api/v1/exports/"+v.Ticket+"/download", nil)
+			r.Header.Set("X-Test-User", "owner1")
+			ext.ServeHTTP(w, r)
+		}()
+		if got := w.Body.Bytes(); len(got) != 3*exportCopyBuffer || !bytes.Equal(got, archive[:len(got)]) {
+			t.Fatalf("wrote %d of %d bytes, want all but the last read (%d)", len(got), len(archive), 3*exportCopyBuffer)
+		}
+		if w := awaitUpload(t, up); w.Code != http.StatusConflict || decodeErr(t, w) != "backup_corrupt" {
+			t.Fatalf("upload answered %d %s, want 409 backup_corrupt", w.Code, w.Body.String())
+		}
+	})
+}
+
+// zeros reads as an endless run of zero bytes.
+type zeros struct{}
+
+func (zeros) Read(p []byte) (int, error) {
+	clear(p)
+	return len(p), nil
+}
+
+func mustRead(t *testing.T, r io.Reader) []byte {
+	t.Helper()
+	b, err := io.ReadAll(r)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return b
+}
+
+// TestExportUploadPace: the Job's upload waits for the browser longer than the
+// body-deadline grace, and then moves at the browser's pace, without being cut
+// off; a body that stops moving for exportStall is.
+func TestExportUploadPace(t *testing.T) {
+	defer func(g time.Duration, r float64, s time.Duration) { bodyGrace, bodyMinRate, exportStall = g, r, s }(bodyGrace, bodyMinRate, exportStall)
+	bodyGrace, bodyMinRate, exportStall = 100*time.Millisecond, 1<<30, 300*time.Millisecond
+
+	t.Run("a slow claim is not a slow body", func(t *testing.T) {
+		a, _, _, ex := exportFixture()
+		ext, in := exportServers(t, a)
+		v := beginExport(t, a.ExternalHandler(), worldPath, "owner1")
+		archive := randomBytes(64 << 10)
+		pr, pw := io.Pipe()
+		go func() {
+			_, _ = pw.Write(archive[:1000])
+			time.Sleep(3 * bodyGrace)
+			_, _ = pw.Write(archive[1000:])
+			pw.Close()
+		}()
+		up := realUpload(t, in, ex.reqs[0], pr, -1)
+		waitExportReady(t, a.ExternalHandler(), v.Ticket, "owner1")
+		time.Sleep(3 * bodyGrace)
+		_, got, err := realDownload(ext, v.Ticket)
+		if err != nil || !bytes.Equal(got, archive) {
+			t.Fatalf("read %d of %d bytes, err %v", len(got), len(archive), err)
+		}
+		if resp := awaitResponse(t, up); resp.StatusCode != http.StatusNoContent {
+			t.Fatalf("upload answered %d, want 204", resp.StatusCode)
+		}
+	})
+
+	t.Run("a stalled body is cut off", func(t *testing.T) {
+		a, _, _, ex := exportFixture()
+		ext, in := exportServers(t, a)
+		v := beginExport(t, a.ExternalHandler(), worldPath, "owner1")
+		pr, pw := io.Pipe()
+		defer pw.Close()
+		go func() { _, _ = pw.Write(make([]byte, 1000)) }() // then nothing, ever
+		up := realUpload(t, in, ex.reqs[0], pr, -1)
+		waitExportReady(t, a.ExternalHandler(), v.Ticket, "owner1")
+		done := make(chan error, 1)
+		go func() { _, _, err := realDownload(ext, v.Ticket); done <- err }()
+		select {
+		case err := <-done:
+			if err == nil {
+				t.Fatal("a stalled archive downloaded as complete")
+			}
+		case <-time.After(5 * time.Second):
+			t.Fatal("a stalled upload was never cut off")
+		}
+		pw.CloseWithError(errors.New("test over"))
+		awaitResponse(t, up)
+	})
+
+	t.Run("a browser that stops reading is cut off", func(t *testing.T) {
+		a, _, _, ex := exportFixture()
+		ext, in := exportServers(t, a)
+		v := beginExport(t, a.ExternalHandler(), worldPath, "owner1")
