test(media): lock XSS escaping on new-media-type user-string sinks (#21 review F1)

The new Telegram media types (Kurigram 2.2.23) escape every user string in
_format_special_media, but only CONTACT first_name + CHECKLIST task text were
regression-tested — the reviewer mutation-proved the other sinks were uncovered
(removing html.escape from VENUE title / GIVEAWAY description / CHECKLIST title kept
the suite green). Given this project's XSS history (#12 fail-open, #13 debug-title),
lock them.

Add 7 regression tests (mirroring test_contact_name_is_escaped): a <script> payload in
VENUE title, VENUE address label, DICE emoji, GAME title, CHECKLIST title, GIVEAWAY
description, and GIVEAWAY_WINNERS prize_description; each asserts the raw payload is
absent and the escaped form present. Mutation-verified: removing the escape on each sink
reds its test. 321 passed (314 + 7). Test-only — no production code touched.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

config.py

@@ -9,6 +9,7 @@
 # pylance: disable=reportMissingImports, reportMissingModuleSource
 
 import os
+import sys
 import logging
 from typing import Any
 
@@ -42,22 +43,105 @@ def get_settings() -> dict[str, Any]:
     tg_api_id = os.getenv("TG_API_ID")
     tg_api_hash = os.getenv("TG_API_HASH")
     if not tg_api_id or not tg_api_hash:
-        print("TG_API_ID and TG_API_HASH must be set")
+        print("TG_API_ID and TG_API_HASH must be set", flush=True)
-        os._exit(1)
+        sys.exit(1)
+
+    # Validate and convert TG_API_ID to integer
+    try:
+        tg_api_id_int = int(tg_api_id)
+    except ValueError:
+        print(f"TG_API_ID must be a valid integer, got: {tg_api_id!r}", flush=True)
+        sys.exit(1)
 
     log_level = os.getenv("LOG_LEVEL", "INFO")
 
+    # Build MTProto proxy config if proxy host is provided.
+    # Telegram MTProto proxy (telegrammessenger/proxy) exposes a SOCKS5 interface,
+    # so Pyrogram connects to it via SOCKS5 scheme.
+    proxy_host = os.getenv("TG_PROXY_HOST")
+    proxy: dict[str, Any] | None = None
+    if proxy_host:
+        # Validate and convert TG_PROXY_PORT to integer
+        try:
+            proxy_port = int(os.getenv("TG_PROXY_PORT") or 1080)
+        except ValueError:
+            print(f"TG_PROXY_PORT must be a valid integer, got: {os.getenv('TG_PROXY_PORT')!r}", flush=True)
+            sys.exit(1)
+        proxy = {
+            "scheme": "SOCKS5",
+            "hostname": proxy_host,
+            "port": proxy_port,
+            "username": os.getenv("TG_PROXY_USERNAME") or None,
+            "password": os.getenv("TG_PROXY_PASSWORD") or None,
+        }
+
+    # Validate and convert API_PORT to integer
+    try:
+        api_port = int(os.getenv("API_PORT") or 8000)
+    except ValueError:
+        print(f"API_PORT must be a valid integer, got: {os.getenv('API_PORT')!r}", flush=True)
+        sys.exit(1)
+
+    # Local helper to parse int env vars with a default and exit on a bad value
+    def _parse_int_env(name: str, default: int, minimum: int = 1) -> int:
+        raw = os.getenv(name)
+        if raw is None or raw.strip() == "":
+            return default
+        try:
+            value = int(raw)
+        except ValueError:
+            print(f"{name} must be a valid integer, got: {raw!r}", flush=True)
+            sys.exit(1)
+        if value < minimum:
+            print(f"{name} must be >= {minimum}, got: {value}", flush=True)
+            sys.exit(1)
+        return value
+
     return {
-        "tg_api_id": int(tg_api_id),
+        "tg_api_id": tg_api_id_int,
         "tg_api_hash": tg_api_hash,
         "session_path": os.getenv("SESSION_PATH", "data") or "data",
         "api_host": os.getenv("API_HOST", "0.0.0.0"),
-        "api_port": int(os.getenv("API_PORT") or 8000),
+        "api_port": api_port,
         "pyrogram_bridge_url": os.getenv("PYROGRAM_BRIDGE_URL", ""),
         "log_level": log_level,
-        "debug": os.getenv("DEBUG", "False") == "True",
+        "debug": os.getenv("DEBUG", "False").strip() in ["True", "true"],
         "token": os.getenv("TOKEN", ""),
+        "trusted_proxies": [ip.strip() for ip in os.getenv("TRUSTED_PROXIES", "").split(",") if ip.strip()],
         "time_based_merge": os.getenv("TIME_BASED_MERGE", "False").strip() in ["True", "true"],
         "show_bridge_link": os.getenv("SHOW_BRIDGE_LINK", "False").strip() in ["True", "true"],
         "show_post_flags": os.getenv("SHOW_POST_FLAGS", "False").strip() in ["True", "true"],
-    } 
+        "proxy": proxy,
+        # Hard cap (seconds) on any single live Telegram RPC held under the global RPC gate,
+        # so a hung MTProto call can never pin the gate (and the whole app) indefinitely.
+        "tg_rpc_timeout": _parse_int_env("TG_RPC_TIMEOUT", 60),
+        "tg_watchdog_enabled": os.getenv("TG_WATCHDOG_ENABLED", "true").strip().lower() not in ["false", "0", "no", "off", "disable", "disabled"],
+        "tg_watchdog_interval": _parse_int_env("TG_WATCHDOG_INTERVAL", 60),
+        "tg_watchdog_timeout": _parse_int_env("TG_WATCHDOG_TIMEOUT", 10),
+        "tg_watchdog_failures": _parse_int_env("TG_WATCHDOG_FAILURES", 3),
+        "tg_watchdog_restart_timeout": _parse_int_env("TG_WATCHDOG_RESTART_TIMEOUT", 90),
+        "tg_watchdog_heartbeat_every": _parse_int_env("TG_WATCHDOG_HEARTBEAT_EVERY", 30),
+        "tg_disconnect_flap_limit": _parse_int_env("TG_DISCONNECT_FLAP_LIMIT", 3),
+        "tg_disconnect_flap_window": _parse_int_env("TG_DISCONNECT_FLAP_WINDOW", 120),
+        # /ping reports TG as unhealthy once the last successful watchdog probe is older than
+        # this many seconds. Default is derived from the watchdog cadence: it is roughly how
+        # long the watchdog itself would take to give up and trigger a restart —
+        # interval * (failures + 1) + timeout. With the defaults (60,3,10) that is 250s, so a
+        # transient slow probe never flaps /ping, but a genuinely stuck session (no successful
+        # probe for ~4 min) surfaces as 503 before/around the time the watchdog restarts.
+        "tg_ping_unhealthy_after": _parse_int_env(
+            "TG_PING_UNHEALTHY_AFTER",
+            _parse_int_env("TG_WATCHDOG_INTERVAL", 60) * (_parse_int_env("TG_WATCHDOG_FAILURES", 3) + 1)
+            + _parse_int_env("TG_WATCHDOG_TIMEOUT", 10),
+        ),
+        # Media download timeout scales with file size (large videos): the per-download
+        # timeout is clamped to [min, max] seconds, with an effective floor of
+        # `media_download_min_speed` bytes/s (timeout ≈ file_size / min_speed).
+        "media_download_timeout_min": _parse_int_env("MEDIA_DOWNLOAD_TIMEOUT_MIN", 120),
+        "media_download_timeout_max": _parse_int_env("MEDIA_DOWNLOAD_TIMEOUT_MAX", 1800),
+        "media_download_min_speed": _parse_int_env("MEDIA_DOWNLOAD_MIN_SPEED", 256 * 1024),
+        # Size of the asyncio default ThreadPoolExecutor. SQLite/python-magic/pickle/os.walk
+        # all run via asyncio.to_thread; the interpreter default (min(32, cpu+4)) is only 5-6
+        # on a 1-2 CPU container, which starves those under load. 32 gives ample headroom.
+        "io_thread_pool_size": _parse_int_env("IO_THREAD_POOL_SIZE", 32),
+    }

dockercompose.yml

@@ -8,6 +8,27 @@ services:
     environment:
       TG_API_ID: XXX
       TG_API_HASH: XXX
+      # TG_PROXY_HOST: 10.0.0.1       # MTProto proxy host (SOCKS5). Prefer a literal IP over a hostname:
+      #                               # a hostname is re-resolved on every reconnect, so flaky DNS during a
+      #                               # network blip can wedge Pyrogram's reconnect.
+      # TG_PROXY_PORT: 1080           # MTProto proxy SOCKS5 port (default: 1080)
+      # TG_PROXY_USERNAME: XXX        # SOCKS5 username (optional)
+      # TG_PROXY_PASSWORD: XXX        # SOCKS5 password (optional)
+      # TG_WATCHDOG_ENABLED: "True"   # Active liveness watchdog: detects 'zombie' sessions and restarts in-process (default: True)
+      # TG_WATCHDOG_INTERVAL: 60      # Seconds between liveness probes (default: 60)
+      # TG_WATCHDOG_TIMEOUT: 10       # Seconds to wait for each get_me probe (default: 10)
+      # TG_WATCHDOG_FAILURES: 3       # Consecutive failed probes before restart (default: 3)
+      # TG_WATCHDOG_HEARTBEAT_EVERY: 30 # Emit an INFO watchdog heartbeat every N successful probes (default: 30)
+      # TG_DISCONNECT_FLAP_LIMIT: 3   # Disconnect events within the flap window before an in-process restart (default: 3)
+      # TG_DISCONNECT_FLAP_WINDOW: 120 # Flap detection window in seconds (default: 120)
+      # TG_CHAT_CACHE_TTL_HOURS: 12   # TTL for cached channel info (title/username/id); removes GetFullChannel from the poll hot path (default: 12)
+      # TG_RPC_CONCURRENCY: 1         # Max concurrent live Telegram RPC calls — global throttle (default: 1)
+      # TG_RPC_MIN_INTERVAL_MS: 500   # Minimum gap between live Telegram RPC starts, ms (default: 500)
+      # TG_RPC_TIMEOUT: 60            # Max seconds a single live Telegram RPC may run before timing out (default: 60)
+      # MEDIA_DOWNLOAD_TIMEOUT_MIN: 120   # Min per-download timeout, seconds — also the timeout for regular (non-large) files (default: 120)
+      # MEDIA_DOWNLOAD_TIMEOUT_MAX: 1800  # Max per-download timeout, seconds — cap for the largest videos (default: 1800)
+      # MEDIA_DOWNLOAD_MIN_SPEED: 262144  # Assumed floor download speed, bytes/s — large-video timeout ≈ file_size / this, clamped to [MIN,MAX] (default: 262144 = 256 KB/s)
+      # IO_THREAD_POOL_SIZE: 32           # Size of the asyncio default threadpool for blocking I/O (SQLite/python-magic/pickle/os.walk); raise on a busy 1-2 CPU box (default: 32)
       PYROGRAM_BRIDGE_URL: https://pgbridge.example.com
       API_PORT: 80
       TOKEN: ХХХ
@@ -32,10 +53,15 @@ services:
       com.centurylinklabs.watchtower.enable: "true"   
       autoheal: true
     healthcheck:
-      test: ["CMD", "curl", "-s", "-f", "-o", "/dev/null", "http://127.0.0.1:80/rss/vvzvlad_lytdybr?limit=1"]
+      # Lightweight process/loop liveness probe. /ping never touches Telegram or the
-      interval: 30m
+      # filesystem, so it answers instantly even while a TG RPC is hung — unlike the old
+      # /rss?limit=1 check, which could exceed the 5s timeout on a cold cache and get the
+      # container restarted mid-download (corrupted temp files). TG liveness is now judged
+      # by the in-process watchdog, which /ping reports via its last-probe age.
+      test: ["CMD", "curl", "-sf", "http://127.0.0.1:80/ping"]
+      interval: 5m
       timeout: 5s
-      retries: 2
+      retries: 3
       start_period: 30s
       start_interval: 5s
 

