diff --git a/cmd/ingestor/tx_last_seen_backfill.go b/cmd/ingestor/tx_last_seen_backfill.go
new file mode 100644
index 00000000..5d5a383e
--- /dev/null
+++ b/cmd/ingestor/tx_last_seen_backfill.go
@@ -0,0 +1,31 @@
+// Package main: chunked tx_last_seen_v1 backfill (issue #1724).
+//
+// Stub for the RED test commit — returns zero counts and nil error so the
+// test compiles + runs to assertions (which then fail, proving the gate).
+// The GREEN commit replaces this body with the real chunked implementation.
+
+package main
+
+import (
+	"context"
+	"database/sql"
+	"time"
+)
+
+type txBackfillProgressFn func(processed, total int64)
+
+func chunkedTxLastSeenBackfill(
+	ctx context.Context,
+	db *sql.DB,
+	batchSize int,
+	yieldDelay time.Duration,
+	progress txBackfillProgressFn,
+) (processed int64, total int64, err error) {
+	// Intentional no-op stub: real implementation lands in the GREEN commit.
+	_ = ctx
+	_ = db
+	_ = batchSize
+	_ = yieldDelay
+	_ = progress
+	return 0, 0, nil
+}
diff --git a/cmd/ingestor/tx_last_seen_backfill_test.go b/cmd/ingestor/tx_last_seen_backfill_test.go
new file mode 100644
index 00000000..b0cf21b3
--- /dev/null
+++ b/cmd/ingestor/tx_last_seen_backfill_test.go
@@ -0,0 +1,322 @@
+// Tests for the chunked tx_last_seen backfill (issue #1724).
+//
+// These tests pin the contract the operator-scale fix MUST honor:
+//
+//  - The loop actually yields the writer between batches (concurrent
+//    reader gets through in bounded latency — NOT just "chunks happen",
+//    which a single-tx fake could satisfy).
+//  - With seedN=12000 + batchSize=5000 the progress callback fires
+//    multiple times (≥3 including terminal) — proves chunking happened.
+//  - ctx cancellation mid-loop returns context.Canceled, partial rows
+//    are committed (visible in the DB), no panic.
+//  - Concurrent INSERT of new last_seen=0 rows during the loop does
+//    NOT cause infinite iteration (maxID snapshot bounds the scan).
+//  - Errors from inner queries (RowsAffected via driver poisoning) and
+//    parameter validation propagate — migration does not silently mark done.
+//  - Orphan transmissions (no matching observations) do NOT trap the
+//    loop in an infinite n=0/n=0 ping-pong: the EXISTS filter skips them.
+
+package main
+
+import (
+	"context"
+	"database/sql"
+	"fmt"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// seedTransmissions inserts n transmissions each with one observation,
+// last_seen=0. Returns when done. Uses the store's raw db.
+func seedTransmissions(t *testing.T, s *Store, n int) {
+	t.Helper()
+	// Wait for the boot-time async migrations to release the single writer
+	// connection before we start a big seed — they're harmless on the empty
+	// store but they race for SetMaxOpenConns(1).
+	s.WaitForAsyncMigrations()
+	tx, err := s.db.Begin()
+	if err != nil {
+		t.Fatalf("seed begin: %v", err)
+	}
+	for i := 0; i < n; i++ {
+		res, err := tx.Exec(`
+			INSERT INTO transmissions (raw_hex, hash, first_seen, route_type, payload_type, payload_version, decoded_json, last_seen)
+			VALUES (?, ?, datetime('now'), 2, 2, 0, '{}', 0)
+		`, "deadbeef", randHex(i))
+		if err != nil {
+			t.Fatalf("seed tx %d: %v", i, err)
+		}
+		id, _ := res.LastInsertId()
+		// Observer row + observation referencing this tx, with a stable timestamp.
+		ts := int64(1_700_000_000 + i)
+		if _, err := tx.Exec(`
+			INSERT INTO observations (transmission_id, observer_idx, timestamp, snr, rssi, score, path_json)
+			VALUES (?, NULL, ?, 0, 0, 0, '[]')
+		`, id, ts); err != nil {
+			t.Fatalf("seed obs %d: %v", i, err)
+		}
+	}
+	if err := tx.Commit(); err != nil {
+		t.Fatalf("seed commit: %v", err)
+	}
+}
+
+func randHex(i int) string {
+	return fmt.Sprintf("%016x", uint64(i)*0x9E3779B97F4A7C15+1)
+}
+
+// TestChunkedBackfill_YieldsToReaderBetweenBatches asserts that a concurrent
+// reader gets through in bounded latency while the backfill is running. A
+// fake single-transaction implementation would NOT satisfy this — the
+// reader would queue behind sqlite_busy_timeout for the full duration.
+func TestChunkedBackfill_YieldsToReaderBetweenBatches(t *testing.T) {
+	s := newTestStore(t)
+	seedTransmissions(t, s, 12_000)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	// Background backfill (modest yield delay so the test is bounded).
