diff --git a/docs/superpowers/plans/2026-05-28-library-delete-resilience.md b/docs/superpowers/plans/2026-05-28-library-delete-resilience.md
new file mode 100644
index 00000000..e8211bd2
--- /dev/null
+++ b/docs/superpowers/plans/2026-05-28-library-delete-resilience.md
@@ -0,0 +1,724 @@
+# Resilient Library Deletion Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Rewrite `delete_library` so it deletes a library in small, autocommitted, deadlock-retrying batches instead of one multi-minute transaction that deadlocks and rolls back on large libraries.
+
+**Architecture:** `FolderRepository.DeleteWithStats` becomes a phased, batched operation (orphan items → folder media_files → remaining shared junctions → folder row). Each batch is one autocommit statement wrapped in a `retryOnDeadlock` helper, so locks are held briefly and the job is resumable. Two small helpers (`retryOnDeadlock`, `deleteInBatches`) are unit-tested in pure Go; the full path is verified against the live dev database.
+
+**Tech Stack:** Go, pgx v5 (`pgxpool`, `pgconn`), PostgreSQL. Spec: `docs/superpowers/specs/2026-05-28-library-delete-resilience-design.md`.
+
+All commands assume the repository root is the cwd.
+
+---
+
+## File Structure
+
+- **Modify** `internal/catalog/folder_repo.go`
+  - Add constants `orphanDeleteBatch`, `folderChildDeleteBatch`.
+  - Add package vars `deadlockMaxAttempts`, `deadlockBaseBackoff` and func `retryOnDeadlock`.
+  - Add func `deleteInBatches`.
+  - Add interface `rowQuerier`; change `collectImageDirs` and `filterUnreferencedImageDirs` to take `rowQuerier` instead of `pgx.Tx`; extract `collectRawImageDirs`.
+  - Rewrite `DeleteWithStats`; add helpers `collectOrphanBatch`, `dirSetToSlice`.
+  - Leave `collectOrphanIDs` and `Delete` unchanged.
+- **Create** `internal/catalog/folder_delete_test.go` — pure-Go unit tests for `retryOnDeadlock` and `deleteInBatches`.
+- **Do not modify** `internal/catalog/library_repo.go` — its `collectImageDirs`/`collectOrphanIDs` calls keep compiling because `pgx.Tx` satisfies `rowQuerier`. (Verified by build.)
+- **Unchanged behavior** in `internal/adminjob/library_delete.go` (executor), `internal/adminjob/runner.go` (heartbeat).
+
+---
+
+## Task 1: Deadlock-retry helper
+
+**Files:**
+- Modify: `internal/catalog/folder_repo.go`
+- Test: `internal/catalog/folder_delete_test.go` (create)
+
+- [ ] **Step 1: Write the failing tests**
+
+Create `internal/catalog/folder_delete_test.go`:
+
+```go
+package catalog
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+
+	"github.com/jackc/pgx/v5/pgconn"
+)
+
+// withFastDeadlockRetry shrinks retry timing/attempts for tests and restores
+// the originals on cleanup. Tests using it must not call t.Parallel().
+func withFastDeadlockRetry(t *testing.T, maxAttempts int) {
+	t.Helper()
+	oldMax, oldBackoff := deadlockMaxAttempts, deadlockBaseBackoff
+	deadlockMaxAttempts = maxAttempts
+	deadlockBaseBackoff = time.Millisecond
+	t.Cleanup(func() {
+		deadlockMaxAttempts = oldMax
+		deadlockBaseBackoff = oldBackoff
+	})
+}
+
+func TestRetryOnDeadlockRetriesThenSucceeds(t *testing.T) {
+	withFastDeadlockRetry(t, 5)
+	calls := 0
+	err := retryOnDeadlock(context.Background(), func() error {