docs/stability-verification.md

@@ -0,0 +1,148 @@
+# Стадия 7 — сквозная верификация плана стабилизации
+
+Дата: 2026-07-05. Ветка: `fix/stage-7-verification` (на базе `fix/stage-6-healthcheck`,
+кумулятивно содержит стадии 1–6). Эта стадия **ничего не меняет в поведении** прод-кода —
+только автоматизирует проверки и фиксирует, что должен наблюдать оператор после деплоя.
+
+## Итог автотестов
+
+Каноничный прогон (изоляция `config` через `sys.modules`): `python -m pytest tests/`.
+
+```
+260 passed
+```
+
+Раскладка по стадиям (все зелёные):
+
+| Стадия | Файл тестов | Кол-во | Что покрывает |
+|--------|-------------|--------|----------------|
+| 1 — анти-зависания | `tests/test_stage1_hangs.py` | 6 | таймаут RPC-гейта без утечки пермита; воркер переживает Exception/FloodWait, `task_done` сбалансирован; отмена в spacing-ожидании не теряет пермит |
+| 2 — статика/большие видео | `tests/test_stage2_static.py` | 15 | атомарный `_download_atomic` (publish-on-rename, чистка `.part.` при таймауте/zero-size/гонке); дедуп конкурентных скачиваний; FloodWait→429; touch mtime у `temp_*`; свипер чистит `.part.`+legacy `.tmp.`; баланс HTTP-семафора |
+| 3 — FileResponse | `tests/test_stage3_fileresponse.py` | 19 | матрица Range (0-499/500-/-500/за EOF→416/мусор→400/мульти-range→206 multipart); сохранены mtime-touch, delete_after-BackgroundTask, MIME-кэш; чистый ASGI-логгер |
+| 4 — гигиена event loop | `tests/test_stage4_eventloop.py` | 19 | ленивый `raw_message`; вынос side-effect IO из `process_message` (bulk upsert media id); рендер-пайплайн ушёл в поток без create_task/get_running_loop; XSS вычищен во всех 4 выводах ровно одним проходом |
+| 5 — батчинг SQLite | `tests/test_stage5_sqlite.py` | 8 | кэш-хит пишет в аккумулятор, а не в SQLite; flush→DB; snapshot-then-clear не теряет поздние апдейты; re-queue при сбое; str(channel)-ключи |
+| 6 — healthcheck | `tests/test_stage6_healthcheck.py` | 11 | `/ping` 200/503 по connected+age; ноль TG RPC; watchdog отключён→чистая проверка коннекта; `watchdog_last_ok_age()` |
+| 7 — интеграция | `tests/test_stage7_integration.py` | 8 | сквозные сценарии (см. ниже) |
+| — регрессия парсера | `tests/test_postparser_*.py` | 174 | заголовки/флаги/автор (существовавшие до плана) |
+
+## Новые интеграционные тесты стадии 7 (DoD → сквозное доказательство)
+
+`tests/test_stage7_integration.py` драйвит **реальные точки входа** (`get_media`, `ping`,
+flush), чтобы поймать регрессии, которые проявляются только при взаимодействии стадий:
+
+1. **Range на уровне маршрута `/media`** — `test_media_route_range_prefix_0_99`,
+   `_range_suffix_last_100`, `_range_unsatisfiable_416`. Прогоняют `get_media` (кэш-хит) →
+   `prepare_file_response` → `FileResponse` через реальный ASGI: `bytes=0-99`→206 с точным
+   Content-Range/длиной и байтами, `bytes=-100`→206 (суффикс), `bytes=999999999-`→416 (`*/size`).
+   Ловит: если кэш-хит перестанет доходить до FileResponse (ре-буферизация тела, ручные
+   заголовки) или сломается связка digest-гейта — Range перестанет работать. (Стадия 3
+   пинит это на `prepare_file_response` напрямую; здесь — сшивка стадий 2+3.)
+2. **`/ping` быстр и без RPC при висящей операции** — `test_ping_prompt_and_rpc_free_while_slow_op_pending`
+   и `_reports_degraded_promptly_while_slow_op_pending`. Пока фейковая медленная корутина
+   (модель зависшего hot-path) висит на `Event`, `ping()` возвращает 200/503 корректно и
+   **до** завершения медленной операции, при нуле вызовов `get_me`/`safe_get_messages`.
+   Ловит: рекаплинг `/ping` к TG RPC или к любому awaitable, который может застопориться.
+3. **Дедуп + очистка при disconnect через реальный `get_media`** —
+   `test_get_media_concurrent_shares_one_download_and_drains_registry` (2 конкурентных
+   запроса → одна скачка, `_inflight` пуст) и `_request_cancel_does_not_stick_registry_or_hang_sibling`
+   (отмена запроса-«клиента» не оставляет застрявший ключ и не вешает соседа). Ловит:
+   возврат скачивания в корутину запроса (отмена убила бы download) или потерю `finally`-pop.
+4. **str(channel)-ключ access-time end-to-end** — `test_media_cache_hit_flush_updates_str_keyed_db_row`.
+   Кэш-хит `/media` записывает str-ключ, flush обновляет ту же строку в реальной SQLite
+   (hit→flush→DB). Сшивает две половины: ключ, который пишет hot-path, и WHERE, по которому
+   апдейтит flush. Ловит любое расхождение ключа аккумулятора и WHERE bulk-UPDATE (напр.
+   перестановку колонок в SQL — проверено мутацией: тест краснеет, `added` остаётся stale).
+   (int/str-хазард самого канала живёт в `download_media_file` и закрыт стадией 5 —
+   `test_cache_hit_keys_channel_as_str`; здесь покрыта связка get_media+flush.)
+
+## Соответствие DoD стадий → доказательство
+
+- **DoD 1** «ни один путь не ждёт TG без таймаута; воркер не умирает молча» →
+  `test_stage1_hangs.py` (гейт-таймаут, живучесть воркера). Живая проверка supervision
+  под нагрузкой — **наблюдение оператора** (лог `supervised_task_crashed/…_exited`).
+- **DoD 2** «клиенту никогда не отдаётся неполный файл; флуд→429; обрезки не живут >часа» →
+  `test_stage2_static.py` целиком + интеграционный дедуп-тест стадии 7.
+- **DoD 3** «Range-тесты зелёные; поведение эндпоинта эквивалентно (± RFC-допущения)» →
+  `test_stage3_fileresponse.py` + Range на уровне `/media` (стадия 7). «Нет потока-на-чанк»
+  — **наблюдение оператора** (нагрузочный запрос большого файла, число io-потоков).
+- **DoD 4** «генерация 100-сообщ. фида не блокирует луп (параллельный /ping <100 мс); XSS
+  зелёный; media id сохраняются» → `test_stage4_eventloop.py` (рендер в потоке, XSS,
+  bulk upsert) + `/ping`-decoupling стадии 7. Живой замер «<100 мс на проде» —
+  **наблюдение оператора**.
+- **DoD 5** «на кэш-хит /media ноль обращений к SQLite; фоновая запись раз в минуту» →
+  `test_stage5_sqlite.py` + str-ключ end-to-end стадии 7.
+- **DoD 6** «во время зависшего TG RPC /ping отвечает мгновенно (503); контейнер не
+  рестартится от медленного фида» → `test_stage6_healthcheck.py` + `/ping`-under-slow-op
+  стадии 7. Отсутствие autoheal-рестартов на холодном кэше — **наблюдение оператора**.
+
+## Ручные curl-сценарии для оператора (пост-деплой, локально в контейнере)
+
+Подставить реальный `{channel}/{post_id}/{fid}/{digest}` (валидная подпись обязательна).
+
+```bash
+# 1. Range-тройка (ожидания: 206 / 206 / 416)
+curl -s -D- -o /dev/null -H "Range: bytes=0-99"      "http://127.0.0.1:80/media/{ch}/{pid}/{fid}/{digest}"
+curl -s -D- -o /dev/null -H "Range: bytes=-100"       "http://127.0.0.1:80/media/{ch}/{pid}/{fid}/{digest}"
+curl -s -D- -o /dev/null -H "Range: bytes=999999999-" "http://127.0.0.1:80/media/{ch}/{pid}/{fid}/{digest}"
+
+# 2. Параллельные запросы одного БОЛЬШОГО видео (>100 MB), пока не в кэше:
+#    оба должны получить ПОЛНЫЙ файл (одинаковый размер), без частичной отдачи.
+URL="http://127.0.0.1:80/media/{ch}/{pid}/{bigfid}/{digest}"
+curl -s -o /tmp/a "$URL" & curl -s -o /tmp/b "$URL" & wait
+ls -l /tmp/a /tmp/b     # размеры совпадают и равны полному файлу; на диске нет *.part.*
+
+# 3. Обрыв на середине стрима — нет утечки тасков/фд:
+#    считать fd до/после и убедиться, что не растут монотонно.
+lsof -p $(pgrep -f api_server) | wc -l          # baseline
+timeout 2 curl -s -o /dev/null "$URL"; sleep 5  # оборвать скачку на середине (несколько раз)
+lsof -p $(pgrep -f api_server) | wc -l          # не выросло относительно baseline
+
+# 4. Генерация большого фида + параллельный /ping (<100 мс во время генерации):
+curl -s -o /dev/null "http://127.0.0.1:80/rss/{big_channel}" &
+for i in $(seq 1 20); do curl -s -o /dev/null -w "%{time_total}\n" "http://127.0.0.1:80/ping"; done
+wait   # все замеры /ping должны быть заметно < 0.100 s
+```
+
+> Сценарии 2 и 3 (реальная скачка большого видео + реальный подсчёт fd через `lsof`) в
+> headless-тестах **намеренно не подделаны** — им нужен живой сервер, реальные загрузки и
+> реальные файловые дескрипторы. Дедуп-инвариант и disconnect-очистка доказаны на уровне
+> реестра/`get_media` (тесты стадии 7 №3), но «нет утечки fd на проде» проверяет оператор.
+
+## Diag-логи для наблюдения после деплоя
+
+Ожидаемая динамика (сравнить с до-деплойным baseline):
+
+- `diag_semaphore_wait` — ожидание HTTP-семафора должно **упасть** (реже/меньше секунд).
+- `diag_download_timing` — время скачивания стабилизируется; нет длинных «зависаний».
+- `diag_sanitize_slow` — почти **исчезнуть** (один sanitize на выходную границу, стадия 4).
+- `watchdog: heartbeat` — **продолжаются** штатно (живость TG-сессии).
+- На момент FloodWait — **нет всплеска 404** на `/media`; вместо этого `media_flood_wait`
+  и ответы **429** с `Retry-After` (стадия 2.3).
+- Признаки supervision: любые `supervised_task_crashed` / `supervised_task_exited` на
+  CRITICAL — сигнал разобраться, но задача при этом перезапускается (не тихая смерть).
+
+## План отката (rollback)
+
+Единица отката — **стадия целиком**, не отдельный коммит. Каждая стадия живёт на своей
+ветке/PR (`fix/stage-1-hangs` … `fix/stage-7-verification`) и, как правило, состоит из
+**двух коммитов**: feature-коммит + фикс review-раунда (последний нередко чинит реальный
+баг — напр. fail-closed XSS в стадии 4, watchdog-gate в стадии 6). Поэтому `git revert`
+одного коммита осиротит фикс review-раунда и даст несогласованный откат. Откатывать нужно
+на гранулярности стадии; порядок стадий = порядок деплоя.
+
+- **Как откатывать стадию** — зависит от того, как ветки влиты в `fix/stability`/`main`:
+  - если стадия влита **squash-merge** (один коммит на стадию) — `git revert <squash-commit>`
+    откатывает её целиком, однозначно;
+  - если стадия влита **merge-коммитом** — `git revert -m 1 <merge-commit>` откатывает весь
+    вклад ветки (оба коммита) разом;
+  - если история линейная (rebase/fast-forward) — реверт **всех** коммитов стадии
+    (`git revert <feature>..<review-fix>` включительно), а не одного.
+- Стадии 1 и 2 — низкорисковые, деплоятся первыми; откатываются независимо от остальных.
+- Стадия 3 (FileResponse) независима — реверт возвращает прежний ручной стриминг.
+- Стадия 4 требует, чтобы 4.2 откатывалась не позже 4.3 (иначе рендер-в-потоке остался бы
+  без безопасного пути сохранения media id) — откатывать стадию 4 целиком.
+- Стадии 5 и 6 независимы, откатываются по отдельности.
+- Стадия 7 — только тесты и этот документ: реверт ничего не меняет в проде.
+
+Прод-деплой, живое наблюдение diag-логов и обновление прод-compose (healthcheck→`/ping`,
+вне репозитория) — **зона ответственности оператора** (пункты 3–4 плана стадии 7).

