diff --git a/CHANGELOG.md b/CHANGELOG.md
index cd56291..69b7a63 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -5,6 +5,23 @@ Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/); version
 
 ## [Unreleased]
 
+### Fixed
+
+- Log-tailer resumes where it stopped instead of jumping to end-of-file on
+  every start (phase 3, `docs/code-review.md`): the read position and a
+  fingerprint of the log's head are persisted (`logtail_state`, migration
+  `0003`), so delivery lines written while the panel was down are parsed and
+  their send-log rows no longer stay `queued` forever. A log that changed
+  identity while the panel was down is read from the start; a first-ever start,
+  with nothing stored, still begins at the end. Container recreate remains a
+  gap — `mail.log` is not in `/data` (`docs/security.md`).
+- Level-2 rate limit no longer overshoots under concurrency: messages that
+  passed the check at MAIL FROM but have not reached the send log yet are
+  counted alongside the stored rows (`internal/milter/inflight.go`), so
+  parallel SMTP sessions cannot each spend the same last slot. Slots are
+  released at end-of-message, on ABORT, and after a 10-minute TTL, so a client
+  that drops mid-transaction cannot hold one — the limiter stays fail-open.
+
 ### Changed
 
 - Phase 1 doc/code hygiene (`docs/code-review.md`): removed ~30 stale
diff --git a/docs/architecture.md b/docs/architecture.md
index 7a9c6ad..0d05215 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -81,7 +81,11 @@ One process, three roles:
    proxy only.
 2. **journal-milter** — unix socket `JOURNAL_MILTER_SOCKET`; records From/To/
    Subject/SASL user at DATA; enforces level-2 rate limits; **fail-open**
-   (`default_action=accept`) so milter failure does not stop mail.
+   (`default_action=accept`) so milter failure does not stop mail. The level-2
+   count is the stored send-log rows plus the messages this process has admitted
+   but not yet written (`internal/milter/inflight.go`), so concurrent sessions
+   cannot each spend the same last slot; a reservation is released at
+   end-of-message, on ABORT, or after a 10-minute TTL.
 3. **log-tailer** — follows `MAIL_LOG`, updates send-log delivery status by
    queue-id. Send-log `queued → sent` transitions depend on this goroutine alone
    (`UpdateStatus` is only called from [internal/logtail](../internal/logtail/logtail.go)).
@@ -99,15 +103,23 @@ panel user cannot read). `follow()` drains the old inode once more before
 switching descriptors; the panel treats a missing log file as an empty tail, not
 an error.
 
-**Known gaps (same class of loss, not fixed by rename rotation):**
+**Read offset is persisted** (`logtail_state` table, migration `0003`): the
+tailer stores its position plus a fingerprint of the log's first 512 bytes, and
+on start resumes from it, parsing the tail written while the panel was down. If
+the fingerprint no longer matches (rotated or recreated in the meantime) it reads
+the current file from the start; re-parsing lines is harmless because
+`UpdateStatus` writes the same status onto the same row. Only a first-ever start,
+with nothing stored, begins at end-of-file, so installing the panel does not
+replay a pre-existing log.
+
+**Remaining gap:**
 
-- **Panel restart** — `follow()` starts at end-of-file; lines written while the
-  panel was down are never parsed; in-flight send-log rows may stay `queued`.
 - **Container recreate** — `/var/log` is ephemeral; the log is lost with the
-  container.
+  container, so the delivery lines for rows still `queued` are gone with it and
+  those rows stay `queued` forever.
 
-Possible follow-ups if these become painful: persist read offset across restarts,
-mount mail log under `/data`, or reconcile stuck rows via `postqueue`.
+Possible follow-ups if this becomes painful: mount the mail log under `/data`, or
+reconcile stuck rows via `postqueue`.
 
 ---
 
@@ -158,7 +170,7 @@ cookie and idle timeout has not expired.
 
 | Path | Contents |
 |---|---|
-| `selfpost.db` | SQLite: domains, apps, admin, sessions, send log, L2 limits |
+| `selfpost.db` | SQLite: domains, apps, admin, sessions, send log, L2 limits, log-tailer offset |
 | `setup-token` | First-run setup token file |
 | `opendkim/` | DKIM keys + tables |
 | `sasl/sasldb2` | Application SASL credentials |
diff --git a/docs/code-review.md b/docs/code-review.md
index 62e27f7..f96406b 100644
--- a/docs/code-review.md
+++ b/docs/code-review.md
@@ -72,7 +72,7 @@ flowchart TB
 - **Layered / ports-and-adapters:** handlers → services (`domain`, `app`) → `store`; инфраструктура изолирована в адаптерах ([`internal/web/web.go`](../internal/web/web.go), [`internal/app/service.go`](../internal/app/service.go)).
 - **Composition root** в [`cmd/panel/main.go`](../cmd/panel/main.go): три роли (HTTP, journal-milter, log-tailer) в одном процессе — оправдано для single-container deployment.
 - **Interface seams** для тестов: `milter.Store`, `app.SenderMaps`, `logtail.StatusStore`.
-- **Embedded migrations** ([`internal/store/store.go`](../internal/store/store.go)) — простой, надёжный подход для 2 миграций.
+- **Embedded migrations** ([`internal/store/store.go`](../internal/store/store.go)) — простой, надёжный подход для 3 миграций.
 - **E2E как отдельный модуль** (`test/e2e/go.mod`) — не загрязняет основной модуль.
 
 ### Замечания (не блокеры)
@@ -215,7 +215,7 @@ E2E покрывает: bootstrap, SMTP AUTH, DKIM, send-log lifecycle, negative
 
 | Gap | Описание | Документировано |
 |-----|----------|-----------------|
-| Send-log `queued` forever | Log-tailer стартует с EOF; после restart пропущенный хвост не дочитывается | [security.md](security.md), [roadmap.md](roadmap.md) |
+| Send-log `queued` forever | **Закрыто для рестарта:** Фаза 3 — offset персистится (`logtail_state`), хвост дочитывается. Остаётся пересоздание контейнера: `mail.log` не в `/data` | [security.md](security.md), [roadmap.md](roadmap.md) |
 | CSRF without tokens | POST без Origin/Sec-Fetch-Site пропускается | [security.md](security.md) |
 | Fail-open L2 rate limit | DB error → mail проходит | [`internal/milter/ratelimit.go`](../internal/milter/ratelimit.go) |
 | Shallow SPF check | Не следует `include:`/`redirect=` | README, `internal/dnscheck/spf.go` |
@@ -235,8 +235,8 @@ E2E покрывает: bootstrap, SMTP AUTH, DKIM, send-log lifecycle, negative
 |---|----------|-----------|--------|
 | L1 | **Предрелизный security review** (§ D) — обязательный гейт | **P0** | **Fable** |
 | L2 | **Шифрование бэкапа и экспорта домена** (R13) — optional, checkbox + password — **выполнено** | P1 | **Opus** + Sonnet |
-| L3 | Send-log gap mitigation — опционально | P2 | Opus |
-| L4 | Transaction wrap для rate limit count+insert — опционально | P3 | Opus |
+| L3 | Send-log gap mitigation — **выполнено** (persist offset, Фаза 3) | P2 | Opus |
+| L4 | Rate limit count+insert — **выполнено** (учёт «в полёте», Фаза 3; транзакция как таковая неприменима) | P3 | Opus |
 