+		calls++
+		if calls < 3 {
+			return &pgconn.PgError{Code: "40P01"}
+		}
+		return nil
+	})
+	if err != nil {
+		t.Fatalf("expected success, got %v", err)
+	}
+	if calls != 3 {
+		t.Fatalf("expected 3 calls, got %d", calls)
+	}
+}
+
+func TestRetryOnDeadlockReturnsNonRetryableImmediately(t *testing.T) {
+	withFastDeadlockRetry(t, 5)
+	sentinel := errors.New("boom")
+	calls := 0
+	err := retryOnDeadlock(context.Background(), func() error {
+		calls++
+		return sentinel
+	})
+	if !errors.Is(err, sentinel) {
+		t.Fatalf("expected sentinel, got %v", err)
+	}
+	if calls != 1 {
+		t.Fatalf("expected 1 call, got %d", calls)
+	}
+}
+
+func TestRetryOnDeadlockGivesUpAfterMaxAttempts(t *testing.T) {
+	withFastDeadlockRetry(t, 4)
+	calls := 0
+	err := retryOnDeadlock(context.Background(), func() error {
+		calls++
+		return &pgconn.PgError{Code: "40P01"}
+	})
+	var pgErr *pgconn.PgError
+	if !errors.As(err, &pgErr) || pgErr.Code != "40P01" {
+		t.Fatalf("expected 40P01 pg error, got %v", err)
+	}
+	if calls != 4 {
+		t.Fatalf("expected 4 calls, got %d", calls)
+	}
+}
+
+func TestRetryOnDeadlockStopsOnCanceledContext(t *testing.T) {
+	withFastDeadlockRetry(t, 5)
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	calls := 0
+	err := retryOnDeadlock(ctx, func() error {
+		calls++
+		return &pgconn.PgError{Code: "40P01"}
+	})
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("expected context.Canceled, got %v", err)
+	}
+	if calls != 1 {
+		t.Fatalf("expected 1 call before cancel, got %d", calls)
+	}
+}
+```
+
+- [ ] **Step 2: Run tests to verify they fail**
+
+Run: `go test ./internal/catalog/ -run TestRetryOnDeadlock -v`
+Expected: compile failure / FAIL — `undefined: retryOnDeadlock`, `deadlockMaxAttempts`, `deadlockBaseBackoff`.
+
+- [ ] **Step 3: Implement the helper**
+
+In `internal/catalog/folder_repo.go`, add near the top of the file (after the `import` block, before the sentinel errors var is fine):
+
+```go
+// Retry parameters for transient serialization/deadlock failures. They are
+// package vars (not consts) only so tests can shrink them; production code
+// never mutates them.
+var (
+	deadlockMaxAttempts = 5
+	deadlockBaseBackoff = 50 * time.Millisecond
+)
+
+// retryOnDeadlock runs op, retrying when Postgres reports a deadlock (40P01) or
+// serialization failure (40001), with exponential backoff. It returns
+// immediately for any other error, and honors context cancellation between
+// attempts.
+func retryOnDeadlock(ctx context.Context, op func() error) error {
+	backoff := deadlockBaseBackoff
+	for attempt := 1; ; attempt++ {
+		err := op()
+		if err == nil {
+			return nil
+		}
+		var pgErr *pgconn.PgError
+		if attempt < deadlockMaxAttempts && errors.As(err, &pgErr) &&
+			(pgErr.Code == "40P01" || pgErr.Code == "40001") {
+			select {
+			case <-ctx.Done():
+				return ctx.Err()
+			case <-time.After(backoff):
+			}
+			backoff *= 2
+			continue
+		}
+		return err
+	}
+}
+```
+
+(`context`, `errors`, `time`, and `github.com/jackc/pgx/v5/pgconn` are already imported in this file.)
+
+- [ ] **Step 4: Run tests to verify they pass**
+
+Run: `go test ./internal/catalog/ -run TestRetryOnDeadlock -v`
+Expected: PASS (4 tests).