file_io.py

@@ -0,0 +1,159 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+# flake8: noqa
+# pylint: disable=missing-function-docstring
+
+import os
+import sqlite3
+from contextlib import contextmanager
+from typing import List
+
+# Path to SQLite database file
+DB_PATH = os.path.join(os.path.abspath("./data"), "media_file_ids.db")
+
+
+def _open_db(db_path: str) -> sqlite3.Connection:
+    """Open a SQLite connection with WAL journal mode enabled."""
+    conn = sqlite3.connect(db_path)
+    conn.execute("PRAGMA journal_mode=WAL")
+    conn.execute("PRAGMA busy_timeout = 5000")  # Wait up to 5 seconds on lock
+    return conn
+
+
+@contextmanager
+def _db_connection(db_path: str):
+    """Context manager that opens a WAL-mode SQLite connection and ensures it is closed."""
+    conn = _open_db(db_path)
+    try:
+        yield conn
+        conn.commit()
+    except Exception:
+        conn.rollback()
+        raise
+    finally:
+        conn.close()
+
+
+def init_db_sync(db_path: str) -> None:
+    """Create the media_file_ids table if it does not exist."""
+    with _db_connection(db_path) as conn:
+        conn.execute(
+            """
+            CREATE TABLE IF NOT EXISTS media_file_ids (
+                channel        TEXT    NOT NULL,
+                post_id        INTEGER NOT NULL,
+                file_unique_id TEXT    NOT NULL,
+                added          REAL    NOT NULL,
+                PRIMARY KEY (channel, post_id, file_unique_id)
+            )
+            """
+        )
+        # Add mime_type column if it does not exist yet (idempotent migration)
+        try:
+            conn.execute("ALTER TABLE media_file_ids ADD COLUMN mime_type TEXT")
+        except sqlite3.OperationalError as e:
+            if "duplicate column name" not in str(e):
+                raise  # Re-raise any OperationalError that is not "duplicate column name"
+
+
+def upsert_media_file_id_sync(db_path: str, channel: str, post_id: int, file_unique_id: str, added: float) -> None:
+    """Insert or replace a single media file ID record."""
+    with _db_connection(db_path) as conn:
+        conn.execute(
+            """INSERT INTO media_file_ids (channel, post_id, file_unique_id, added)
+               VALUES (?, ?, ?, ?)
+               ON CONFLICT(channel, post_id, file_unique_id)
+               DO UPDATE SET added = excluded.added""",
+            (channel, post_id, file_unique_id, added),
+        )
+
+
+def upsert_media_file_ids_bulk_sync(db_path: str, entries: List[tuple]) -> None:
+    """Insert or replace multiple media file ID records in a single transaction.
+
+    entries: iterable of (channel, post_id, file_unique_id, added) tuples.
+    Uses executemany for batched upserts (one connection, one commit).
+    """
+    if not entries:
+        return
+    with _db_connection(db_path) as conn:
+        conn.executemany(
+            """INSERT INTO media_file_ids (channel, post_id, file_unique_id, added)
+               VALUES (?, ?, ?, ?)
+               ON CONFLICT(channel, post_id, file_unique_id)
+               DO UPDATE SET added = excluded.added""",
+            entries,
+        )
+
+
+def update_media_file_access_sync(db_path: str, channel: str, post_id: int, file_unique_id: str, added: float) -> None:
+    """Update the access timestamp for an existing media file ID record."""
+    with _db_connection(db_path) as conn:
+        conn.execute(
+            "UPDATE media_file_ids SET added = ? WHERE channel = ? AND post_id = ? AND file_unique_id = ?",
+            (added, channel, post_id, file_unique_id),
+        )
+
+
+def update_media_file_access_bulk_sync(db_path: str, entries: List[tuple]) -> None:
+    """Update the access timestamp for multiple existing media file ID records.
+
+    entries: iterable of (channel, post_id, file_unique_id, added) tuples.
+    Uses executemany (one connection, one commit) so a batch of cache-hit access
+    updates costs a single SQLite transaction instead of one connect+UPDATE per hit.
+    Rows that do not exist are simply not matched by the WHERE clause (no-op), mirroring
+    the single-row update_media_file_access_sync. An empty batch is a no-op.
+    """
+    if not entries:
+        return
+    with _db_connection(db_path) as conn:
+        conn.executemany(
+            "UPDATE media_file_ids SET added = ? WHERE channel = ? AND post_id = ? AND file_unique_id = ?",
+            # Reorder each (channel, post_id, file_unique_id, added) tuple to match the
+            # UPDATE's placeholder order (added first, then the WHERE key columns).
+            [(added, channel, post_id, file_unique_id)
+             for (channel, post_id, file_unique_id, added) in entries],
+        )
+
+
+def get_all_media_file_ids_sync(db_path: str) -> List[dict]:
+    """Return all rows from media_file_ids as a list of dicts."""
+    with _db_connection(db_path) as conn:
+        conn.row_factory = sqlite3.Row
+        cursor = conn.execute("SELECT channel, post_id, file_unique_id, added FROM media_file_ids")
+        rows = cursor.fetchall()
+    return [dict(row) for row in rows]
+
+
+def remove_media_file_ids_sync(db_path: str, entries: List[tuple]) -> None:
+    """Remove media file ID records identified by (channel, post_id, file_unique_id) tuples."""
+    with _db_connection(db_path) as conn:
+        conn.executemany(
+            "DELETE FROM media_file_ids WHERE channel = ? AND post_id = ? AND file_unique_id = ?",
+            entries,
+        )
+
+
+def get_mime_type_sync(db_path: str, channel: str, post_id: int, file_unique_id: str) -> str | None:
+    """Return the cached MIME type for a given media key, or None if not stored yet."""
+    with _db_connection(db_path) as conn:
+        cursor = conn.execute(
+            "SELECT mime_type FROM media_file_ids WHERE channel = ? AND post_id = ? AND file_unique_id = ?",
+            (channel, post_id, file_unique_id),
+        )
+        row = cursor.fetchone()
+    if row is None:
+        return None
+    return row[0]  # May be None if the column value was never set
+
+
+def set_mime_type_sync(db_path: str, channel: str, post_id: int, file_unique_id: str, mime_type: str) -> None:
+    """Persist a detected MIME type for an existing media file ID record."""
+    with _db_connection(db_path) as conn:
+        conn.execute(
+            "UPDATE media_file_ids SET mime_type = ? WHERE channel = ? AND post_id = ? AND file_unique_id = ?",
+            (mime_type, channel, post_id, file_unique_id),
+        )
+
+