 ---
 
@@ -246,6 +246,7 @@ E2E покрывает: bootstrap, SMTP AUTH, DKIM, send-log lifecycle, negative
 
 - [`0001_init.sql`](../internal/store/migrations/0001_init.sql) — initial schema
 - [`0002_sessions.sql`](../internal/store/migrations/0002_sessions.sql) — sessions (plan B.1)
+- [`0003_logtail_state.sql`](../internal/store/migrations/0003_logtail_state.sql) — log-tailer read offset (Фаза 3)
 - Механизм: `PRAGMA user_version`, embedded FS, transactional apply — **чистый**, без legacy branches в коде.
 
 ### Архивная документация
@@ -360,7 +361,7 @@ Go `html/template` + HTMX polling + [`panel.css`](../internal/web/static/panel.c
 |---|--------|--------|
 | R9 | Inbound relay (Phase O1) | **Opus** |
 | R10 | Domain-admin role | Opus |
-| R11 | Send-log gap fix (persist offset) | Opus |
+| R11 | Send-log gap fix (persist offset) — **выполнено в Фазе 3**, из 2.x снято | Opus |
 | R12 | Split internal/web subpackages | Sonnet/Opus |
 
 ### CI/infra
@@ -423,10 +424,22 @@ Go `html/template` + HTMX polling + [`panel.css`](../internal/web/static/panel.c
 2. Sonnet: CSS custom properties для dark mode
 3. Haiku: consolidate main max-width rules
 
-### Фаза 3 — Operational improvements (P2–P3, optional)
+### Фаза 3 — Operational improvements (P2–P3, optional) — **выполнено 2026-08-06**
 
-1. Opus: send-log read offset persistence
-2. Opus: rate limit count transaction wrap
+1. Opus: send-log read offset persistence — **сделано**: `logtail_state`
+   (миграция `0003`) хранит offset + отпечаток головы лога; при совпадении
+   отпечатка чтение продолжается, при несовпадении файл читается с начала,
+   первый запуск (записи нет) — с конца, как раньше.
+2. Opus: rate limit count transaction wrap — **сделано иначе**: буквальная
+   транзакция невозможна, count живёт на MAIL FROM, insert — на end-of-message,
+   это разные стадии SMTP-транзакции. Overshoot закрыт учётом сообщений «в
+   полёте» (`internal/milter/inflight.go`): к счёту из БД добавляются
+   резервации, взятые прошедшими проверку сессиями и снимаемые после записи в
+   send-log, на ABORT или по TTL 10 минут.
+
+Остаток по send-log (не закрывается персистом offset): при пересоздании
+контейнера `mail.log` теряется вместе с ним — принятый риск в
+[security.md](security.md).
 