+
+- [ ] **Step 5: Format and commit**
+
+```bash
+gofmt -w internal/catalog/folder_repo.go internal/catalog/folder_delete_test.go
+git add internal/catalog/folder_repo.go internal/catalog/folder_delete_test.go
+git commit -m "feat(catalog): add deadlock-retry helper for batched deletes"
+```
+
+---
+
+## Task 2: Batched-delete loop helper
+
+**Files:**
+- Modify: `internal/catalog/folder_repo.go`
+- Test: `internal/catalog/folder_delete_test.go`
+
+- [ ] **Step 1: Write the failing tests**
+
+Append to `internal/catalog/folder_delete_test.go`:
+
+```go
+func TestDeleteInBatchesLoopsUntilUnderBatchSize(t *testing.T) {
+	withFastDeadlockRetry(t, 5)
+	counts := []int64{5, 5, 2}
+	idx := 0
+	total, err := deleteInBatches(context.Background(), 5, func(context.Context) (int64, error) {
+		n := counts[idx]
+		idx++
+		return n, nil
+	})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if total != 12 {
+		t.Fatalf("expected total 12, got %d", total)
+	}
+	if idx != 3 {
+		t.Fatalf("expected 3 batches, got %d", idx)
+	}
+}
+
+func TestDeleteInBatchesStopsImmediatelyWhenFirstBatchUnderSize(t *testing.T) {
+	withFastDeadlockRetry(t, 5)
+	calls := 0
+	total, err := deleteInBatches(context.Background(), 5, func(context.Context) (int64, error) {
+		calls++
+		return 0, nil
+	})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if total != 0 || calls != 1 {
+		t.Fatalf("expected total 0 and 1 call, got total=%d calls=%d", total, calls)
+	}
+}
+
+func TestDeleteInBatchesReturnsError(t *testing.T) {
+	withFastDeadlockRetry(t, 5)
+	sentinel := errors.New("delete failed")
+	_, err := deleteInBatches(context.Background(), 5, func(context.Context) (int64, error) {
+		return 0, sentinel
+	})
+	if !errors.Is(err, sentinel) {
+		t.Fatalf("expected sentinel, got %v", err)
+	}
+}
+```
+
+- [ ] **Step 2: Run tests to verify they fail**
+
+Run: `go test ./internal/catalog/ -run TestDeleteInBatches -v`
+Expected: FAIL — `undefined: deleteInBatches`.
+
+- [ ] **Step 3: Implement the helper**
+
+In `internal/catalog/folder_repo.go`, add after `retryOnDeadlock`:
+
+```go
+// deleteInBatches repeatedly runs deleteBatch (each a single autocommit
+// statement) until a batch removes fewer than batchSize rows. Each batch is
+// retried on deadlock. It returns the total number of rows deleted.
+func deleteInBatches(
+	ctx context.Context,
+	batchSize int,
+	deleteBatch func(ctx context.Context) (int64, error),
+) (int64, error) {
+	var total int64
+	for {
+		var affected int64
+		if err := retryOnDeadlock(ctx, func() error {
+			n, e := deleteBatch(ctx)
+			affected = n
+			return e
+		}); err != nil {
+			return total, err
+		}
+		total += affected
+		if affected < int64(batchSize) {
+			return total, nil
+		}
+	}
+}
+```
+
+- [ ] **Step 4: Run tests to verify they pass**
+
+Run: `go test ./internal/catalog/ -run TestDeleteInBatches -v`
+Expected: PASS (3 tests).
+
+- [ ] **Step 5: Format and commit**
+
+```bash
+gofmt -w internal/catalog/folder_repo.go internal/catalog/folder_delete_test.go
+git add internal/catalog/folder_repo.go internal/catalog/folder_delete_test.go
+git commit -m "feat(catalog): add deleteInBatches loop helper"
+```
+
+---
+
+## Task 3: Generalize image-dir helpers to a querier interface
+
+This lets the image-dir helpers run on either a `*pgxpool.Pool` (new batched path) or a `pgx.Tx` (existing `library_repo.go` path), and splits raw collection from filtering so the batched path can collect cheaply per batch and filter once.
+
+**Files:**
+- Modify: `internal/catalog/folder_repo.go:547-635` (the `collectImageDirs` / `filterUnreferencedImageDirs` block)
+
+- [ ] **Step 1: Add the `rowQuerier` interface**
+
+In `internal/catalog/folder_repo.go`, add (near the other type declarations, e.g. just above `collectOrphanIDs`):
+
+```go
+// rowQuerier is satisfied by both *pgxpool.Pool and pgx.Tx, letting read
+// helpers run inside or outside an explicit transaction.
+type rowQuerier interface {
+	Query(ctx context.Context, sql string, args ...any) (pgx.Rows, error)
+}
+```
+
+- [ ] **Step 2: Split `collectImageDirs` into raw collection + filtering**
+
+Replace the existing `collectImageDirs` function (currently `internal/catalog/folder_repo.go:549-584`) with these two functions:
+
+```go
+// collectImageDirs returns S3 directory prefixes for images belonging to the
+// given content IDs that are not still referenced by other surviving content.
+func collectImageDirs(ctx context.Context, q rowQuerier, contentIDs []string) ([]string, error) {
+	dirs, err := collectRawImageDirs(ctx, q, contentIDs)
+	if err != nil {
+		return nil, err
+	}
+	return filterUnreferencedImageDirs(ctx, q, dirs, contentIDs)
+}
+
+// collectRawImageDirs returns the deduped S3 directory prefixes referenced by
+// the given content IDs (items, their seasons, and their episodes), without
+// filtering out dirs still used by other content.