plan.md

@@ -0,0 +1,136 @@
+
+### Где и почему может виснуть
+
+- Синхронный диск/JSON в обработчиках запросов
+  - `_save_media_file_ids` делает `open/json.load/json.dump` прямо в пути запроса, без `to_thread`. При большом `data/media_file_ids.json` это блокирует event loop.
+```950:979:post_parser.py
+                        if os.path.exists(file_path):
+                            with open(file_path, 'r', encoding='utf-8') as f:
+                                existing_data = json.load(f)
+...
+                        with open(file_path, 'w', encoding='utf-8') as f:
+                            json.dump(existing_data, f, ensure_ascii=False, indent=2)
+```
+  - Кеш истории также читает/пишет файлы синхронно:
+```71:99:tg_cache.py
+        with open(cache_file, 'rb') as f:
+            cache_data = pickle.load(f)
+...
+        with open(cache_file, 'wb') as f:
+            pickle.dump(cache_data, f)
+```
+  - Итог: под нагрузкой любой долгий sync I/O останавливает обработку всех запросов на время операции.
+
+- CPU‑тяжёлое формирование RSS/HTML в event loop
+  - Парсинг/санитизация/генерация RSS идут синхронно в корутинах (bleach, feedgen, склейка больших строк). На больших лимитах это “съедает” цикл.
+```185:299:rss_generator.py
+final_posts = await _render_messages_groups(...)
+...
+rss_feed = fg.rss_str(pretty=True)
+```
+```511:533:rss_generator.py
+final_posts = await _render_messages_groups(...)
+...
+html = '\n<hr class="post-divider">\n'.join(html_posts)
+```
+
+- Глобальный `python-magic` в потоках
+  - Один общий `Magic()` используется из разных потоков — это не потокобезопасно, возможны зависания внутри libmagic.
+```198:206:api_server.py
+media_type = await asyncio.to_thread(magic_mime.from_file, file_path)
+```
+
+- BaseHTTPMiddleware над стримингом файлов
+  - `BaseHTTPMiddleware` оборачивает `call_next`. Для `FileResponse` это может ломать/буферизовать стриминг и усиливать задержки.
+```48:62:api_server.py
+class RequestLoggingMiddleware(BaseHTTPMiddleware):
+    async def dispatch(self, request: Request, call_next):
+        ...
+        response = await call_next(request)
+```
+
+- Отсутствие таймаутов на вызовах Telegram API
+  - Если `get_messages/download_media` “подвисли”, запрос к `/media` висит долго.
+```748:758:api_server.py
+file_path, delete_after = await download_media_file(...)
+```
+```232:258:api_server.py
+message = await client.client.get_messages(...)
+file_path = await client.client.download_media(...)
+```
+
+- Шумное логирование в middleware
+  - Лог всех заголовков и запросов при каждом хите может создавать I/O‑бутылочное горлышко под нагрузкой.
+```49:59:api_server.py
+logger.info(f"Request: {request.method} {request.url}")
+logger.info(f"Headers: {dict(request.headers)}")
+```
+
+- Конкурентная запись и разрастание `media_file_ids.json`
+  - Записи без блокировок → гонки и порча JSON. В фоне потом ремонт json-repair (тяжёлый CPU).
+```503:511:api_server.py
+except json.JSONDecodeError:
+    media_files = await asyncio.to_thread(fix_corrupted_json, file_path)
+```
+
+- Один воркер uvicorn
+  - Любая тяжёлая задача “монополизирует” процесс — остальное, включая отдачу файлов, ждёт.
+```114:121:api_server.py
+uvicorn.run(..., reload=True, loop="uvloop")
+```
+
+### Что сделать (короткий чеклист)
+
+- [x] Убрать sync I/O из пути запроса
+  - [x] Обернуть чтение/запись JSON и pickle в `asyncio.to_thread`.
+  - [x] Для `media_file_ids.json` везде использовать единые sync‑хелперы с временным файлом и заменить прямые `open/json.dump`:
+    - [x] Вынести в отдельный модуль функции уже имеющихся `read_json_file_sync`/`write_json_file_sync` и звать их через `to_thread`.
+  - [x] Добавить один `asyncio.Lock` на операции записи `media_file_ids.json` чтобы не было гонок.
+
+- Разгрузить CPU из event loop
+  - В RSS/HTML:
+    - [x] После того как данные получены из Telegram, вынести “склейку HTML”, bleach‑санитизацию и `fg.rss_str` в `asyncio.to_thread`.
+    - Ограничить `limit` по умолчанию до 50–80. Для больших значений — отдавать 429 или 400 с рекомендацией.
+    - Опционально — кэшировать итоговые RSS/HTML на пару минут.
+
+- Исправить `python-magic`
+  - Создавать `Magic()` per‑call внутри `to_thread` ИЛИ защищать глобальный объект `threading.Lock`.
+  - Если не критично — падать на `mimetypes` и не вызывать `magic` для известных расширений.
+
+- Middleware для логов
+  - В проде отключить текущий `RequestLoggingMiddleware` или перевести на ASGI‑middleware без вмешательства в стриминг.
+  - Снизить уровень и объём логов (заголовки — только на DEBUG, семплирование).
+
+- Таймауты телеграма
+  - Оборачивать `get_messages`/`download_media` в `asyncio.wait_for(..., timeout=30–60)` с понятной ошибкой/503.
+  - На FloodWait в `/media` — отдавать 429 с `Retry-After`.
+
+- Запись `media_file_ids.json`
+  - Причесать размер: периодически чистить/ротация по размеру/возрасту.
+  - В фоне перепись файла уже есть — не трогать event loop; убедиться, что на записи тоже используется temp‑файл + `os.replace` (как в `write_json_file_sync`).
+
+- Процессный параллелизм
+  - Запускать с несколькими воркерами, без `reload`:
+    - пример: `uvicorn api_server:app --host 0.0.0.0 --port 80 --workers 2 --loop uvloop --http httptools`
+  - Или за Gunicorn: `--workers 2 --worker-class uvicorn.workers.UvicornWorker`.
+
+- Пул потоков
+  - Явно ограничить/увеличить пул для `to_thread` если есть много фоновых задач, чтобы не выжирался из‑за долгих операций.
+
+### Минимальные целевые правки (по месту)
+
+- `_save_media_file_ids` — переписать на `await asyncio.to_thread(read_json_file_sync/ write_json_file_sync)` + общий `asyncio.Lock`. Убрать прямые `open/json.dump`.
+- `tg_cache` — обернуть `pickle.load/dump` в `to_thread`.
+- `prepare_file_response` — перестать шарить `magic_mime`; создавать локальный `magic.Magic(mime=True)` внутри `to_thread` или защитить глобальный объект `Lock`.
+- `RequestLoggingMiddleware` — удалить/заменить на ASGI‑middleware и логировать только метод+путь, без заголовков по умолчанию.
+- `get_messages/download_media` — таймауты через `asyncio.wait_for`.
+- `uvicorn.run` — убрать `reload=True`, добавить `workers=2+`.
+
+### Как быстро проверить
+- Нагрузить `/rss` c большим `limit` параллельно с `/media` и измерить задержку отдачи файлов до/после:
+  - до фиксов статик “замирает” во время генерации;
+  - после выноса CPU/I/O в `to_thread` и увеличения воркеров — файлы продолжают отдаваться.
+
+Коротко: основные причины — синхронный диск и CPU в корутинах, общий `python-magic`, и единичный воркер. Перенос I/O/CPU в `to_thread`, фиксы гонок JSON, таймауты на Telegram и 2–4 воркера решат подвисания.
+
+- Нашёл проблемные места: sync I/O и CPU в обработчиках, `python-magic`, middleware, таймауты и один воркер. Дал точечные действия для устранения подвисаний.