 ---
 
diff --git a/docs/progress.md b/docs/progress.md
index 7d3321b..00c227f 100644
--- a/docs/progress.md
+++ b/docs/progress.md
@@ -52,6 +52,7 @@
 - **§ D выполнен (2026-08-06):** предрелизная ревизия безопасности моделью Fable — диф от аудита v1.0 (Фаза 11, `bd64e80`) до HEAD + полный проход по чек-листу [security.md](security.md) (бывшее ТЗ 7.6). Эксплуатируемых находок нет; одна правка defence-in-depth (`--` перед логином в argv `saslpasswd2`, `internal/app/sasl.go` + тест). Принятые риски не пополнились. Детали — [implementation-plan.md](implementation-plan.md) § D и CHANGELOG `[Unreleased]/Security`. Локально `go vet`/`go test ./internal/app/...` чистые; падения `internal/domain` (`TestWriteLoadPrivateKeyRoundtrip`, `TestRenderTables`) и `internal/logtail` (`TestFollowTailsAndRotates`) — Windows-специфика (права файлов/`\` в путях/rename открытого файла), на Linux CI зелено.
 - **Фаза 1 выполнена (2026-08-06)** ([code-review.md](code-review.md) § Фаза 1 — doc/code hygiene, P1): cleanup ~30 stale «Phase N» комментариев в коде и shell-скриптах; исправлен stale-комментарий в `handlers_domains.go`; ADR CSRF (Origin vs токены) добавлен в [security.md](security.md); known-limitations по log-tailer уже был в [architecture.md](architecture.md) § Log tailer — отдельного действия не потребовалось; `docs/logo` в [roadmap.md](roadmap.md) закрыт (каталога нет, критерию соответствует); `gofmt -l` добавлен в CI (`.github/workflows/test.yml`). `gofmt`/`go vet`/`go test ./...` чистые в обоих модулях (dev-server).
 - **Фаза 1.5 выполнена (2026-08-06)** ([code-review.md](code-review.md) § Фаза 1.5 — шифрование резервных копий, P1): новый пакет `internal/secretfile` — конверт `magic SELFPOST1 | type | scrypt-параметры | salt | nonce-prefix` + поток 64 KiB чанков AES-256-GCM, каждый с AAD `header+counter+last`, поэтому обрезка, перестановка и подмена не открываются (стриминг в обе стороны — полный бэкап не держится в памяти). Панель: чекбокс «Encrypt with a password» в форме полного бэкапа и экспорта домена (общий партиал `templates/encrypt_fields.html`, показ/очистка полей — `panel.js`, без inline-скриптов), импорт домена принимает `.spde` (шифрование определяется по magic, не по расширению) с полем пароля. CLI `selfpost-backup`: пишет `.spbk` при заданном пароле и умеет `-decrypt` (иначе зашифрованный бэкап нечем распаковать при restore); пароль — только `SELFPOST_BACKUP_PASSWORD` / `-password-file`, никогда argv. Умолчание не изменилось: галочка снята — прежние `.tar.gz` / `.json` байт в байт. Тесты: round-trip по размерам (0, границы чанка, несколько чанков), неверный пароль, обрезка, перестановка чанков, порча байта, чужие KDF-параметры; валидация формы пароля; round-trip CLI create→decrypt→tar. Docs: README § *Encrypting a backup or export*, [security.md](security.md) § «Резервная копия и экспорт домена» + принятый риск (шифрование опционально), [architecture.md](architecture.md) § Persistence. `gofmt`/`go vet`/`go test ./...` чистые (кроме известных Windows-падений `internal/domain`, `internal/logtail`). E2E-сценарий не добавлялся: в `test/e2e/` бэкапа не было и раньше, а прогнать новый тест локально нечем (нет Docker) — кандидат при следующем прогоне на dev-сервере.
+- **Фаза 3 выполнена (2026-08-06)** ([code-review.md](code-review.md) § Фаза 3 — operational improvements, P2–P3): (1) log-tailer сохраняет позицию чтения — таблица `logtail_state` (миграция `0003`, `internal/store/logtail.go`) хранит offset + отпечаток первых 512 байт лога, `internal/logtail/offset.go` решает откуда стартовать: отпечаток совпал → продолжаем с offset (дочитывается хвост, написанный пока панель лежала); не совпал (лог сменился/пересоздан) → читаем файл с начала (повторный разбор безвреден, `UpdateStatus` идемпотентен); записи нет вовсе (первый запуск) → с конца, как раньше. Запись offset — не чаще раза в 5 с, плюс форс при ротации и на выключении; сохраняется позиция *потреблённых* байт (минус недочитанная частичная строка). (2) L2-лимит перестал промахиваться при параллельных сессиях: между проверкой на MAIL FROM и вставкой строки на end-of-message сообщение не видно в БД, поэтому N одновременных сессий пропускали друг друга — теперь к счёту из БД добавляются «в полёте» (`internal/milter/inflight.go`, общий на процесс реестр резерваций); резервация освобождается после записи в send-log, на ABORT и по TTL 10 минут (у go-milter нет колбэка на закрытие соединения, а вечная резервация — это fail-closed-дрейф, которого у лимитера быть не должно). Транзакция «count+insert», как предлагал review, невозможна буквально: эти два шага разнесены по разным стадиям SMTP-транзакции. Тесты: restart/rotation-resume для tailer'а, четыре сценария резерваций для лимита. `gofmt`/`go vet` чистые; `go test ./...` — падения только известные Windows-специфичные (`internal/domain`, `TestFollowTailsAndRotates`). Не проверено на стенде (нет Docker локально) — кандидат на следующий прогон на dev-сервере.
 - **Дальше:** релизный гейт (Фаза 0) закрыт по существу — e2e C.4 и ревизия § D пройдены; остаются только шаги, которые делаются в момент резки версии (бамп тега образа + Codeberg URL, git tag) по явной команде пользователя. Остальные пункты из [code-review.md](code-review.md) — polish (фазы 2–3).
 - **Принятые риски** — [security.md](security.md). **Опционально v1.x / 2.x** — [roadmap.md](roadmap.md) (хвост документации, send-log gaps, Фаза O1+, роль администратора домена).
 - **Прод:** `selfpost.example.com`, реальный Let's Encrypt сертификат, живой e2e (DKIM/SPF pass). Контейнер там всё ещё на образе v1.0 — Фаза 14 в него не выкатывалась. При апгрейде: админа один раз разлогинит (сменилось имя cookie), а от reverse-proxy требуется передача исходного `Host` (Apache-фрагмент из `deploy/` это делает).
diff --git a/docs/roadmap.md b/docs/roadmap.md
index ef3bf91..40193af 100644
--- a/docs/roadmap.md
+++ b/docs/roadmap.md
@@ -60,11 +60,12 @@ CHANGELOG `[Unreleased]/Security`, а разделы B.1–B.3 и C.4 вырез
 (`docs/logo` уже отсутствует как каталог — критерию «либо содержит файлы, либо
 отсутствует» удовлетворяет; отдельного действия не требуется.)
 
-**Send-log vs `mail.log` (опционально).** После рестарта панели или пересоздания
-контейнера строки send-log могут навсегда остаться `queued` — log-tailer не
-дочитывает пропущенный хвост, `mail.log` не в `/data`. Rename-ротация (B.2)
-это не лечит. Кандидаты, если станет больно: persist позиции чтения, volume для
-лога, сверка зависших строк через `postqueue`. As-built и принятый риск:
+**Send-log vs `mail.log` (частично закрыто).** Persist позиции чтения сделан
+(Фаза 3 [code-review.md](code-review.md), таблица `logtail_state`): после
+рестарта панели log-tailer дочитывает пропущенный хвост. Остаётся пересоздание
+контейнера — `mail.log` не в `/data` и теряется вместе с ним, такие строки
+навсегда останутся `queued`. Кандидаты, если станет больно: volume для лога,
+сверка зависших строк через `postqueue`. As-built и принятый риск:
 [architecture.md](architecture.md) § Log tailer, [security.md](security.md).
 
 ---
diff --git a/docs/security.md b/docs/security.md
index 295c4dd..3f5f587 100644
--- a/docs/security.md
+++ b/docs/security.md
@@ -118,10 +118,12 @@ Hardening сверх обязательного (security-заголовки, п
   безвозвратно нерасшифровываемый архив хуже незашифрованного: пароль SelfPost
   не хранит. Триггером сделать шифрование обязательным считать появление
   второго администратора (тогда «кто скачал» перестаёт быть одним человеком).
-- **Send-log может навсегда остаться `queued` после рестарта панели или
-  пересоздания контейнера.** Log-tailer стартует с конца `mail.log`; файл не в
-  `/data` и теряется при recreate. Rename-ротация это не лечит. См. [architecture.md](architecture.md)
-  § Log tailer — known gaps.
+- **Send-log может навсегда остаться `queued` после пересоздания контейнера.**
+  Рестарт панели это больше не вызывает: log-tailer сохраняет позицию чтения
+  (`logtail_state`, миграция `0003`) и после старта дочитывает хвост, написанный
+  пока панель лежала. Но `mail.log` живёт не в `/data` и теряется вместе с
+  контейнером — строки, чьи delivery-строки были только в нём, останутся
+  `queued`. См. [architecture.md](architecture.md) § Log tailer.
 