+		realUpload(t, in, ex.reqs[0], io.LimitReader(zeros{}, 1<<30), -1)
+		waitExportReady(t, a.ExternalHandler(), v.Ticket, "owner1")
+		req, _ := http.NewRequest("GET", ext.URL+"/api/v1/exports/"+v.Ticket+"/download", nil)
+		req.Header.Set("X-Test-User", "owner1")
+		resp, err := http.DefaultClient.Do(req)
+		if err != nil {
+			t.Fatal(err)
+		}
+		t.Cleanup(func() { resp.Body.Close() }) // before the servers close, so a held handler ends
+		// Read nothing: the socket buffers fill and the next write waits.
+		reg := a.exportTickets()
+		for deadline := time.Now().Add(5 * time.Second); ; {
+			reg.mu.Lock()
+			state := reg.byTicket[v.Ticket].state
+			reg.mu.Unlock()
+			if state == exportSpent {
+				break
+			}
+			if time.Now().After(deadline) {
+				t.Fatalf("a browser that stopped reading held the download open: state %q", state)
+			}
+			time.Sleep(10 * time.Millisecond)
+		}
+		if _, err := io.ReadAll(resp.Body); err == nil {
+			t.Fatal("the cut-off download read as complete")
+		}
+	})
+}
+
+// TestK8sExportJobs: export Jobs show in the jobs list by what they archive,
+// and never count as world work the backup scheduler waits for.
+func TestK8sExportJobs(t *testing.T) {
+	scheme := runtime.NewScheme()
+	if err := clientgoscheme.AddToScheme(scheme); err != nil {
+		t.Fatal(err)
+	}
+	job := func(id, mode string) *batchv1.Job {
+		j, err := worldexport.ExportJob(worldexport.JobParams{
+			Server: "survival", ID: id, Mode: mode,
+			WorldPVC: "world-survival-0", BackupPVC: "felis-backups", BackupRef: "/backups/a.tar.gz",
+			TargetURL: exportBase + "/x", Token: "t", Namespace: "minecraft", Image: "felis:1",
+		})
+		if err != nil {
+			t.Fatal(err)
+		}
+		return j
+	}
+	world, backupJob := job("1111111111111111", worldexport.ModeWorld), job("2222222222222222", worldexport.ModeBackup)
+	restoring := &batchv1.Job{ObjectMeta: metav1.ObjectMeta{Namespace: "minecraft", Name: "restore-survival-cc",
+		Labels: map[string]string{jobServerLabel: "survival", jobManagedByLabel: jobManagedByRestore}}}
+	c := fake.NewClientBuilder().WithScheme(scheme).WithObjects(world, backupJob, restoring).
+		WithStatusSubresource(&batchv1.Job{}).Build()
+	k := NewK8sJobStatus(c, "minecraft")
+	ctx := context.Background()
+
+	if n, err := k.RunningWorldJobs(ctx); err != nil || n != 1 {
+		t.Fatalf("RunningWorldJobs = %d, %v; want the restore only", n, err)
+	}
+	jobs, err := k.LatestJobs(ctx, "survival")
+	if err != nil {
+		t.Fatal(err)
+	}
+	kinds := map[string]string{}
+	for _, j := range jobs {
+		kinds[j.Name] = j.Kind + "/" + j.State
+	}
+	want := map[string]string{world.Name: "export_world/running", backupJob.Name: "export_backup/running", restoring.Name: "restore/running"}
+	if len(kinds) != len(want) {
+		t.Fatalf("jobs = %v, want %v", kinds, want)
+	}
+	for name, k := range want {
+		if kinds[name] != k {
+			t.Errorf("%s = %q, want %q", name, kinds[name], k)
+		}
+	}
+
+	// The kinds agree with maintenance.JobKind: a world export holds the world,
+	// a backup export does not.
+	if kind, holds := maintenance.JobKind(world); kind != maintenance.KindExport || !holds {
+		t.Errorf("JobKind(world export) = %q, %v", kind, holds)
+	}
+	if _, holds := maintenance.JobKind(backupJob); holds {
+		t.Error("a backup export holds the world")
+	}
+}