plans/2026-07-05-stability-fix-plan.md

@@ -0,0 +1,347 @@
+# План стабилизации pyrogram-bridge: статика + зависания
+
+Дата: 2026-07-05. База: аудит api_server.py / telegram_client.py / tg_throttle.py / tg_cache.py / file_io.py / rss_generator.py / post_parser.py.
+
+Проверенные факты, на которых строится план:
+- `FloodWait` — подкласс `RPCError` (проверено на установленном Kurigram).
+- Starlette 0.45.3 поддерживает Range в `FileResponse` из коробки (проверено по исходникам в venv).
+- Дефолтный executor `asyncio.to_thread` = min(32, cpu+4) потоков; в контейнере на 1–2 CPU это 5–6.
+- В venv стоит Kurigram 2.2.4, в requirements закреплён 2.2.22 — локальное окружение не соответствует прод-образу.
+
+## Целевые инварианты (что должно стать верным после ремонта)
+
+1. Любой файл в кэше с финальным именем (`{file_unique_id}` или `temp_{file_unique_id}`) — гарантированно полный. Частичным может быть только `*.part.*`.
+2. Каждый Telegram RPC ограничен таймаутом. Глобальный RPC-гейт не может удерживаться дольше таймаута.
+3. Временные ошибки Telegram (FloodWait) никогда не превращаются в постоянные HTTP-ответы (404).
+4. Event loop не выполняет CPU-работу > ~50 мс за раз.
+5. Каждая фоновая задача supervised: её смерть видна в логах на CRITICAL и/или она перезапускается.
+6. Healthcheck не зависит от Telegram RPC и от обхода файловой системы.
+
+## Процесс
+
+- Ветка `fix/stability` от `main`, коммит на каждую стадию (на `main` не коммитим).
+- Порядок стадий = порядок деплоя: после стадий 1 и 2 уже можно выкатываться и наблюдать.
+- После каждой стадии: pytest + ручные сценарии из стадии 7 + код-ревью.
+- Перед началом: пересоздать/обновить venv под requirements (Kurigram 2.2.22), прогнать 174 существующих теста как baseline. Добавить dev-зависимости: pytest-asyncio, httpx (для TestClient).
+
+---
+
+## Стадия 1 — устранение зависаний (минимальный диф, максимальный эффект)
+
+### 1.1 Таймауты на все RPC под глобальным гейтом
+
+Файлы: tg_cache.py, config.py.
+
+- Добавить в config `tg_rpc_timeout` (env `TG_RPC_TIMEOUT`, default 60).
+- `cached_get_chat_history` (tg_cache.py:142-144): итерацию истории собрать в корутину и обернуть в `asyncio.wait_for` **внутри** `async with tg_rpc()`:
+  ```python
+  async with tg_rpc():
+      async def _collect():
+          # Full paginated history; bounded by the outer wait_for
+          return [m async for m in client.get_chat_history(channel_id, limit=limit)]
+      messages = await asyncio.wait_for(_collect(), timeout=Config["tg_rpc_timeout"])
+  ```
+- `cached_get_chat` (tg_cache.py:216-217): `await asyncio.wait_for(client.get_chat(channel_id), timeout=...)` внутри гейта.
+
+**Где не наебаться:**
+- `wait_for` должен оборачивать **только RPC**, не `_sem.acquire()`. Ожидание в очереди гейта — легитимный backpressure; если холдер ограничен таймаутом, очередь всегда дренируется. Обернёшь acquire — получишь ложные таймауты при штатной очереди из 47 фидов miniflux.
+- `async for` нельзя завернуть в wait_for напрямую — только через промежуточную корутину (как в эскизе).
+- CancelledError не глотать: `except Exception` в вызывающих местах её и так не ловит (это BaseException в 3.11) — не «улучшать» до `except BaseException`.
+
+### 1.2 Таймауты и гейт на остальные живые RPC
+
+- `_reply_enrichment` (rss_generator.py:505): обернуть `client.get_messages` в `async with tg_rpc()` + `wait_for`.
+- `PostParser.get_post` (post_parser.py:98): `wait_for(..., 30)`; заодно удалить `print(...)` (post_parser.py:95).
+- `/health` (api_server.py:906): `wait_for(client.client.get_me(), 10)`.
+
+### 1.3 Починка фонового воркера и очереди
+
+Файл: api_server.py.
+
+- `background_download_worker` (620-637): `task_done()` вызывать только если элемент был реально получен:
+  ```python
+  while True:
+      item = await download_queue.get()   # cancellation propagates cleanly here
+      channel, post_id, file_unique_id = item
+      try:
+          async with BACKGROUND_DOWNLOAD_SEMAPHORE:
+              await download_media_file(channel, post_id, file_unique_id)
+          await asyncio.sleep(2)
+      except errors.FloodWait as e:
+          logger.warning(f"bg_download_floodwait: sleeping {e.value}s")
+          await asyncio.sleep(min(int(e.value) + 5, 900))
+      except Exception as e:
+          logger.error(f"Background download error ...: {e}")
+      finally:
+          download_queue.task_done()
+  ```
+- `download_new_files` (605-609): `await download_queue.put(...)` → `download_queue.put_nowait(...)` — иначе `except asyncio.QueueFull` остаётся мёртвым кодом, а заполненная очередь навсегда блокирует весь `cache_media_files` (включая удаление старых файлов).
+
+**Где не наебаться:**
+- Сейчас смерть воркера убивает и уборку кэша (цепочка: воркер умер → очередь заполнилась → `await put()` завис → свипер больше не крутится). После фикса проверить тестом: воркер, у которого download постоянно бросает исключение, продолжает жить и `queue.join()` завершается.
+- FloodWait в воркере надо ловить **до** Exception и спать, иначе воркер будет молотить телеграм под флудом.
+
+### 1.4 Supervision фоновых задач
+
+- В lifespan повесить `add_done_callback` на `background_task` и `worker_task`: если таск завершился не через CancelledError — лог CRITICAL с exception + перезапуск таска (обёртка `_supervised(factory)` с ограничением частоты рестартов, например не чаще раза в 60 с).
+
+### Тесты стадии 1
+- Гейт: мок «зависшего» RPC (asyncio.Event, который никогда не сеттится) → первый вызов отваливается по таймауту, второй проходит; пермит не утёк.
+- Воркер: download бросает Exception/FloodWait → воркер жив, task_done сбалансирован.
+- `_TgRpcGate`: отмена во время ожидания spacing не теряет пермит (уже реализовано — закрепить тестом).
+
+### DoD стадии 1
+Ни один путь кода не ждёт Telegram без таймаута; воркер не умирает молча; pytest зелёный.
+
+---
+
+## Стадия 2 — статика: большие видео, FloodWait, семафор
+
+### 2.1 Единый путь скачивания с атомарным rename (главный фикс флаки)
+
+Файл: api_server.py, telegram_client.py.
+
+Сейчас большие видео (>100MB) качаются напрямую в финальное имя `temp_{fid}` (379-396): конкурентный запрос видит частичный файл и отдаёт его; при таймауте обрезок не удаляется и целый час отдаётся как «готовый».
+
+- Выделить хелпер:
+  ```python
+  async def _download_atomic(file_id: str, final_path: str, timeout: float) -> str:
+      # Download to a unique partial path, validate, atomically rename.
+      part_path = f"{final_path}.part.{uuid.uuid4().hex}"
+      try:
+          await client.safe_download_media(file_id, part_path, timeout=timeout)
+          if not os.path.exists(part_path) or os.path.getsize(part_path) == 0:
+              raise ZeroSizeFileError(...)
+          if not os.path.exists(final_path):
+              os.rename(part_path, final_path)  # atomic on POSIX
+          return final_path
+      finally:
+          # Always clean up our partial file (timeout, cancel, race loser)
+          if os.path.exists(part_path):
+              try: os.remove(part_path)
+              except OSError: pass
+      ```
+- Использовать его и для обычных файлов, и для больших видео (финальное имя больших остаётся `temp_{fid}` — семантика «не кэшировать постоянно, чистить по TTL» сохраняется).
+- `safe_download_media` научить принимать `timeout` параметром; для больших видео таймаут от размера: `min(1800, max(120, file_size // (256*1024)))` (≈256 KB/s минимально допустимая скорость), env-конфигурируемо.
+- Проверку `if os.path.exists(temp_file_path): return` (382-384) заменить: существующий `temp_{fid}` теперь **гарантированно полный** (появляется только через rename) — отдавать можно смело.
+
+**Где не наебаться:**
+- Единый суффикс `.part.{hex}` вместо старого `.tmp.{hex}`. Обновить regex свипера (api_server.py:536) на новый суффикс, **оставив** и старый паттерн на переходный период (на диске могут лежать старые обрезки).
+- Инвариант «финальное имя = полный файл» ломается, если хоть один путь пишет мимо `.part.` — после правки grep-ом убедиться, что `download_media` больше нигде не получает финальный путь.
+- В `download_new_files` (598-600) проверка «temp существует → скипаем» остаётся корректной: полный файл есть — качать не надо. Частичные `.part.` под неё не попадают — это правильно.
+- Первый запрос большого видео всё ещё отвечает только после полного скачивания (минуты). Это осознанное ограничение; прогрессивный стриминг через `client.stream_media` — отдельная большая фича, в этот план не входит (пометить как future work).
+
+### 2.2 Дедупликация конкурентных скачиваний (in-flight registry)
+
+- Модульный `_inflight: dict[tuple[str, int, str], asyncio.Future]`.
+- Первый запрос: создаёт future, запускает скачивание **отдельным** `asyncio.create_task` (не в контексте HTTP-запроса!), в finally сеттит result/exception и удаляет ключ. Остальные: `await`ят future.
+
+**Где не наебаться (классическая ловушка):**
+- Если качать прямо в корутине первого HTTP-запроса, отключение его клиента отменит скачивание, future никогда не завершится, и все ожидающие повиснут. Поэтому: скачивание — в detached task; ожидающие делают `await asyncio.wait_for(fut, timeout)`.
+- В finally у task обязательно и `set_exception`, и `pop` ключа — иначе после первой ошибки ключ навсегда «занят» отработавшим future.
+- `set_exception` на future, который никто не await-ит, даст "exception was never retrieved" warning — допустимо, но можно гасить через `fut.exception()` в done-callback.
+
+### 2.3 FloodWait → 429 в /media
+
+- В `get_media` добавить обработчик **до** `except errors.RPCError` (api_server.py:1001):
+  ```python
+  except errors.FloodWait as e:
+      retry_after = min(int(e.value) + random.randint(1, 30), 300)
+      return Response(status_code=429, content="Telegram flood wait",
+                      headers={"Retry-After": str(retry_after)})
+  ```
+
+**Где не наебаться:** порядок except-веток решает: FloodWait — подкласс RPCError, поставишь после — не сработает никогда. Закрепить тестом (мок download_media_file бросает FloodWait → ответ 429, не 404).
+
+### 2.4 Не удалять большое видео «из-под зрителя»
+
+- При каждой отдаче файла с именем `temp_*` — `await asyncio.to_thread(os.utime, path)` (touch mtime). Свипер (порог 1 час по mtime) перестанет удалять активно просматриваемые файлы.
+
+### 2.5 Ограничить ожидание HTTP-семафора
+
+- Вместо `async with HTTP_DOWNLOAD_SEMAPHORE`: явный `acquire` под `wait_for(30)`; таймаут → 503 + `Retry-After: 30`. Release — только если acquire удался (try/finally вокруг критической секции, а не вокруг acquire).
+
+### Тесты стадии 2
+- Конкурентные запросы большого видео с медленным фейковым download: второй запрос НЕ получает частичный файл (либо ждёт future, либо 503/504), после завершения оба получают полный.
+- Таймаут download → `.part.` удалён, финального имени нет.
+- FloodWait → 429 c Retry-After.
+- Свипер: чистит и `.part.`, и старые `.tmp.`, не трогает свежие.
+
+### DoD стадии 2
+Ни при каком сценарии клиенту не может быть отдан неполный файл; флуд отдаёт 429; обрезки не живут дольше часа.