+func collectRawImageDirs(ctx context.Context, q rowQuerier, contentIDs []string) ([]string, error) {
+	imgRows, err := q.Query(ctx, `
+		SELECT poster_path, backdrop_path, logo_path FROM media_items WHERE content_id = ANY($1)
+		UNION ALL
+		SELECT poster_path, '', '' FROM seasons WHERE series_id = ANY($1)
+		UNION ALL
+		SELECT still_path, '', '' FROM episodes WHERE series_id = ANY($1)
+	`, contentIDs)
+	if err != nil {
+		return nil, fmt.Errorf("collecting image paths: %w", err)
+	}
+	defer imgRows.Close()
+	dirSet := make(map[string]struct{})
+	for imgRows.Next() {
+		var p1, p2, p3 string
+		if err := imgRows.Scan(&p1, &p2, &p3); err != nil {
+			return nil, fmt.Errorf("scanning image path: %w", err)
+		}
+		for _, p := range []string{p1, p2, p3} {
+			if p != "" && !strings.Contains(p, "://") {
+				if dir := pathDir(p); dir != "" {
+					dirSet[dir] = struct{}{}
+				}
+			}
+		}
+	}
+	if err := imgRows.Err(); err != nil {
+		return nil, fmt.Errorf("iterating image paths: %w", err)
+	}
+	dirs := make([]string, 0, len(dirSet))
+	for dir := range dirSet {
+		dirs = append(dirs, dir)
+	}
+	return dirs, nil
+}
+```
+
+- [ ] **Step 3: Change `filterUnreferencedImageDirs` to take `rowQuerier`**
+
+In `internal/catalog/folder_repo.go`, change the signature (currently `internal/catalog/folder_repo.go:586`) from:
+
+```go
+func filterUnreferencedImageDirs(ctx context.Context, tx pgx.Tx, dirs, deletingContentIDs []string) ([]string, error) {
+```
+
+to:
+
+```go
+func filterUnreferencedImageDirs(ctx context.Context, q rowQuerier, dirs, deletingContentIDs []string) ([]string, error) {
+```
+
+and change the single `tx.Query(ctx, ...)` call inside it to `q.Query(ctx, ...)`. Leave the SQL and the rest of the body unchanged.
+
+- [ ] **Step 4: Verify the build (proves `library_repo.go` still compiles with `pgx.Tx`)**
+
+Run: `go build ./... && go vet ./internal/catalog/`
+Expected: no errors. (`library_repo.go` passes a `pgx.Tx` to `collectImageDirs`, which now accepts `rowQuerier`; `pgx.Tx` satisfies it.)
+
+- [ ] **Step 5: Run the catalog tests**
+
+Run: `go test ./internal/catalog/...`
+Expected: PASS (existing tests plus Tasks 1–2 helpers).
+
+- [ ] **Step 6: Format and commit**
+
+```bash
+gofmt -w internal/catalog/folder_repo.go
+git add internal/catalog/folder_repo.go
+git commit -m "refactor(catalog): make image-dir helpers querier-agnostic"
+```
+
+---
+
+## Task 4: Rewrite `DeleteWithStats` as a batched, phased operation
+
+**Files:**
+- Modify: `internal/catalog/folder_repo.go:415-511` (the `DeleteWithStats` body) and add two helpers + two constants.
+
+- [ ] **Step 1: Add the batch-size constants**
+
+In `internal/catalog/folder_repo.go`, add near the retry vars from Task 1:
+
+```go
+const (
+	// orphanDeleteBatch is small because each media_items row cascades across
+	// ~15 child tables.
+	orphanDeleteBatch = 1000
+	// folderChildDeleteBatch covers the lighter folder-scoped media_files and
+	// junction deletes.
+	folderChildDeleteBatch = 5000
+)
+```
+
+- [ ] **Step 2: Replace the `DeleteWithStats` body**
+
+Replace the entire existing `DeleteWithStats` function (`internal/catalog/folder_repo.go:415-511`) with:
+
+```go
+func (r *FolderRepository) DeleteWithStats(
+	ctx context.Context,
+	id int,
+	progress func(current, total int, message string),
+) (*DeleteFolderStats, error) {
+	stats := &DeleteFolderStats{}
+
+	// Phase 0: preflight reads (no long-lived transaction).