 ## ADR: CSRF через проверку Origin, без токенов
 
diff --git a/internal/logtail/logtail.go b/internal/logtail/logtail.go
index 70c2984..4829f3e 100644
--- a/internal/logtail/logtail.go
+++ b/internal/logtail/logtail.go
@@ -20,11 +20,14 @@ import (
 	"codeberg.org/mix/selfpost/internal/store"
 )
 
-// StatusStore is the slice of the store the log-tailer needs. *store.Store
-// satisfies it.
+// StatusStore is the slice of the store the log-tailer needs: advancing
+// delivery statuses, pruning the retention window, and remembering how far into
+// mail.log it has read. *store.Store satisfies it.
 type StatusStore interface {
 	UpdateStatus(queueID, recipient, status string) (int64, error)
 	DeleteSendLogBefore(cutoff time.Time) (int64, error)
+	LogtailState(path string) (store.LogtailState, bool, error)
+	SaveLogtailState(path string, st store.LogtailState) error
 }
 
 // pollInterval is how often the tail loop checks for new bytes / rotation. It
@@ -73,12 +76,13 @@ func parseDelivery(line string) (queueID, recipient, status string, ok bool) {
 }
 
 // Run follows path and updates send-log statuses until ctx is cancelled, while
-// a background sweep prunes rows older than retentionDays. It returns nil on a
-// clean shutdown.
+// a background sweep prunes rows older than retentionDays. Reading resumes at
+// the offset the previous run persisted, so a restart parses the delivery lines
+// written while the panel was down. It returns nil on a clean shutdown.
 func Run(ctx context.Context, path string, st StatusStore, retentionDays int) error {
 	go retentionLoop(ctx, st, retentionDays)
 
-	return follow(ctx, path, func(line string) {
+	return follow(ctx, path, &tracker{st: st, path: path}, func(line string) {
 		queueID, recipient, status, ok := parseDelivery(line)
 		if !ok {
 			return
@@ -166,10 +170,11 @@ func TailLines(path string, n int) ([]string, error) {
 }
 
 // follow tails path line by line, calling handle for each complete line, until
-// ctx is cancelled. It starts at end-of-file (so a restart does not reprocess
-// history) and reopens the file when it is rotated (inode change from
+// ctx is cancelled. Where it starts is tr's decision (a persisted offset, the
+// start of a file that changed while the panel was down, or end-of-file on a
+// first ever run); it reopens the file when it is rotated (inode change from
 // logrotate's create, or truncation from copytruncate) so nothing is missed.
-func follow(ctx context.Context, path string, handle func(string)) error {
+func follow(ctx context.Context, path string, tr *tracker, handle func(string)) error {
 	var (
 		f       *os.File
 		r       *bufio.Reader
@@ -199,7 +204,7 @@ func follow(ctx context.Context, path string, handle func(string)) error {
 
 	// The container may start before Postfix has created mail.log; wait for it.
 	for {
-		if err := openAt(0, io.SeekEnd); err == nil {
+		if err := openAt(0, io.SeekStart); err == nil {
 			break
 		}
 		select {
@@ -208,6 +213,10 @@ func follow(ctx context.Context, path string, handle func(string)) error {
 		case <-time.After(pollInterval):
 		}
 	}
+	if _, err := f.Seek(tr.resume(f), io.SeekStart); err != nil {
+		log.Printf("log-tailer: seek %s: %v", path, err)
+	}
+	r.Reset(f) // the reader buffered from the pre-seek position
 	defer func() {
 		if f != nil {
 			f.Close()
@@ -231,11 +240,20 @@ func follow(ctx context.Context, path string, handle func(string)) error {
 		}
 	}
 
+	// read returns how many bytes of the open file have actually been consumed:
+	// the descriptor position less the partial line bufio handed back at EOF,
+	// which is re-read (and completed) on the next drain or the next start.
+	read := func() int64 {
+		pos, _ := f.Seek(0, io.SeekCurrent)
+		return pos - int64(len(pending))
+	}
+
 	ticker := time.NewTicker(pollInterval)
 	defer ticker.Stop()
 	for {
 		select {
 		case <-ctx.Done():
+			tr.record(f, read(), true) // shutdown: the next start resumes here
 			return nil
 		case <-ticker.C:
 			drain()
@@ -252,8 +270,12 @@ func follow(ctx context.Context, path string, handle func(string)) error {
 				drain()
 				if err := openAt(0, io.SeekStart); err != nil {
 					log.Printf("log-tailer: reopen %s: %v", path, err)
+					continue
 				}
+				tr.adopt(f)
+				continue
 			}
+			tr.record(f, read(), false)
 		}
 	}
 }
diff --git a/internal/logtail/logtail_test.go b/internal/logtail/logtail_test.go
index 6abaf1f..62c3d47 100644
--- a/internal/logtail/logtail_test.go
+++ b/internal/logtail/logtail_test.go
@@ -4,6 +4,7 @@ import (
 	"context"
 	"os"
 	"path/filepath"
+	"strings"
 	"sync"
 	"testing"
 	"time"
@@ -77,10 +78,16 @@ func TestParseDelivery(t *testing.T) {
 	}
 }
 
-// captureStore records UpdateStatus calls for the follow integration test.
+// captureStore records UpdateStatus calls for the follow integration test and
+// keeps the persisted read offset in memory, so a "restart" in a test is a
+// second Run against the same captureStore.
 type captureStore struct {
 	mu    sync.Mutex
 	calls []string
+
+	state     store.LogtailState
+	haveState bool
+	stateErr  error
 }
 
 func (c *captureStore) UpdateStatus(queueID, recipient, status string) (int64, error) {
@@ -92,12 +99,37 @@ func (c *captureStore) UpdateStatus(queueID, recipient, status string) (int64, e
 
 func (c *captureStore) DeleteSendLogBefore(time.Time) (int64, error) { return 0, nil }
 
+func (c *captureStore) LogtailState(string) (store.LogtailState, bool, error) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if c.stateErr != nil {
+		return store.LogtailState{}, false, c.stateErr
+	}
+	return c.state, c.haveState, nil
+}
+
+func (c *captureStore) SaveLogtailState(_ string, st store.LogtailState) error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if c.stateErr != nil {
+		return c.stateErr
+	}
+	c.state, c.haveState = st, true
+	return nil
+}
+
 func (c *captureStore) snapshot() []string {
 	c.mu.Lock()
 	defer c.mu.Unlock()
 	return append([]string(nil), c.calls...)
 }
 
+func (c *captureStore) reset() {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	c.calls = nil
+}
+
 // TestFollowTailsAndRotates writes delivery lines to a log file, then rotates
 // it (rename + fresh create, as logrotate does) and writes more, asserting the
 // tailer picks up lines from both the original and rotated file.
@@ -142,6 +174,80 @@ func TestFollowTailsAndRotates(t *testing.T) {
 	}
 }
 