+
+---
+
+## Стадия 3 — отдача файлов через FileResponse, чистка HTTP-слоя
+
+### 3.1 Заменить самодельный стриминг на `FileResponse`
+
+Файл: api_server.py (`prepare_file_response`, 203-335).
+
+Starlette 0.45.3 сам обрабатывает Range/If-Range/206/416/multipart, ставит Accept-Ranges/ETag/Last-Modified и читает файл эффективно (без нашего to_thread на каждые 64KB):
+
+```python
+return FileResponse(
+    file_path,
+    media_type=media_type,
+    filename=os.path.basename(file_path),        # sets Content-Disposition: inline; filename*
+    content_disposition_type="inline",
+    headers={"Cache-Control": "public, max-age=86400, immutable"},
+    background=background,
+)
+```
+
+Оставить: пре-чек существования (404), MIME-логику (magic + SQLite-кэш типа). Удалить: ручной парсинг Range, `file_chunk_generator`, ручные заголовки.
+
+**Где не наебаться:**
+- **Сначала тесты, потом свап.** Написать httpx TestClient-тесты на текущее поведение (bytes=0-499, bytes=500-, bytes=-500, start за EOF → 416, мусорный заголовок), затем перейти на FileResponse и осознанно принять расхождения (например, Starlette на кривой заголовок может отдать 200-полный вместо 416 — это допустимо по RFC 7233).
+- FileResponse проверяет файл на этапе отправки: пре-чек на 404 оставить обязательно.
+- `filename*=UTF-8''...` FileResponse формирует сам — руками Content-Disposition не собирать, иначе задвоится.
+
+### 3.2 Убрать BaseHTTPMiddleware
+
+- `RequestLoggingMiddleware` (api_server.py:54-64) удалить. Логирование запросов: либо `--access-log` uvicorn на debug-уровне, либо чистый ASGI-middleware из ~10 строк (без BaseHTTPMiddleware): меньше оверхеда на каждый чанк стриминга и никаких сюрпризов с отменой/фоновыми тасками.
+
+### 3.3 Расширить дефолтный executor
+
+- В lifespan: `loop.set_default_executor(ThreadPoolExecutor(max_workers=32, thread_name_prefix="io"))`. Даже после ухода per-chunk чтений остаются SQLite/magic/pickle/os.walk — 5–6 дефолтных потоков в контейнере мало.
+
+### DoD стадии 3
+Range-тесты зелёные; profiling-запрос большого файла не порождает поток-на-чанк; поведение эндпоинта эквивалентно (кроме осознанных RFC-допущений).
+
+---
+
+## Стадия 4 — гигиена event loop (генерация фидов)
+
+### 4.1 `raw_message` — лениво
+
+- `process_message(..., include_raw: bool = False)`: `str(message)` (полная сериализация каждого поста! post_parser.py:523) выполнять только для JSON-выдачи и debug-HTML. В фидах — никогда.
+
+### 4.2 Убрать side-effect IO из process_message (обезвредить ловушку до 4.3)
+
+Сейчас `_generate_html_media` → `_save_media_file_ids` (post_parser.py:676, 977-1028) делает `asyncio.get_running_loop()` + `create_task`. **Если перенести рендер в поток (4.3) без этой правки, `get_running_loop()` бросит RuntimeError, ветка упадёт в `except` с логом — и записи media id молча перестанут сохраняться, а фоновый прогрев кэша умрёт.**
+
+- Рефакторинг: `_save_media_file_ids` не пишет в БД, а складывает `(channel, post_id, file_unique_id, ts)` в `self._pending_media_ids` инстанса PostParser (инстанс создаётся на запрос — потокобезопасно, т.к. рендер идёт в одном потоке).
+- После рендера вызывающий код (rss_generator / get_post) один раз делает `await asyncio.to_thread(upsert_media_file_ids_bulk_sync, DB_PATH, entries)` — новая функция в file_io.py с `executemany`. Это заодно даёт батчинг апсертов (часть стадии 5).
+- Счётчик `_persist_pending_count` и create_task-механику удалить.
+
+### 4.3 Рендер фида — в поток
+
+- `_create_time_based_media_groups`, `_create_messages_groups`, `_trim_messages_groups`, `_render_messages_groups` — внутри нет ни одного await; превратить в обычные sync-функции и выполнять одним `await asyncio.to_thread(_render_pipeline, ...)` из `generate_channel_rss` / `generate_channel_html`.
+- `copy.deepcopy(messages)` (rss_generator.py:38) уезжает в поток вместе со всем остальным.
+
+**Где не наебаться:**
+- GIL: перенос CPU-работы в поток не делает её бесплатной — луп перестаёт стоять колом, но замедляется. Поэтому 4.1/4.4 (сокращение работы) обязательны, а не опциональны.
+- Внутри потока не должно остаться ни `create_task`, ни `get_running_loop` (это ровно 4.2), ни обращений к asyncio-примитивам.
+- deepcopy живых Pyrogram-объектов сегодня работает — код не менять, только переместить; отдельным тестом убедиться, что deepcopy Message из pickle-кэша не падает. Полный отказ от deepcopy (не-мутирующая группировка через map message_id → group_id) — отдельный опциональный рефакторинг, в эту стадию не тащить.
+
+### 4.4 Сократить sanitize-проходы
+
+Сейчас на каждое сообщение bleach вызывается 4+ раз (body: post_parser.py:672, media: 752, footer: 833, reactions: 920), плюс финальный проход по всему фиду. Bleach — самая дорогая CPU-часть (есть diag_sanitize_slow > 50 мс на вызов).
+
+- Правило: **один sanitize на каждую выходную границу**, не на каждый фрагмент:
+  - RSS/HTML-фиды: финальный проход уже есть (rss_generator.py:443, 643) → внутренние per-message проходы убрать.
+  - `/html/{channel}/{post_id}`: единственный проход добавить в `_format_html` (или в эндпоинте) — сейчас он полагается на внутренние.
+  - `/json/...`: html-поля внутри JSON должны оставаться чищеными → один проход по body+footer в `process_message`.
+
+**Где не наебаться (XSS!):**
+- Перед удалением внутренних проходов построить карту «выходная точка → где чистится», и только потом резать. Ни один выходной путь не должен остаться без ровно одного прохода.
+- Добавить тесты безопасности: мок-сообщение с `<script>`, `onerror=`, `javascript:`-ссылкой → во всех трёх выводах (rss, html, json) — вычищено.
+- Помнить, что `debug=true` HTML вставляет `raw_message` в `<pre>` без экранирования — при 4.1 заодно прогнать через `html.escape`.
+
+### DoD стадии 4
+Генерация фида на 100 сообщений не блокирует луп заметно (проверка: параллельный запрос `/ping` из стадии 6 отвечает < 100 мс во время генерации); XSS-тесты зелёные; media id продолжают сохраняться (тест на bulk upsert после рендера).
+
+---
+
+## Стадия 5 — SQLite: батчинг вместо записи на каждый чих
+
+### 5.1 Access-time аккумулятор
+
+- Сейчас каждый кэш-хит /media = поток + connect + UPDATE (api_server.py:963-973). Заменить на модульный `dict[(channel, post_id, fid)] = ts` (обновление словаря на лупе — дёшево и атомарно в asyncio), и периодический flush-таск раз в 60 с: `executemany UPDATE` одним соединением.
+- Flush также в shutdown (lifespan) — не терять хвост.
+- Убрать fire-and-forget `asyncio.create_task(_update_access(...))` целиком.
+
+### 5.2 Батч-апсерты media id
+
+- Сделаны в 4.2 (`upsert_media_file_ids_bulk_sync`).
+
+**Где не наебаться:**
+- Flush-таск — под supervision из 1.4.
+- Ключи как и раньше: channel — строкой (`str(channel)`), не смешивать str/int в ключах — иначе UPDATE не найдёт строку и timestamp тихо перестанет обновляться (файлы начнут выпадать из кэша через 20 дней при живом трафике).
+- Порядок в тесте: хит → flush → значение `added` в БД обновилось.
+
+### DoD стадии 5
+На кэш-хит /media — ноль обращений к SQLite в горячем пути (кроме кэша MIME); фоновая запись раз в минуту.
+
+---
+
+## Стадия 6 — healthcheck и деплой
+
+### 6.1 Лёгкий `/ping`
+
+- Новый эндпоинт без токена, без TG RPC, без os.walk:
+  ```python
+  @app.get("/ping")
+  async def ping():
+      age = client.watchdog_last_ok_age()   # seconds since last successful probe, None if never
+      healthy = client.client.is_connected and (age is None or age < unhealthy_threshold)
+      return JSONResponse({"status": "ok" if healthy else "degraded", ...},
+                          status_code=200 if healthy else 503)
+  ```
+- В TelegramClient добавить публичный метод возраста `_wd_last_ok_monotonic`.
+
+### 6.2 Обновить docker-compose пример
+
+- healthcheck → `curl -sf http://127.0.0.1:80/ping`, interval 5m, timeout 5s.
+- Прод-компоуз (вне репозитория) обновить руками — отметить в отчёте деплоя.
+
+**Где не наебаться:**
+- Текущий healthcheck (`/rss/...?limit=1`, timeout 5 с) сам создаёт флаки: на холодном кэше генерация легко > 5 с → autoheal рестартует контейнер посреди докачек → битые temp-файлы. После перехода на /ping «живость TG» проверяет вотчдог, а healthcheck — только живость процесса/лупа. Не оставлять старый URL в проде.
+- `/ping` не должен дергать `get_me()` — иначе вернули ту же проблему через другую дверь.
+
+### DoD стадии 6
+Во время зависшего TG RPC `/ping` отвечает мгновенно (503 degraded), контейнер не рестартится от медленного фида.
+
+---
+
+## Стадия 7 — сквозная верификация
+
+1. Полный pytest (старые 174 + новые).
+2. Ручные сценарии (локально, curl):
+   - Range: `curl -H "Range: bytes=0-99" / "bytes=-100" / "bytes=999999999-"` → 206/206/416.
+   - Параллельные запросы одного большого видео (fake slow client) → нет частичной отдачи.
+   - Отключение клиента на середине стрима → нет утечки тасков/фд (смотреть логи и lsof).
+   - Генерация фида на 100+ сообщений + параллельный /ping → ping < 100 мс.
+3. Деплой на прод и наблюдение по существующим diag-логам:
+   - `diag_semaphore_wait`, `diag_download_timing` — ожидание должно упасть;
+   - `diag_sanitize_slow` — должен почти исчезнуть;
+   - `watchdog: heartbeat` — продолжаются;
+   - отсутствие 404-всплесков на /media в момент FloodWait (теперь 429).
+4. Rollback-план: каждая стадия — отдельный коммит (или PR) → откатывается индивидуально.
+
+---
+
+## Сводка главных ловушек (checklist перед каждым ревью)
+
+1. `wait_for` — вокруг RPC, не вокруг acquire гейта/семафора.
+2. In-flight future: скачивание в detached task, `set_exception` + `pop` в finally, иначе вечно занятый ключ или зависшие ожидающие.
+3. `_save_media_file_ids` внутри потока = молчаливая потеря записей (RuntimeError → except → лог). Сначала 4.2, потом 4.3.
+4. Убирая sanitize-проходы — карта выходных точек; каждый выход имеет ровно один проход; XSS-тесты.
+5. Перенос в поток ≠ ускорение (GIL): обязательно сокращать объём работы (lazy raw_message, один sanitize).
+6. FileResponse: сначала зафиксировать текущую Range-семантику тестами, потом менять.
+7. Переименование `.tmp.` → `.part.`: обновить regex свипера, старый паттерн оставить на переходный период.
+8. `task_done()` только после успешного `get()`; `put_nowait` вместо `await put()` там, где ловится QueueFull.
+9. Ключи SQLite: channel всегда `str(...)` — рассинхрон типов тихо ломает обновление timestamp.
+10. Большие видео: touch mtime при отдаче, иначе свипер удалит файл под зрителем.
+11. Venv ≠ прод: выровнять Kurigram до 2.2.22 до начала работ.
+12. Ветка `fix/stability`; на `main` не коммитить.
+
+## Порядок и зависимости
+
+- Стадия 1 → деплой возможен сразу (низкий риск, убирает зависания).
+- Стадия 2 → деплой вторым (убирает флаки статики). Зависит от 1.3 (FloodWait в воркере).
+- Стадия 3 — независима, но лучше после 2 (меньше конфликтов в prepare_file_response/download-путях).
+- Стадия 4 требует 4.2 строго до 4.3; 4.4 можно отдельным коммитом.
+- Стадия 5 частично делается в 4.2; аккумулятор access-time — независим.
+- Стадия 6 — в любой момент, но эффект от неё максимален после 1–2.