+	if err := r.pool.QueryRow(ctx, `SELECT name FROM media_folders WHERE id = $1`, id).Scan(&stats.LibraryName); err != nil {
+		if errors.Is(err, pgx.ErrNoRows) {
+			return nil, ErrFolderNotFound
+		}
+		return nil, fmt.Errorf("loading folder before delete: %w", err)
+	}
+	if err := r.pool.QueryRow(ctx, `SELECT COUNT(*) FROM media_files WHERE media_folder_id = $1`, id).Scan(&stats.MediaFiles); err != nil {
+		return nil, fmt.Errorf("counting media files: %w", err)
+	}
+	if err := r.pool.QueryRow(ctx, `SELECT COUNT(*) FROM media_item_libraries WHERE media_folder_id = $1`, id).Scan(&stats.MediaItemLinks); err != nil {
+		return nil, fmt.Errorf("counting media item links: %w", err)
+	}
+	var orphanTotal int
+	if err := r.pool.QueryRow(ctx, `
+		SELECT COUNT(*)
+		FROM media_item_libraries mil
+		WHERE mil.media_folder_id = $1
+		AND NOT EXISTS (
+			SELECT 1 FROM media_item_libraries other
+			WHERE other.content_id = mil.content_id
+			AND other.media_folder_id <> $1
+		)`, id).Scan(&orphanTotal); err != nil {
+		return nil, fmt.Errorf("counting orphaned items: %w", err)
+	}
+
+	// Phase 1: delete orphaned media_items in detect-then-delete batches.
+	// Cascade removes their junctions, provider IDs, episodes, seasons, etc.
+	if progress != nil {
+		progress(0, orphanTotal, "Deleting orphaned items")
+	}
+	rawDirs := make(map[string]struct{})
+	for {
+		ids, err := r.collectOrphanBatch(ctx, id, orphanDeleteBatch)
+		if err != nil {
+			return nil, err
+		}
+		if len(ids) == 0 {
+			break
+		}
+		dirs, err := collectRawImageDirs(ctx, r.pool, ids)
+		if err != nil {
+			return nil, err
+		}
+		for _, d := range dirs {
+			rawDirs[d] = struct{}{}
+		}
+		if err := retryOnDeadlock(ctx, func() error {
+			_, e := r.pool.Exec(ctx, `DELETE FROM media_items WHERE content_id = ANY($1)`, ids)
+			return e
+		}); err != nil {
+			return nil, fmt.Errorf("deleting orphaned items: %w", err)
+		}
+		stats.OrphanedItems += len(ids)
+		if progress != nil {
+			progress(stats.OrphanedItems, orphanTotal, "Deleting orphaned items")
+		}
+	}
+
+	// Filter accumulated image dirs once, now that orphans are gone, against
+	// any surviving content. Empty deleting-set means "exclude nothing".
+	if len(rawDirs) > 0 {
+		filtered, err := filterUnreferencedImageDirs(ctx, r.pool, dirSetToSlice(rawDirs), []string{})
+		if err != nil {
+			return nil, err
+		}
+		stats.OrphanedImageDirs = filtered
+	}
+
+	// Phase 2: delete this folder's media_files (folder-tied, not item-tied).
+	if progress != nil {
+		progress(orphanTotal, orphanTotal, "Deleting media files")
+	}
+	if _, err := deleteInBatches(ctx, folderChildDeleteBatch, func(ctx context.Context) (int64, error) {
+		tag, e := r.pool.Exec(ctx, `
+			DELETE FROM media_files
+			WHERE id IN (
+				SELECT id FROM media_files WHERE media_folder_id = $1 LIMIT $2
+			)`, id, folderChildDeleteBatch)
+		if e != nil {
+			return 0, e
+		}
+		return tag.RowsAffected(), nil
+	}); err != nil {
+		return nil, fmt.Errorf("deleting media files: %w", err)
+	}
+
+	// Phase 3: delete remaining folder memberships (shared items kept; only the
+	// membership in this folder is removed).