+// TestFollowResumesAfterRestart covers the persisted read offset: a restart
+// must parse the delivery lines written while the tailer was down (rows that
+// would otherwise stay "queued" forever), without re-parsing what it already
+// read, and must fall back to reading the whole file when the log was rotated
+// or recreated in the meantime.
+func TestFollowResumesAfterRestart(t *testing.T) {
+	old := pollInterval
+	pollInterval = 10 * time.Millisecond
+	t.Cleanup(func() { pollInterval = old })
+
+	dir := t.TempDir()
+	path := filepath.Join(dir, "mail.log")
+	// A head longer than fingerprintSize, so the file stays identifiable across
+	// the restart; the lines themselves predate the first start and are ignored.
+	seed := strings.Repeat("host postfix/qmgr[1]: seed line, not a delivery\n", 20)
+	if err := os.WriteFile(path, []byte(seed), 0o644); err != nil {
+		t.Fatalf("seed log: %v", err)
+	}
+
+	cs := &captureStore{}
+	stop := startRun(t, path, cs)
+	appendLine(t, path, "host postfix/smtp[1]: Q1: to=<a@example.net>, dsn=2.0.0, status=sent (ok)")
+	waitFor(t, func() bool { return contains(cs.snapshot(), "Q1|a@example.net|sent") })
+	stop() // persists the offset past Q1
+
+	// Down: Postfix keeps delivering.
+	appendLine(t, path, "host postfix/smtp[1]: Q2: to=<b@example.net>, dsn=2.0.0, status=sent (ok)")
+
+	cs.reset()
+	stop = startRun(t, path, cs)
+	waitFor(t, func() bool { return contains(cs.snapshot(), "Q2|b@example.net|sent") })
+	if contains(cs.snapshot(), "Q1|a@example.net|sent") {
+		t.Fatal("resumed run re-parsed Q1: offset was not honoured")
+	}
+	stop()
+
+	// Down again, and this time the log is replaced (logrotate + fresh create).
+	// The stored offset belongs to a file that no longer exists, so the new one
+	// must be read from the start.
+	if err := os.WriteFile(path, []byte(strings.Repeat("host postfix/qmgr[1]: fresh log after rotation\n", 20)+
+		"host postfix/smtp[1]: Q3: to=<c@example.net>, dsn=5.1.1, status=bounced (nope)\n"), 0o644); err != nil {
+		t.Fatalf("recreate log: %v", err)
+	}
+
+	cs.reset()
+	stop = startRun(t, path, cs)
+	waitFor(t, func() bool { return contains(cs.snapshot(), "Q3|c@example.net|bounced") })
+	stop()
+}
+
+// startRun launches the tailer and returns a function that cancels it and waits
+// for a clean return, the way a panel restart bookends a run.
+func startRun(t *testing.T, path string, cs *captureStore) func() {
+	t.Helper()
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan error, 1)
+	go func() { done <- Run(ctx, path, cs, 90) }()
+	// follow() opens and seeks on start; give it a moment before the caller
+	// appends, so the append is not raced by the initial open.
+	time.Sleep(50 * time.Millisecond)
+	return func() {
+		t.Helper()
+		cancel()
+		select {
+		case err := <-done:
+			if err != nil {
+				t.Fatalf("Run: %v", err)
+			}
+		case <-time.After(2 * time.Second):
+			t.Fatal("Run did not return after cancel")
+		}
+	}
+}
+
 func appendLine(t *testing.T, path, line string) {
 	t.Helper()
 	f, err := os.OpenFile(path, os.O_APPEND|os.O_WRONLY, 0o644)
diff --git a/internal/logtail/offset.go b/internal/logtail/offset.go
new file mode 100644
index 0000000..f8c8000
--- /dev/null
+++ b/internal/logtail/offset.go
@@ -0,0 +1,136 @@
+package logtail
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"errors"
+	"io"
+	"log"
+	"os"
+	"time"
+
+	"codeberg.org/mix/selfpost/internal/store"
+)
+
+const (
+	// fingerprintSize is how many bytes from the head of the log identify it.
+	// Postfix writes a timestamped line per event, so the first 512 bytes are
+	// effectively unique per log generation — enough to tell "the file we were
+	// reading" from "a fresh one created by logrotate while we were down",
+	// which os.SameFile cannot answer across a restart.
+	fingerprintSize = 512
+	// persistInterval throttles the offset write. Losing up to this much
+	// progress on a crash only means re-parsing a few lines (UpdateStatus is
+	// idempotent), which is much cheaper than a database write per poll tick.
+	persistInterval = 5 * time.Second
+)
+
+// tracker persists the tailer's read position so a restart resumes where the
+// previous run stopped instead of jumping to end-of-file — the "send-log rows
+// stay queued forever" gap (architecture.md § Log tailer). It is used from the
+// follow loop only, so it needs no locking.
+type tracker struct {
+	st   StatusStore
+	path string
+
+	fp       string // fingerprint of the file currently open ("" if too short)
+	saved    int64  // last offset written to the store
+	lastSave time.Time
+}
+
+// resume returns the byte offset the tailer should start reading f at, having
+// recorded f's fingerprint for later saves.
+//
+// The rules, in order: no stored state at all (first ever start) means start at
+// the end, so installing the panel does not replay a pre-existing log; a stored
+// state whose fingerprint still matches means continue from it, parsing the
+// tail written while the panel was down; anything else means the file is not
+// the one the offset referred to (rotated, recreated or truncated in the
+// meantime), so read it from the start. Re-parsing lines already seen is
+// harmless: UpdateStatus writes the same status onto the same row.
+func (t *tracker) resume(f *os.File) int64 {
+	size, err := fileSize(f)
+	if err != nil {
+		log.Printf("log-tailer: stat %s: %v (reading from the start)", t.path, err)
+		return 0
+	}
+	t.fp = fingerprintOf(f)
+
+	prev, ok, err := t.st.LogtailState(t.path)
+	if err != nil {
+		log.Printf("log-tailer: read stored offset: %v (starting at end)", err)
+		return size
+	}
+	switch {
+	case !ok:
+		t.saved = size
+		return size
+	case prev.Fingerprint != "" && prev.Fingerprint == t.fp && prev.Offset <= size:
+		t.saved = prev.Offset
+		if prev.Offset < size {
+			log.Printf("log-tailer: resuming %s at offset %d (%d bytes to catch up)",
+				t.path, prev.Offset, size-prev.Offset)
+		}
+		return prev.Offset
+	default:
+		log.Printf("log-tailer: %s changed while the panel was down; reading from the start", t.path)
+		t.saved = 0
+		return 0
+	}
+}
+
+// adopt re-fingerprints after the follow loop switched to a rotated-in file and
+// persists the fresh start immediately, so a restart right after a rotation
+// does not resume at the old file's offset.
+func (t *tracker) adopt(f *os.File) {
+	t.fp = fingerprintOf(f)
+	t.saved = -1 // force the write below even if the old offset happened to be 0
+	t.record(f, 0, true)
+}
+
+// record persists offset, at most once per persistInterval unless force is set
+// (rotation and shutdown, where the write must not be skipped).
+func (t *tracker) record(f *os.File, offset int64, force bool) {
+	if offset == t.saved {
+		return
+	}
+	if !force && time.Since(t.lastSave) < persistInterval {
+		return
+	}
+	if t.fp == "" {
+		// The log was shorter than a fingerprint when we opened it; now that it
+		// has grown, an identifiable one may be available.
+		t.fp = fingerprintOf(f)
+	}
+	if err := t.st.SaveLogtailState(t.path, store.LogtailState{Fingerprint: t.fp, Offset: offset}); err != nil {
+		log.Printf("log-tailer: save offset: %v", err)
+		return
+	}
+	t.saved = offset
+	t.lastSave = time.Now()
+}
+
+// fingerprintOf hashes the head of the file. It returns "" for a file too short
+// to identify — the head would still change as Postfix appends, so such a
+// fingerprint could not be compared meaningfully on the next start.
+func fingerprintOf(f *os.File) string {
+	buf := make([]byte, fingerprintSize)
+	n, err := f.ReadAt(buf, 0)
+	if err != nil && !errors.Is(err, io.EOF) {
+		log.Printf("log-tailer: fingerprint read: %v", err)
+		return ""
+	}
+	if n < fingerprintSize {
+		return ""
+	}
+	sum := sha256.Sum256(buf)
+	return hex.EncodeToString(sum[:])
+}
+
+func fileSize(f *os.File) (int64, error) {
+	info, err := f.Stat()
+	if err != nil {
+		return 0, err
+	}
+	return info.Size(), nil
+}
diff --git a/internal/milter/inflight.go b/internal/milter/inflight.go
new file mode 100644
index 0000000..78ca52b
--- /dev/null
+++ b/internal/milter/inflight.go
@@ -0,0 +1,99 @@
+package milter
+
+import (
+	"sync"
+	"time"
+)
+
+// reservationTTL bounds how long a message may stay reserved. A reservation is
+// released at end-of-message or on ABORT, but a client that simply drops the
+// connection after MAIL FROM produces neither callback (go-milter has no
+// connection-close hook), and a reservation that never expired would count
+// against the limit forever — a fail-closed drift this milter must not have.
+// The TTL is generously longer than any realistic DATA transfer, so a message
+// still being received is never dropped from the count.
+const reservationTTL = 10 * time.Minute
+
+// reservation is one message that passed the level-2 check and has not been
+// written to the send log yet.
+type reservation struct {
+	key string
+	at  time.Time
+}
+
+// inflight counts messages that are between the limit check (MAIL FROM) and the
+// send-log insert (end-of-message). The stored count alone cannot see them, so
+// without this several concurrent SMTP sessions each read the same pre-insert
+// count, each conclude they are under the ceiling, and the limit is overshot by
+// however many were in flight. Counting reservations closes that window without
+// writing placeholder rows the operator would see in the UI.
+//
+// One instance is shared by every session of the process, hence the mutex.
+// Methods tolerate a nil receiver so a session built without one (tests) simply
+// behaves as it did before.
+type inflight struct {
+	mu sync.Mutex
+	m  map[string]map[*reservation]struct{}
+}
+
+// count returns how many reservations for key were taken within the limit's
+// window (at or after since), pruning any that outlived reservationTTL.
+func (f *inflight) count(key string, since time.Time) int64 {
+	if f == nil {
+		return 0
+	}
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	set := f.m[key]
+	cutoff := time.Now().Add(-reservationTTL)
+	var n int64
+	for r := range set {
+		if r.at.Before(cutoff) {
+			delete(set, r)
+			continue
+		}
+		if !r.at.Before(since) {
+			n++
+		}
+	}
+	if len(set) == 0 {
+		delete(f.m, key)
+	}
+	return n
+}
+
+// reserve claims a slot for key until the message is recorded or released.
+func (f *inflight) reserve(key string) *reservation {
+	if f == nil {
+		return nil
+	}
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	if f.m == nil {
+		f.m = make(map[string]map[*reservation]struct{})
+	}
+	if f.m[key] == nil {
+		f.m[key] = make(map[*reservation]struct{})
+	}
+	r := &reservation{key: key, at: time.Now()}
+	f.m[key][r] = struct{}{}
+	return r
+}
+
+// release drops a reservation, either because the message reached the send log
+// (where the stored count takes over) or because it never will.
+func (f *inflight) release(r *reservation) {
+	if f == nil || r == nil {
+		return
+	}
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	set := f.m[r.key]
+	delete(set, r)
+	if len(set) == 0 {
+		delete(f.m, r.key)
+	}
+}
diff --git a/internal/milter/milter.go b/internal/milter/milter.go
index 09dbcb9..2ae29d4 100644
--- a/internal/milter/milter.go
+++ b/internal/milter/milter.go
@@ -45,13 +45,18 @@ type Store interface {
 type session struct {
 	milter.NoOpMilter
 	rec Store
+	// flight is shared by every session of the process; it holds the messages
+	// that passed the level-2 check but are not in the send log yet. Nil is a
+	// valid zero value (no in-flight accounting).
+	flight *inflight
 