plans/performance-blocking-analysis.md

@@ -0,0 +1,619 @@
+# Critical Performance and Blocking Analysis: Pyrogram Bridge
+
+**Date**: 2026-01-01  
+**Severity**: Critical  
+**Status**: Analysis Complete - Awaiting Implementation Approval
+
+---
+
+## Executive Summary
+
+The pyrogram-bridge project experiences critical blocking issues preventing static file delivery to clients. The root cause is **concurrent authentication failures in Pyrogram** when multiple download requests occur simultaneously, causing the entire event loop to hang while authentication repeatedly fails with `KeyError: 0`.
+
+### Key Findings:
+1. **Primary Issue**: Pyrogram auth key creation failures block all concurrent operations
+2. **Impact**: Multiple HTTP requests wait indefinitely for downloads that never complete
+3. **Pattern**: Background downloads and on-demand downloads compete for authentication resources
+4. **Scale**: Affects ~30-40% of download attempts during peak loads
+
+---
+
+## 1. Log Analysis and Pattern Identification
+
+### 1.1 Critical Error Pattern
+
+The logs show a repeating pattern of authentication failures:
+
+```
+Line 14:  pyrogram.session.auth - INFO - Retrying due to KeyError: 0
+Line 15:  pyrogram.connection.connection - INFO - Disconnected
+Line 16:  pyrogram.session.auth - INFO - Start creating a new auth key on DC4
+Line 17:  pyrogram.connection.connection - INFO - Connecting...
+Line 18:  pyrogram.connection.connection - INFO - Connected! Production DC4 - IPv4
+[~11 seconds pass]
+Line 23:  pyrogram.session.auth - INFO - Retrying due to KeyError: 0
+```
+
+**Key Observations**:
+- Authentication retries occur every ~10-12 seconds
+- Each retry cycle involves: disconnect → reconnect → KeyError
+- After 5-6 retry attempts, downloads fail with zero-size files
+- HTTP requests that arrive during this period hang indefinitely
+
+### 1.2 Timeline Analysis
+
+#### Request Flow (Line 10-71):
+```
+13:47:52 - Request arrives for media file
+13:47:52 - Valid digest checked
+13:48:01 - First KeyError: 0 in pyrogram auth
+13:48:02 - Disconnected, starts creating new auth key
+13:48:02 - Connected to DC4
+13:48:12 - Second KeyError: 0
+[Multiple retry cycles...]
+13:48:59 - Final KeyError: 0 exception raised
+13:48:59 - ERROR: Downloaded file is zero size
+13:49:00 - File successfully downloads (after auth stabilizes)
+```
+
+**Timeline Duration**: 67 seconds from request to successful delivery  
+**Expected Duration**: 2-5 seconds for cached download
+
+### 1.3 Concurrent Request Impact
+
+Lines 194-244 show multiple simultaneous requests:
+- 10 parallel HTTP GET requests arrive within 6 seconds (lines 194-234)
+- Background cache task attempts 3 concurrent downloads
+- Pyrogram authentication locks prevent progress on all requests
+- Successful cached files are served (lines 242-244)
+- New downloads hang for 60+ seconds
+
+---
+
+## 2. Root Cause Analysis
+
+### 2.1 Pyrogram Authentication Architecture
+
+#### Problem: Non-Reentrant Auth Key Creation
+
+From [`telegram_client.py:30-35`](telegram_client.py:30):
+
+```python
+self.client = Client(
+    name="pyro_bridge",
+    api_id=settings["tg_api_id"],
+    api_hash=settings["tg_api_hash"],
+    workdir=settings["session_path"],
+)
+```
+
+**Issue**: Single Pyrogram client instance shared across all operations:
+- One client for API server requests (line 68 in [`api_server.py`](api_server.py:68))
+- Same client for background downloads (line 80 in [`api_server.py`](api_server.py:80))
+- No session pooling or connection reuse strategy
+
+#### The KeyError: 0 Root Cause
+
+The `KeyError: 0` occurs in Pyrogram's TLObject deserialization:
+
+```python
+File "/usr/local/lib/python3.11/site-packages/pyrogram/raw/core/tl_object.py", line 33
+return cast(TLObject, objects[int.from_bytes(b.read(4), "little")]).read(b, *args)
+                      ~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+KeyError: 0
+```
+
+This happens when:
+1. Multiple concurrent operations attempt to establish auth keys
+2. Telegram sends a response that doesn't match expected protocol format
+3. Likely caused by race condition in auth key negotiation
+4. Results in 10-second retry cycle per attempt
+
+### 2.2 Critical Blocking Points
+
+#### 2.2.1 Download Media Function (Lines 216-341 in [`api_server.py`](api_server.py:216))
+
+```python
+async def download_media_file(channel: Union[str, int], post_id: int, file_unique_id: str) -> tuple[Union[str, None], bool]:
+    # ... setup code ...
+    
+    message = await client.client.get_messages(channel_id, post_id)  # BLOCKING POINT 1
+    
+    # ... cache check ...
+    
+    file_id = await find_file_id_in_message(message, file_unique_id)
+    file_path = await client.client.download_media(file_id, file_name=cache_path)  # BLOCKING POINT 2
+```
+
+**Problems**:
+1. **Direct await on Pyrogram calls**: No timeout protection
+2. **No concurrency limits**: Unlimited simultaneous downloads
+3. **Shared client state**: All operations block when auth fails
+4. **No circuit breaker**: Failed auth retries indefinitely
+
+#### 2.2.2 Background Cache Task (Lines 494-532 in [`api_server.py`](api_server.py:494))
+
+```python
+async def cache_media_files() -> None:
+    delay = 60
+    while True:
+        # ... cache cleanup ...
+        await download_new_files(updated_media_files, cache_dir)  # BLOCKING POINT 3
+        await asyncio.sleep(delay)
+```
+
+**Problems**:
+1. Runs concurrently with HTTP request handlers
+2. Can trigger multiple downloads simultaneously (line 457)
+3. Shares same Pyrogram client with request handlers
+4. No coordination with active requests
+
+#### 2.2.3 Race Condition Visualization
+
+```mermaid
+graph TD
+    A[HTTP Request 1] --> B[download_media_file]
+    C[HTTP Request 2] --> D[download_media_file]
+    E[Background Task] --> F[download_new_files]
+    
+    B --> G[client.client.get_messages]
+    D --> H[client.client.get_messages]
+    F --> I[client.client.get_messages]
+    
+    G --> J{Shared Pyrogram Client}
+    H --> J
+    I --> J
+    
+    J --> K[Auth Key Creation]
+    K --> L{KeyError: 0}
+    L --> M[10s Retry]
+    M --> K
+```
+
+### 2.3 Async/Sync Execution Issues
+
+#### Issue 1: Blocking Operations in Async Context
+
+[`api_server.py:324`](api_server.py:324):
+```python
+file_path = await client.client.download_media(file_id, file_name=cache_path)
+```
+
+Pyrogram's `download_media()` is async but internally performs:
+- Network I/O (potentially blocking on slow connections)
+- File system writes (not truly async)
+- Auth key operations (blocks entire client)
+
+#### Issue 2: No Timeout Protection
+
+None of the Pyrogram calls have timeout wrappers:
+```python
+# No timeout on message fetch
+message = await client.client.get_messages(channel_id, post_id)
+
+# No timeout on file download
+file_path = await client.client.download_media(file_id, file_name=cache_path)
+```
+
+A single stuck operation blocks all others sharing the client.
+
+#### Issue 3: Thread Pool Misuse
+
+[`api_server.py:201`](api_server.py:201):
+```python
+media_type = await asyncio.to_thread(magic_mime.from_file, file_path)
+```
+
+Good practice here (CPU-bound work in thread pool), but Pyrogram operations should also be isolated.
+
+---
+
+## 3. Architectural Problems
+
+### 3.1 Violated Best Practices
+
+#### ❌ Single Client Instance for All Operations
+
+**Current**: One Pyrogram client handles all requests  
+**Problem**: Client state shared across all operations  
+**Best Practice**: Separate clients or connection pooling
+
+#### ❌ No Concurrency Control
+
+**Current**: Unlimited simultaneous downloads  
+**Problem**: Resource exhaustion and auth conflicts  
+**Best Practice**: Semaphore-limited concurrent operations
+
+#### ❌ No Circuit Breaker Pattern
+
+**Current**: Infinite retries on auth failure  
+**Problem**: Cascading failures block all requests  
+**Best Practice**: Fast-fail after N attempts, exponential backoff
+
+#### ❌ Synchronous I/O in Async Context
+
+**Current**: File writes happen in async functions  
+**Problem**: Blocks event loop during large file writes  
+**Best Practice**: Use `aiofiles` or thread pool for I/O
+
+#### ❌ No Request Coordination
+
+**Current**: Background task and request handlers compete  
+**Problem**: Auth conflicts from concurrent operations  
+**Best Practice**: Queue-based download coordination
+
+### 3.2 Connection Management Issues
+
+#### Current Architecture:
+```
+FastAPI (uvicorn) → Single TelegramClient → Single Pyrogram Client
+                          ↓
+                    All operations share this client
+```
+
+#### Problems:
+1. **No session isolation**: All requests share auth state
+2. **No connection pool**: Single TCP connection to Telegram
+3. **No retry strategy**: Each operation independently retries
+4. **No health checking**: No way to detect dead connections
+
+---
+
+## 4. Proposed Solutions
+
+### 4.1 Immediate Fixes (Critical Priority)
+
+#### Fix 1: Add Concurrency Limiter
+
+**File**: [`api_server.py`](api_server.py:68)
+