+	if progress != nil {
+		progress(orphanTotal, orphanTotal, "Removing library memberships")
+	}
+	if _, err := deleteInBatches(ctx, folderChildDeleteBatch, func(ctx context.Context) (int64, error) {
+		tag, e := r.pool.Exec(ctx, `
+			DELETE FROM media_item_libraries
+			WHERE ctid IN (
+				SELECT ctid FROM media_item_libraries WHERE media_folder_id = $1 LIMIT $2
+			)`, id, folderChildDeleteBatch)
+		if e != nil {
+			return 0, e
+		}
+		return tag.RowsAffected(), nil
+	}); err != nil {
+		return nil, fmt.Errorf("deleting media item links: %w", err)
+	}
+
+	// Phase 4: delete the now-lightweight folder row. Tolerate 0 rows so a
+	// resumed run that already removed it still succeeds.
+	if err := retryOnDeadlock(ctx, func() error {
+		_, e := r.pool.Exec(ctx, `DELETE FROM media_folders WHERE id = $1`, id)
+		return e
+	}); err != nil {
+		return nil, fmt.Errorf("deleting folder: %w", err)
+	}
+
+	if progress != nil {
+		progress(orphanTotal, orphanTotal, "Library deletion completed")
+	}
+	return stats, nil
+}
+
+// collectOrphanBatch returns up to limit content IDs whose only library
+// membership is the given folder.
+func (r *FolderRepository) collectOrphanBatch(ctx context.Context, folderID, limit int) ([]string, error) {
+	rows, err := r.pool.Query(ctx, `
+		SELECT mil.content_id
+		FROM media_item_libraries mil
+		WHERE mil.media_folder_id = $1
+		AND NOT EXISTS (
+			SELECT 1 FROM media_item_libraries other
+			WHERE other.content_id = mil.content_id
+			AND other.media_folder_id <> $1
+		)
+		LIMIT $2`, folderID, limit)
+	if err != nil {
+		return nil, fmt.Errorf("finding orphaned items: %w", err)
+	}
+	defer rows.Close()
+	var ids []string
+	for rows.Next() {
+		var contentID string
+		if err := rows.Scan(&contentID); err != nil {
+			return nil, fmt.Errorf("scanning orphan content_id: %w", err)
+		}
+		ids = append(ids, contentID)
+	}
+	return ids, rows.Err()
+}
+
+// dirSetToSlice returns the keys of set as a slice (nil if empty).
+func dirSetToSlice(set map[string]struct{}) []string {
+	if len(set) == 0 {
+		return nil
+	}
+	dirs := make([]string, 0, len(set))
+	for d := range set {
+		dirs = append(dirs, d)
+	}
+	return dirs
+}
+```
+
+Note: this removes the old single-transaction body (`tx, err := r.pool.Begin(...)`, the inline orphan-collection loop, the `collectImageDirs` call, the CASCADE folder drop, and `tx.Commit`). `collectOrphanIDs` and `Delete` are untouched. `collectImageDirs` remains (still used by `library_repo.go`).
+
+- [ ] **Step 3: Verify the build and vet**
+
+Run: `go build ./... && go vet ./internal/catalog/`
+Expected: no errors. If `go vet` flags an unused function, confirm it is genuinely unused before removing — `collectImageDirs`, `collectOrphanIDs`, and `pathDir` must all remain (still referenced).
+
+- [ ] **Step 4: Run the full catalog + adminjob test suites**
+
+Run: `go test ./internal/catalog/... ./internal/adminjob/...`
+Expected: PASS. (No DB-backed test exercises `DeleteWithStats`; the helper unit tests and existing tests must stay green.)
+
+- [ ] **Step 5: Lint**
+
+Run: `golangci-lint run ./internal/catalog/...`
+Expected: no findings. (If `golangci-lint` is unavailable, run `gofmt -l internal/catalog/` and ensure it prints nothing.)
+
+- [ ] **Step 6: Format and commit**
+
+```bash
+gofmt -w internal/catalog/folder_repo.go
+git add internal/catalog/folder_repo.go
+git commit -m "fix(catalog): delete libraries in deadlock-retrying batches
+
+Replaces the single multi-minute delete transaction with phased, batched
+autocommit deletes (orphan items, media files, memberships, folder row),
+each retried on deadlock. Holds only short locks, survives concurrent
+writers, and is resumable on failure."
+```
+
+---
+
+## Task 5: Verify on the dev server
+
+This is the integration test (no catalog DB harness exists). Follow the deployment-debugging runbook. The Audiobooks library (folder id 8) is currently stuck with ~249K items / ~339K files.