+	backfillDone := make(chan error, 1)
+	go func() {
+		_, _, err := chunkedTxLastSeenBackfill(ctx, s.db, 2000, 50*time.Millisecond, nil)
+		backfillDone <- err
+	}()
+
+	// Give the backfill a moment to actually start its first chunk.
+	time.Sleep(20 * time.Millisecond)
+
+	// Concurrent reader: must succeed in bounded time. A single-tx
+	// implementation that holds the writer for the full duration would
+	// either time out or take seconds to acquire the lock.
+	readDeadline := time.Now().Add(2 * time.Second)
+	var readLatency time.Duration
+	readStart := time.Now()
+	var rowCount int64
+	for time.Now().Before(readDeadline) {
+		queryStart := time.Now()
+		row := s.db.QueryRowContext(ctx, `SELECT COUNT(*) FROM transmissions`)
+		var c int64
+		if err := row.Scan(&c); err != nil {
+			t.Fatalf("reader scan: %v", err)
+		}
+		readLatency = time.Since(queryStart)
+		rowCount = c
+		if readLatency < 500*time.Millisecond {
+			// Got a fast read while backfill running → good signal.
+			break
+		}
+	}
+	t.Logf("reader latency: %v, count=%d, total wall=%v", readLatency, rowCount, time.Since(readStart))
+	if readLatency > 500*time.Millisecond {
+		t.Errorf("reader latency=%v exceeded bound — backfill is not yielding the writer", readLatency)
+	}
+
+	// Backfill should complete.
+	select {
+	case err := <-backfillDone:
+		if err != nil {
+			t.Fatalf("backfill error: %v", err)
+		}
+	case <-time.After(25 * time.Second):
+		t.Fatalf("backfill did not complete")
+	}
+
+	// Verify all rows got populated.
+	var remaining int64
+	if err := s.db.QueryRow(`SELECT COUNT(*) FROM transmissions WHERE last_seen = 0`).Scan(&remaining); err != nil {
+		t.Fatalf("post-check: %v", err)
+	}
+	if remaining != 0 {
+		t.Errorf("remaining last_seen=0 rows: %d, want 0", remaining)
+	}
+}
+
+// TestChunkedBackfill_MinBatchCount asserts the progress callback fires
+// at least 3 times for seedN=12000 / batchSize=5000 (≥2 in-flight chunks
+// + 1 terminal). A single-tx fake would fire 1 or 2 times only.
+func TestChunkedBackfill_MinBatchCount(t *testing.T) {
+	s := newTestStore(t)
+	seedTransmissions(t, s, 12_000)
+
+	var callbacks int64
+	progress := func(processed, total int64) {
+		atomic.AddInt64(&callbacks, 1)
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+	processed, total, err := chunkedTxLastSeenBackfill(ctx, s.db, 5000, 5*time.Millisecond, progress)
+	if err != nil {
+		t.Fatalf("backfill: %v", err)
+	}
+	if total != 12_000 {
+		t.Errorf("total=%d, want 12000", total)
+	}
+	if processed != 12_000 {
+		t.Errorf("processed=%d, want 12000", processed)
+	}
+	got := atomic.LoadInt64(&callbacks)
+	if got < 3 {
+		t.Errorf("progress callbacks=%d, want >=3 (proves real chunking)", got)
+	}
+}
+
+// TestChunkedBackfill_CtxCancelMidLoop cancels ctx between batches and asserts
+// the function returns context.Canceled, partial rows are committed (visible
+// in DB), and no panic occurs.
+func TestChunkedBackfill_CtxCancelMidLoop(t *testing.T) {
+	s := newTestStore(t)
+	seedTransmissions(t, s, 10_000)
+
+	ctx, cancel := context.WithCancel(context.Background())
+
+	// Cancel shortly after the first batch starts.
+	go func() {
+		time.Sleep(30 * time.Millisecond)
+		cancel()
+	}()
+
+	_, _, err := chunkedTxLastSeenBackfill(ctx, s.db, 1000, 80*time.Millisecond, nil)
+	if err == nil {
+		t.Fatalf("expected context error, got nil")
+	}
+	if err != context.Canceled && ctx.Err() != context.Canceled {
+		t.Fatalf("expected context.Canceled, got %v", err)
+	}
+
+	var done int64
+	if err := s.db.QueryRow(`SELECT COUNT(*) FROM transmissions WHERE last_seen != 0`).Scan(&done); err != nil {
+		t.Fatalf("post-check: %v", err)
+	}
+	if done == 0 {
+		t.Errorf("expected partial commits (>0), got 0 — backfill is not committing per chunk")
+	}
+	if done >= 10_000 {
+		t.Errorf("cancel did not take effect, all %d rows committed", done)
+	}
+}
+
+// TestChunkedBackfill_ConcurrentInsertTerminates inserts new last_seen=0
+// rows during the loop. The maxID snapshot must bound the scan so the loop
+// terminates.
+func TestChunkedBackfill_ConcurrentInsertTerminates(t *testing.T) {
+	s := newTestStore(t)
+	seedTransmissions(t, s, 3000)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Second)
+	defer cancel()
+
+	// Concurrent inserter: keeps adding new last_seen=0 rows in tight loop.