internal/api/jobstatus.go

@@ -14,7 +14,7 @@ import (
 // object an operator with kubectl could read. The route is the API-side outlet.
 type AsyncJob struct {
 	Name       string    `json:"name"`
-	Kind       string    `json:"kind"`  // "backup" | "restore"
+	Kind       string    `json:"kind"`  // "backup" | "restore" | "export_world" | "export_backup"
 	State      string    `json:"state"` // "running" | "succeeded" | "failed"
 	Message    string    `json:"message,omitempty"`
 	StartedAt  time.Time `json:"started_at,omitzero"`

internal/api/k8sjobstatus.go

@@ -25,6 +25,7 @@ const (
 
 	jobManagedByBackup  = "felis-backup"
 	jobManagedByRestore = "felis-restore"
+	jobManagedByExport  = "felis-export"
 
 	// jobBackupReasonLabel marks the backup Job of a scheduled restore point
 	// (backupjob.LabelReason).
@@ -45,9 +46,9 @@ func NewK8sJobStatus(c client.Client, namespace string) *K8sJobStatus {
 	return &K8sJobStatus{c: c, namespace: namespace}
 }
 
-// LatestJobs lists this server's backup/restore Jobs newest-first, capped so a
-// long history cannot balloon the response. Jobs the label selector catches but
-// another component created (unknown managed-by) are dropped.
+// LatestJobs lists this server's backup, restore and export Jobs newest-first,
+// capped so a long history cannot balloon the response. Jobs the label selector
+// catches but another component created (unknown managed-by) are dropped.
 func (k *K8sJobStatus) LatestJobs(ctx context.Context, serverName string) ([]AsyncJob, error) {
 	var list batchv1.JobList
 	if err := k.c.List(ctx, &list, client.InNamespace(k.namespace),
@@ -147,7 +148,10 @@ func jobToAsyncJob(j *batchv1.Job) (AsyncJob, bool) {
 }
 
 // RunningWorldJobs counts the backup and restore Jobs of every server that
-// have yet to finish (BackupScheduler waits for them).
+// have yet to finish (BackupScheduler waits for them). Export Jobs are left
+// out: one moves at its browser's pace, a download left running for an hour
+// must not hold every world's restore point back, and the export limits
+// (exportMaxActive) already bound the load they add.
 func (k *K8sJobStatus) RunningWorldJobs(ctx context.Context) (int, error) {
 	var list batchv1.JobList
 	if err := k.c.List(ctx, &list, client.InNamespace(k.namespace), client.HasLabels{jobManagedByLabel}); err != nil {
@@ -156,7 +160,7 @@ func (k *K8sJobStatus) RunningWorldJobs(ctx context.Context) (int, error) {
 	n := 0
 	for i := range list.Items {
 		j := &list.Items[i]
-		if _, ok := jobOutcome(j); ok && !maintenance.JobFinished(j) {
+		if _, ok := jobOutcome(j); ok && j.Labels[jobManagedByLabel] != jobManagedByExport && !maintenance.JobFinished(j) {
 			n++
 		}
 	}
@@ -225,6 +229,13 @@ func jobOutcome(j *batchv1.Job) (AsyncJob, bool) {
 		kind = "backup"
 	case jobManagedByRestore:
 		kind = "restore"
+	case jobManagedByExport:
+		// As maintenance.JobKind reads it: only a Job that says it reads a
+		// backup is not a world export.
+		kind = "export_world"
+		if j.Labels[maintenance.LabelExportMode] == maintenance.ExportModeBackup {
+			kind = "export_backup"
+		}
 	default:
 		return AsyncJob{}, false
 	}

internal/api/maintenance.go

@@ -11,9 +11,9 @@ import (
 )
 