+
+- [ ] **Step 1: Deploy the branch to dev**
+
+Run: `make dev-deploy`
+Expected: builds `silo:dev-local`, recreates the `silo` service. Then confirm health:
+`ssh "$DEV_HOST" 'curl -s http://localhost:8090/api/v1/ready'`
+
+- [ ] **Step 2: Capture the starting counts**
+
+Run (psql via the dev compose `postgres` service, user/db `continuum`):
+
+```sql
+SELECT
+  (SELECT COUNT(*) FROM media_files WHERE media_folder_id = 8) AS files,
+  (SELECT COUNT(*) FROM media_item_libraries WHERE media_folder_id = 8) AS links,
+  (SELECT EXISTS (SELECT 1 FROM media_folders WHERE id = 8)) AS folder_exists;
+```
+
+Expected: non-zero files/links, `folder_exists = t`.
+
+- [ ] **Step 3: Trigger the deletion**
+
+Re-trigger the library deletion through the normal path (admin UI "delete library" for Audiobooks, or the admin delete endpoint). Confirm a new job is queued:
+
+```sql
+SELECT id, job_type, status, requested_at
+FROM admin_jobs WHERE job_type = 'delete_library'
+ORDER BY requested_at DESC LIMIT 3;
+```
+
+- [ ] **Step 4: Monitor to completion**
+
+Watch the job and the draining counts (re-run periodically):
+
+```sql
+SELECT id, status, progress_current, progress_total, error_message, heartbeat_at
+FROM admin_jobs WHERE job_type = 'delete_library'
+ORDER BY requested_at DESC LIMIT 1;
+```
+
+```sql
+SELECT COUNT(*) FROM media_item_libraries WHERE media_folder_id = 8;
+```
+
+Also confirm deadlock retries (if any) are recovered rather than fatal:
+`ssh "$DEV_HOST" "$DEV_COMPOSE_CMD logs postgres 2>&1 | grep -i deadlock | tail -20"`
+Expected: the job reaches `status = completed`; counts trend to 0; any deadlocks are transient (the job does not fail).
+
+- [ ] **Step 5: Confirm full removal**
+
+```sql
+SELECT
+  (SELECT COUNT(*) FROM media_files WHERE media_folder_id = 8) AS files,
+  (SELECT COUNT(*) FROM media_item_libraries WHERE media_folder_id = 8) AS links,
+  (SELECT EXISTS (SELECT 1 FROM media_folders WHERE id = 8)) AS folder_exists;
+```
+
+Expected: `files = 0`, `links = 0`, `folder_exists = f`.
+
+- [ ] **Step 6: Report**
+
+Summarize: starting counts, time to completion, number of batches/retries observed, and final state. No commit (verification only).
+
+---
+
+## Self-Review
+
+**Spec coverage:**
+- Phased batched deletion (Phases 0–4) → Task 4. ✔
+- `retryOnDeadlock` (40P01/40001, 5 attempts, backoff, ctx-aware) → Task 1. ✔
+- Batch sizes `orphanDeleteBatch=1000`, `folderChildDeleteBatch=5000` → Task 4 Step 1. ✔
+- Autocommit per batch / no long transaction → Task 4 Step 2 (uses `r.pool.Exec`, no `Begin`). ✔
+- Stats + progress + heartbeat reliance → Task 4 Step 2; runner heartbeat unchanged. ✔
+- Resumability (folder row last, tolerate 0 rows) → Task 4 Phase 4. ✔
+- Unit tests for retry + batch loop → Tasks 1–2. ✔
+- `collectImageDirs` shared with `library_repo.go` preserved → Task 3 (interface widening), Task 4 Step 3 build check. ✔
+- Dev integration verification → Task 5. ✔
+- Out of scope (writer coordination, runner/executor changes, schema changes) → not present in any task. ✔
+
+**Placeholder scan:** No TBD/TODO; every code step has complete code; every command has expected output. ✔
+
+**Type consistency:** `retryOnDeadlock(ctx, func() error) error`, `deleteInBatches(ctx, int, func(ctx) (int64,error)) (int64,error)`, `rowQuerier{ Query(ctx,string,...any)(pgx.Rows,error) }`, `collectRawImageDirs`/`collectImageDirs`/`filterUnreferencedImageDirs(ctx, rowQuerier, ...)`, `collectOrphanBatch(ctx,int,int)([]string,error)`, `dirSetToSlice(map[string]struct{})[]string`, consts `orphanDeleteBatch`/`folderChildDeleteBatch`, vars `deadlockMaxAttempts`/`deadlockBaseBackoff` — names used consistently across tasks. ✔