+	stopInserts := make(chan struct{})
+	insertsDone := make(chan struct{})
+	go func() {
+		defer close(insertsDone)
+		i := 100_000
+		for {
+			select {
+			case <-stopInserts:
+				return
+			default:
+			}
+			_, err := s.db.Exec(`
+				INSERT INTO transmissions (raw_hex, hash, first_seen, route_type, payload_type, payload_version, decoded_json, last_seen)
+				VALUES ('aa', ?, datetime('now'), 2, 2, 0, '{}', 0)
+			`, randHex(i))
+			if err != nil {
+				return
+			}
+			id, _ := func() (int64, error) {
+				var lid int64
+				err := s.db.QueryRow(`SELECT last_insert_rowid()`).Scan(&lid)
+				return lid, err
+			}()
+			_, _ = s.db.Exec(`
+				INSERT INTO observations (transmission_id, observer_idx, timestamp, snr, rssi, score, path_json)
+				VALUES (?, NULL, ?, 0, 0, 0, '[]')
+			`, id, int64(1_700_000_000+i))
+			i++
+			time.Sleep(time.Millisecond)
+		}
+	}()
+
+	_, _, err := chunkedTxLastSeenBackfill(ctx, s.db, 500, 5*time.Millisecond, nil)
+	close(stopInserts)
+	<-insertsDone
+
+	if err != nil {
+		t.Fatalf("backfill should terminate, got err=%v", err)
+	}
+}
+
+// TestChunkedBackfill_ParamValidation: batchSize<=0 must reject (no <0 sentinel).
+func TestChunkedBackfill_ParamValidation(t *testing.T) {
+	s := newTestStore(t)
+	s.WaitForAsyncMigrations()
+	ctx := context.Background()
+	if _, _, err := chunkedTxLastSeenBackfill(ctx, s.db, 0, 100*time.Millisecond, nil); err == nil {
+		t.Errorf("batchSize=0 must error")
+	}
+	if _, _, err := chunkedTxLastSeenBackfill(ctx, s.db, -1, 100*time.Millisecond, nil); err == nil {
+		t.Errorf("batchSize=-1 must error")
+	}
+	if _, _, err := chunkedTxLastSeenBackfill(ctx, s.db, 100, -1, nil); err == nil {
+		t.Errorf("negative yieldDelay must error")
+	}
+}
+
+// TestChunkedBackfill_OrphanTxTerminates: transmission row with no matching
+// observation must NOT trap the loop. Using EXISTS in the WHERE clause skips
+// the row; the chunk returns n=0 after eligible rows are exhausted; loop ends.
+func TestChunkedBackfill_OrphanTxTerminates(t *testing.T) {
+	s := newTestStore(t)
+	s.WaitForAsyncMigrations()
+	// One orphan tx (no observation row at all)
+	if _, err := s.db.Exec(`
+		INSERT INTO transmissions (raw_hex, hash, first_seen, route_type, payload_type, payload_version, decoded_json, last_seen)
+		VALUES ('00', 'orphan', datetime('now'), 2, 2, 0, '{}', 0)
+	`); err != nil {
+		t.Fatalf("seed orphan: %v", err)
+	}
+	// Plus one normal tx so total>0.
+	seedTransmissions(t, s, 5)
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	processed, _, err := chunkedTxLastSeenBackfill(ctx, s.db, 10, 5*time.Millisecond, nil)
+	if err != nil {
+		t.Fatalf("backfill: %v", err)
+	}
+	// Only the 5 non-orphan rows should be touched; orphan retains last_seen=0.
+	if processed != 5 {
+		t.Errorf("processed=%d, want 5 (orphan must be skipped)", processed)
+	}
+	var orphanLs int64
+	_ = s.db.QueryRow(`SELECT last_seen FROM transmissions WHERE hash = 'orphan'`).Scan(&orphanLs)
+	if orphanLs != 0 {
+		t.Errorf("orphan tx last_seen=%d, want 0 (no observation to source MAX from)", orphanLs)
+	}
+}
+
+// TestChunkedBackfill_ErrorPropagation_BadDB: a closed DB must surface as
+// an error — migration must NOT silently report success.
+func TestChunkedBackfill_ErrorPropagation_BadDB(t *testing.T) {
+	s := newTestStore(t)
+	s.WaitForAsyncMigrations()
+	// Open a second handle to the same path then close it — using a fresh
+	// closed *sql.DB gives a deterministic ExecContext failure.
+	bad, err := sql.Open("sqlite", ":memory:")
+	if err != nil {
+		t.Fatalf("open: %v", err)
+	}
+	bad.Close()
+	ctx := context.Background()
+	if _, _, err := chunkedTxLastSeenBackfill(ctx, bad, 100, 0, nil); err == nil {
+		t.Errorf("closed DB must produce an error; got nil — migration would silently mark done")
+	}
+	// Sanity: live store still works (no global mutation).
+	if _, _, err := chunkedTxLastSeenBackfill(ctx, s.db, 100, 0, nil); err != nil {
+		t.Fatalf("sanity: live store backfill should succeed, got %v", err)
+	}
+}