 	clientIP string // captured once per connection
 
-	login   string
-	from    string
-	rcpts   []string
-	subject string
+	login    string
+	from     string
+	rcpts    []string
+	subject  string
+	reserved []*reservation // level-2 slots held by the current message
 }
 
 // Connect captures the client IP, which comes from the addr parameter rather
@@ -72,6 +77,7 @@ func (s *session) Connect(host, family string, port uint16, addr net.IP, m *milt
 // tempfail before recipients are even offered (spec 7.4). Enforcement is
 // fail-open — see overLimit.
 func (s *session) MailFrom(from string, m *milter.Modifier) (milter.Response, error) {
+	s.releaseReservations() // a previous transaction that ended without EOM/ABORT
 	s.from = cleanAddress(from)
 	s.login = macro(m, "auth_authen")
 	s.rcpts = nil
@@ -127,9 +133,22 @@ func decodeSubject(v string) string {
 // rows are written. We accept (this milter is done) without ever rejecting.
 func (s *session) Body(m *milter.Modifier) (milter.Response, error) {
 	s.record(macro(m, "i"))
+	// The rows are in the send log now, so the stored count sees this message
+	// and its level-2 slots are no longer needed.
+	s.releaseReservations()
 	return milter.RespAccept, nil
 }
 
+// Abort ends the current transaction without an end-of-message (client RSET, or
+// Postfix rejecting the message for its own reasons). No send-log row will be
+// written, so the level-2 slots this message held must go back.
+func (s *session) Abort(m *milter.Modifier) error {
+	s.releaseReservations()
+	s.rcpts = nil
+	s.subject = ""
+	return nil
+}
+
 // macro reads a milter macro, tolerating Postfix's convention of wrapping
 // multi-character macro names in curly braces (e.g. {auth_authen}) while
 // single-character names (e.g. i) arrive bare. go-milter stores whatever name
@@ -191,8 +210,9 @@ func domainOf(addr string) string {
 // Serve runs the journal-milter on ln until ctx is cancelled. Each connection
 // gets a fresh session bound to rec. It returns nil on a clean shutdown.
 func Serve(ctx context.Context, ln net.Listener, rec Store) error {
+	flight := &inflight{} // shared: the level-2 window spans all connections
 	srv := &milter.Server{
-		NewMilter: func() milter.Milter { return &session{rec: rec} },
+		NewMilter: func() milter.Milter { return &session{rec: rec, flight: flight} },
 		Actions:   0,                // read-only: we make no message modifications
 		Protocol:  milter.OptNoBody, // the journal needs headers/EOM, not the body
 	}
diff --git a/internal/milter/milter_test.go b/internal/milter/milter_test.go
index 999a43a..9474e6b 100644
--- a/internal/milter/milter_test.go
+++ b/internal/milter/milter_test.go
@@ -331,6 +331,130 @@ func TestRateLimitNoIPKeyDoesNotApply(t *testing.T) {
 	}
 }
 
+// mailFromIn is mailFrom with an explicit shared in-flight registry, so a test
+// can play several concurrent SMTP sessions of one process against each other.
+func mailFromIn(t *testing.T, rec Store, fl *inflight, ip, from, login string) (*session, milter.Response) {
+	t.Helper()
+	s := &session{rec: rec, flight: fl}
+	if _, err := s.Connect("h", "tcp4", 0, net.ParseIP(ip), mods(nil)); err != nil {
+		t.Fatalf("Connect: %v", err)
+	}
+	resp, err := s.MailFrom(from, mods(map[string]string{"auth_authen": login}))
+	if err != nil {
+		t.Fatalf("MailFrom: %v", err)
+	}
+	return s, resp
+}
+
+func limitedRecorder(count int64) *fakeRecorder {
+	return &fakeRecorder{
+		limits: map[string]store.RateLimit{
+			store.RateLimitScopeDomain + "|example.com": activeLimit(limitIP),
+		},
+		counts: map[string]int64{store.RateLimitScopeDomain + "|example.com": count},
+	}
+}
+
+// Messages between MAIL FROM and end-of-message are not in the send log yet, so
+// counting the stored rows alone lets concurrent sessions each pass the same
+// check and overshoot the ceiling. The last free slot may only be taken once.
+func TestRateLimitCountsInFlightMessages(t *testing.T) {
+	rec := limitedRecorder(4) // one below the ceiling of 5
+	fl := &inflight{}
+
+	if _, resp := mailFromIn(t, rec, fl, limitIP, "a@example.com", "app1"); resp != milter.RespContinue {
+		t.Fatalf("first message = %v, want Continue (4/5 stored)", resp)
+	}
+	// Same window, nothing written yet: the first message holds the fifth slot.
+	if _, resp := mailFromIn(t, rec, fl, limitIP, "b@example.com", "app1"); resp != milter.RespTempFail {
+		t.Fatalf("concurrent message = %v, want TempFail (would overshoot)", resp)
+	}
+	if len(rec.rejected) != 1 {
+		t.Fatalf("want one rejected send-log row, got %+v", rec.rejected)
+	}
+}
+
+// Once the message is recorded the stored count sees it, so its reservation
+// must be given back — otherwise it would be counted twice and the ceiling
+// would drift closed.
+func TestReservationReleasedAtEndOfMessage(t *testing.T) {
+	rec := limitedRecorder(4)
+	fl := &inflight{}
+
+	s, resp := mailFromIn(t, rec, fl, limitIP, "a@example.com", "app1")
+	if resp != milter.RespContinue {
+		t.Fatalf("first message = %v, want Continue", resp)
+	}
+	if _, err := s.Body(mods(map[string]string{"i": "Q1"})); err != nil {
+		t.Fatalf("Body: %v", err)
+	}
+	if n := fl.count(store.RateLimitScopeDomain+"|example.com", time.Now().Add(-time.Hour)); n != 0 {
+		t.Fatalf("in-flight count after EOM = %d, want 0", n)
+	}
+}
+
+// A transaction the client abandons (RSET, or a Postfix-side rejection) never
+// reaches the send log, so its slot must not stay claimed.
+func TestReservationReleasedOnAbort(t *testing.T) {
+	rec := limitedRecorder(4)
+	fl := &inflight{}
+
+	s, resp := mailFromIn(t, rec, fl, limitIP, "a@example.com", "app1")
+	if resp != milter.RespContinue {
+		t.Fatalf("first message = %v, want Continue", resp)
+	}
+	if err := s.Abort(mods(nil)); err != nil {
+		t.Fatalf("Abort: %v", err)
+	}
+	if _, resp := mailFromIn(t, rec, fl, limitIP, "b@example.com", "app1"); resp != milter.RespContinue {
+		t.Fatalf("after abort = %v, want Continue (slot released)", resp)
+	}
+}
+
+// A refused message must not leave the slots it claimed for the limits checked
+// before the one that tripped, or every refusal would tighten the ceiling.
+func TestRefusalReleasesEarlierReservation(t *testing.T) {
+	rec := &fakeRecorder{
+		limits: map[string]store.RateLimit{
+			store.RateLimitScopeDomain + "|example.com": activeLimit(limitIP),
+			store.RateLimitScopeApp + "|app1":           activeLimit(limitIP),
+		},
+		counts: map[string]int64{
+			store.RateLimitScopeDomain + "|example.com": 0, // domain: plenty of room
+			store.RateLimitScopeApp + "|app1":           5, // app: at the ceiling
+		},
+	}
+	fl := &inflight{}
+	if _, resp := mailFromIn(t, rec, fl, limitIP, "a@example.com", "app1"); resp != milter.RespTempFail {
+		t.Fatalf("app over limit = %v, want TempFail", resp)
+	}
+	if n := fl.count(store.RateLimitScopeDomain+"|example.com", time.Now().Add(-time.Hour)); n != 0 {
+		t.Fatalf("domain reservation left behind after refusal: %d", n)
+	}
+}
+
+// The in-flight count only covers the limit's own window: a reservation older
+// than it (a session stuck mid-DATA for longer than the window) must not be
+// counted against a window it no longer belongs to.
+func TestInflightIgnoresReservationsOutsideWindow(t *testing.T) {
+	fl := &inflight{}
+	r := fl.reserve("domain|example.com")
+	r.at = time.Now().Add(-time.Minute)
+
+	if n := fl.count("domain|example.com", time.Now().Add(-time.Hour)); n != 1 {
+		t.Fatalf("count inside window = %d, want 1", n)
+	}
+	if n := fl.count("domain|example.com", time.Now().Add(-time.Second)); n != 0 {
+		t.Fatalf("count outside window = %d, want 0", n)
+	}
+	// Past the TTL the reservation is dropped even for a wide window, so a
+	// client that vanished after MAIL FROM cannot hold a slot forever.
+	r.at = time.Now().Add(-2 * reservationTTL)
+	if n := fl.count("domain|example.com", time.Now().Add(-3*reservationTTL)); n != 0 {
+		t.Fatalf("expired reservation still counted: %d", n)
+	}
+}
+
 func TestDomainOf(t *testing.T) {
 	cases := map[string]string{
 		"user@Example.COM": "example.com",
diff --git a/internal/milter/ratelimit.go b/internal/milter/ratelimit.go
index a2ad2fd..049c68a 100644
--- a/internal/milter/ratelimit.go
+++ b/internal/milter/ratelimit.go
@@ -17,6 +17,10 @@ import (
 // can never block mail — Postfix's level-1 anvil limit (spec 5) remains the
 // backstop, and it does not depend on this milter at all. Only a clean count at
 // or above a configured ceiling returns true.
+//
+// A message that passes reserves a slot per applicable limit, released once it
+// reaches the send log (or is abandoned) — see inflight for why the stored
+// count alone is not enough.
 func (s *session) overLimit() bool {
 	if s.clientIP == "" {
 		return false // no client IP to key on; level-2 does not apply
@@ -25,6 +29,7 @@ func (s *session) overLimit() bool {
 		{store.RateLimitScopeDomain, domainOf(s.from)},
 		{store.RateLimitScopeApp, s.login},
 	}
+	var taken []*reservation
 	for _, c := range checks {
 		if c.ref == "" {
 			continue
@@ -45,15 +50,34 @@ func (s *session) overLimit() bool {
 			log.Printf("journal-milter: rate-limit count %s %q: %v (fail-open)", c.scope, c.ref, err)
 			continue
 		}
+		key := c.scope + "|" + c.ref
+		n += s.flight.count(key, since)
 		if n >= int64(rl.MaxMessages) {
 			log.Printf("journal-milter: %s %q over limit: %d/%d in %ds from %s — refusing 4xx",
 				c.scope, c.ref, n, rl.MaxMessages, rl.WindowSeconds, s.clientIP)
+			// The message is refused, so the slots claimed for the limits
+			// checked before this one must not stay claimed.
+			for _, r := range taken {
+				s.flight.release(r)
+			}
 			return true
 		}
+		taken = append(taken, s.flight.reserve(key))
 	}
+	s.reserved = append(s.reserved, taken...)
 	return false
 }
 