 // maintenanceError maps the world-volume lock's refusals onto their 409s:
-// maintenance_in_progress while a restore, backup or file write holds the volume,
-// not_stopped (with the caller's wording) while the server is not fully down.
-// Anything else passes through unchanged.
+// maintenance_in_progress while a restore, backup, file write or world export
+// holds the volume, not_stopped (with the caller's wording) while the server is
+// not fully down. Anything else passes through unchanged.
 func maintenanceError(err error, notStopped string) error {
 	var busy *MaintenanceBusyError
 	switch {
@@ -37,6 +37,8 @@ func maintenanceLabel(kind string) string {
 		return "a backup"
 	case maintenance.KindFileWrite:
 		return "a file write"
+	case maintenance.KindExport:
+		return "a world export"
 	case maintenance.KindReap:
 		return "the idle-world reaper"
 	}

internal/api/middleware.go

@@ -210,6 +210,18 @@ func (b *deadlineBody) Read(p []byte) (int, error) {
 	return n, err
 }
 
+// takeBodyDeadline lifts withBodyDeadline from a request whose body is read at
+// someone else's pace (an export's upload waits for the browser, then moves at
+// its speed), and returns the controller the handler bounds it with instead.
+func takeBodyDeadline(w http.ResponseWriter, r *http.Request) *http.ResponseController {
+	if b, ok := r.Body.(*deadlineBody); ok {
+		r.Body = b.ReadCloser
+		_ = b.rc.SetReadDeadline(time.Time{})
+		return b.rc
+	}
+	return http.NewResponseController(w)
+}
+
 // requireInternal enforces service-token auth for the internal face and stashes
 // the caller the token belongs to, which callersOnly checks against the route.
 // It never applies Zero Trust (spec §14 red line).

internal/api/openapi_parity_test.go

@@ -40,6 +40,8 @@ func TestOpenAPISchemasMatchWireStructs(t *testing.T) {
 		"MyServerView":          MyServerView{},
 		"AllowlistEntry":        AllowlistEntry{},
 		"BackupView":            BackupView{},
+		"ExportTicket":          exportTicketView{},
+		"ExportStatus":          exportStatusView{},
 		"Schedule":              Schedule{},
 		"Build":                 build.Build{},
 		"Image":                 build.Image{},

internal/api/pgrepo.go

@@ -958,11 +958,11 @@ func (p *PGRepo) LatestBackup(ctx context.Context, serverName string) (*BackupRe
 // BackupByID returns a single present backup by its id, or ErrNotFound.
 func (p *PGRepo) BackupByID(ctx context.Context, id string) (*BackupRecord, error) {
 	const q = `SELECT id, server_name, COALESCE(former_owner, ''), backup_ref, COALESCE(size_bytes, 0),
-		corrupt_at IS NOT NULL
+		corrupt_at IS NOT NULL, COALESCE(sha256, '')
 		FROM world_backups WHERE id = $1 AND status = 'present'`
 	var b BackupRecord
 	switch err := p.db.QueryRowContext(ctx, q, id).Scan(
-		&b.ID, &b.ServerName, &b.FormerOwner, &b.BackupRef, &b.SizeBytes, &b.Corrupt); {
+		&b.ID, &b.ServerName, &b.FormerOwner, &b.BackupRef, &b.SizeBytes, &b.Corrupt, &b.SHA256); {
 	case errors.Is(err, sql.ErrNoRows):
 		return nil, ErrNotFound
 	case err != nil:

internal/api/repo.go

@@ -143,6 +143,10 @@ type BackupRecord struct {
 	SizeBytes   int64
 	// Corrupt reports that the archive failed a read-back (see BackupView).
 	Corrupt bool
+	// SHA256 is the digest recorded when the archive was written, which an
+	// export checks the bytes against as they stream. Empty for an archive from
+	// before digests were kept.
+	SHA256 string
 }
 
 // StaffUser is the login-side projection of a users row (spec §B passwordless