+// releaseReservations gives back every slot this message holds. It runs once
+// the message is in the send log (where the stored count sees it), and whenever
+// the transaction ends without getting there.
+func (s *session) releaseReservations() {
+	for _, r := range s.reserved {
+		s.flight.release(r)
+	}
+	s.reserved = nil
+}
+
 // recordRejected writes a send-log row for a message refused by a level-2 limit
 // (spec 7.4, "опционально фиксирует ... для видимости в UI"), so the rejection
 // shows up in the monitoring screen. Only MAIL-stage fields are known; the write
diff --git a/internal/store/logtail.go b/internal/store/logtail.go
new file mode 100644
index 0000000..e9d1fc8
--- /dev/null
+++ b/internal/store/logtail.go
@@ -0,0 +1,54 @@
+package store
+
+import (
+	"database/sql"
+	"errors"
+	"fmt"
+	"time"
+)
+
+// LogtailState is the persisted read position of the mail.log tailer: how far
+// into the log it had got, and a fingerprint of the file that offset refers to.
+// It lets the tailer resume after a panel restart instead of jumping to
+// end-of-file and losing the delivery lines written while it was down.
+type LogtailState struct {
+	Fingerprint string
+	Offset      int64
+}
+
+// LogtailState returns the stored read position for path. ok is false the first
+// time a path is followed (nothing persisted yet), which the tailer treats as
+// "start at the end".
+func (s *Store) LogtailState(path string) (LogtailState, bool, error) {
+	var st LogtailState
+	err := s.db.QueryRow(
+		`SELECT fingerprint, read_offset FROM logtail_state WHERE path = ?`,
+		path,
+	).Scan(&st.Fingerprint, &st.Offset)
+	if errors.Is(err, sql.ErrNoRows) {
+		return LogtailState{}, false, nil
+	}
+	if err != nil {
+		return LogtailState{}, false, fmt.Errorf("read logtail state for %q: %w", path, err)
+	}
+	return st, true, nil
+}
+
+// SaveLogtailState records the tailer's read position for path, replacing any
+// previous one. It is called on a timer while tailing, so it is a single small
+// upsert rather than a transaction.
+func (s *Store) SaveLogtailState(path string, st LogtailState) error {
+	_, err := s.db.Exec(
+		`INSERT INTO logtail_state (path, fingerprint, read_offset, updated_at)
+		 VALUES (?, ?, ?, ?)
+		 ON CONFLICT(path) DO UPDATE SET
+		     fingerprint = excluded.fingerprint,
+		     read_offset = excluded.read_offset,
+		     updated_at  = excluded.updated_at`,
+		path, st.Fingerprint, st.Offset, time.Now().UTC().Format(time.RFC3339),
+	)
+	if err != nil {
+		return fmt.Errorf("save logtail state for %q: %w", path, err)
+	}
+	return nil
+}
diff --git a/internal/store/migrations/0003_logtail_state.sql b/internal/store/migrations/0003_logtail_state.sql
new file mode 100644
index 0000000..c1ea93a
--- /dev/null
+++ b/internal/store/migrations/0003_logtail_state.sql
@@ -0,0 +1,12 @@
+-- Log-tailer read position. Without it the tailer starts at end-of-file on
+-- every start, so delivery lines written while the panel was down are never
+-- parsed and their send-log rows stay "queued" forever. One row per followed
+-- path; fingerprint identifies the file the offset belongs to (the head bytes
+-- of the log), so a rotated or recreated mail.log is detected across a restart,
+-- where os.SameFile cannot help.
+CREATE TABLE logtail_state (
+    path        TEXT PRIMARY KEY,
+    fingerprint TEXT NOT NULL,
+    read_offset INTEGER NOT NULL,
+    updated_at  TEXT NOT NULL
+);