+**Problem**: Unlimited concurrent Pyrogram operations cause auth conflicts
+
+**Solution**:
+```python
+# Add at global level after client initialization
+DOWNLOAD_SEMAPHORE = asyncio.Semaphore(3)  # Max 3 concurrent downloads
+
+async def download_media_file(channel: Union[str, int], post_id: int, file_unique_id: str) -> tuple[Union[str, None], bool]:
+    async with DOWNLOAD_SEMAPHORE:  # Limit concurrent downloads
+        base_cache_dir = os.path.abspath("./data/cache")
+        # ... rest of function
+```
+
+**Impact**: Reduces auth conflicts by limiting concurrent Telegram API calls
+
+#### Fix 2: Add Timeout Protection
+
+**File**: [`api_server.py`](api_server.py:233)
+
+**Problem**: Operations hang indefinitely on auth failures
+
+**Solution**:
+```python
+async def download_media_file(channel: Union[str, int], post_id: int, file_unique_id: str) -> tuple[Union[str, None], bool]:
+    async with DOWNLOAD_SEMAPHORE:
+        try:
+            # Add timeout wrapper for Telegram operations
+            message = await asyncio.wait_for(
+                client.client.get_messages(channel_id, post_id),
+                timeout=30.0  # 30 second timeout
+            )
+            
+            # ... file ID lookup ...
+            
+            file_path = await asyncio.wait_for(
+                client.client.download_media(file_id, file_name=cache_path),
+                timeout=120.0  # 2 minute timeout for downloads
+            )
+            
+        except asyncio.TimeoutError:
+            logger.error(f"Timeout downloading {channel}/{post_id}/{file_unique_id}")
+            raise HTTPException(status_code=504, detail="Download timeout")
+```
+
+**Impact**: Prevents indefinite hangs, allows other requests to proceed
+
+#### Fix 3: Separate Background Download Queue
+
+**File**: [`api_server.py`](api_server.py:425)
+
+**Problem**: Background task competes with request handlers
+
+**Solution**:
+```python
+# Add queue for background downloads
+download_queue = asyncio.Queue(maxsize=100)
+
+async def download_new_files(media_files: list, cache_dir: str) -> None:
+    """Queue files for background download instead of downloading immediately"""
+    for file_data in media_files:
+        # ... validation ...
+        
+        cache_path = os.path.join(post_dir, file_unique_id)
+        if not os.path.exists(cache_path):
+            try:
+                await download_queue.put((channel, post_id, file_unique_id))
+            except asyncio.QueueFull:
+                logger.warning(f"Download queue full, skipping {channel}/{post_id}/{file_unique_id}")
+                break
+
+async def background_download_worker():
+    """Worker that processes downloads from queue"""
+    while True:
+        try:
+            channel, post_id, file_unique_id = await download_queue.get()
+            logger.info(f"Background download: {channel}/{post_id}/{file_unique_id}")
+            
+            async with DOWNLOAD_SEMAPHORE:  # Use same semaphore
+                await download_media_file(channel, post_id, file_unique_id)
+                await asyncio.sleep(2)  # Rate limiting
+                
+        except Exception as e:
+            logger.error(f"Background download error: {e}")
+        finally:
+            download_queue.task_done()
+```
+
+**Impact**: Coordinates background and on-demand downloads, prevents conflicts
+
+#### Fix 4: Improve Error Handling for KeyError: 0
+
+**File**: [`telegram_client.py`](telegram_client.py:62)
+
+**Problem**: No handling for Pyrogram auth errors
+
+**Solution**:
+```python
+# Add in TelegramClient class
+async def safe_get_messages(self, channel_id, post_id, max_retries=2):
+    """Wrapper with retry logic for auth errors"""
+    for attempt in range(max_retries):
+        try:
+            return await asyncio.wait_for(
+                self.client.get_messages(channel_id, post_id),
+                timeout=30.0
+            )
+        except Exception as e:
+            if "KeyError" in str(e) and attempt < max_retries - 1:
+                logger.warning(f"Auth error on attempt {attempt + 1}, retrying in 5s...")
+                await asyncio.sleep(5)
+                continue
+            raise
+    
+async def safe_download_media(self, file_id, file_name, max_retries=2):
+    """Wrapper with retry logic for download errors"""
+    for attempt in range(max_retries):
+        try:
+            return await asyncio.wait_for(
+                self.client.download_media(file_id, file_name=file_name),
+                timeout=120.0
+            )
+        except Exception as e:
+            if "KeyError" in str(e) and attempt < max_retries - 1:
+                logger.warning(f"Download auth error on attempt {attempt + 1}, retrying...")
+                await asyncio.sleep(5)
+                continue
+            raise
+```
+
+**Impact**: Isolates auth errors, provides controlled retry behavior
+
+### 4.2 Medium-Term Improvements (High Priority)
+
+#### Improvement 1: Connection Pool Architecture
+
+**New File**: `telegram_pool.py`
+
+```python
+import asyncio
+from typing import List
+from pyrogram import Client
+from config import get_settings
+
+class TelegramClientPool:
+    """Pool of Pyrogram clients for load distribution"""
+    
+    def __init__(self, pool_size: int = 3):
+        self.pool_size = pool_size
+        self.clients: List[Client] = []
+        self.current_index = 0
+        self.lock = asyncio.Lock()
+        
+    async def initialize(self):
+        """Create and start pool of clients"""
+        settings = get_settings()
+        
+        for i in range(self.pool_size):
+            client = Client(
+                name=f"pyro_bridge_{i}",
+                api_id=settings["tg_api_id"],
+                api_hash=settings["tg_api_hash"],
+                workdir=settings["session_path"],
+            )
+            await client.start()
+            self.clients.append(client)
+            logger.info(f"Initialized client {i+1}/{self.pool_size}")
+    
+    async def get_client(self) -> Client:
+        """Get next available client (round-robin)"""
+        async with self.lock:
+            client = self.clients[self.current_index]
+            self.current_index = (self.current_index + 1) % self.pool_size
+            return client
+    
+    async def shutdown(self):
+        """Stop all clients"""
+        for client in self.clients:
+            await client.stop()
+```
+
+**Usage in [`api_server.py`](api_server.py:68)**:
+```python
+# Replace single client with pool
+client_pool = TelegramClientPool(pool_size=3)
+
+@asynccontextmanager
+async def lifespan(_: FastAPI):
+    setup_logging(Config["log_level"])
+    
+    await client_pool.initialize()  # Start pool
+    background_task = asyncio.create_task(cache_media_files())
+    yield
+    background_task.cancel()
+    await client_pool.shutdown()  # Stop pool
+
+# In download functions:
+async def download_media_file(...):
+    client = await client_pool.get_client()  # Get from pool
+    message = await client.get_messages(channel_id, post_id)
+```
+
+**Benefits**:
+- Distributes load across multiple connections
+- Auth failures affect only subset of requests
+- Better throughput for concurrent operations
+
+#### Improvement 2: Async File I/O
+
+**File**: [`api_server.py`](api_server.py:324)
+
+**Current**: Blocking file writes in async context
+
+**Solution**: Add `aiofiles` to [`requirements.txt`](requirements.txt:1):
+```
+aiofiles==24.1.0
+```
+
+Update download to stream to disk asynchronously:
+```python
+import aiofiles
+
+async def download_media_file(...):
+    # ... setup ...
+    
+    # Stream download to avoid memory issues
+    async for chunk in client.stream_media(file_id):
+        async with aiofiles.open(cache_path, 'ab') as f:
+            await f.write(chunk)
+```
+
+#### Improvement 3: Circuit Breaker Pattern
+
+**New File**: `circuit_breaker.py`
+
+```python
+import asyncio
+import time
+from enum import Enum
+from typing import Callable, Any
+
+class CircuitState(Enum):
+    CLOSED = "closed"  # Normal operation
+    OPEN = "open"      # Failures detected, reject requests
+    HALF_OPEN = "half_open"  # Testing if service recovered
+
+class CircuitBreaker:
+    def __init__(self, failure_threshold: int = 5, timeout: int = 60):
+        self.failure_threshold = failure_threshold
+        self.timeout = timeout
+        self.failure_count = 0
+        self.last_failure_time = 0
+        self.state = CircuitState.CLOSED
+    
+    async def call(self, func: Callable, *args, **kwargs) -> Any:
+        if self.state == CircuitState.OPEN:
+            if time.time() - self.last_failure_time > self.timeout:
+                self.state = CircuitState.HALF_OPEN
+            else:
+                raise Exception("Circuit breaker open - service unavailable")
+        
+        try:
+            result = await func(*args, **kwargs)
+            if self.state == CircuitState.HALF_OPEN:
+                self.state = CircuitState.CLOSED
+                self.failure_count = 0
+            return result
+        except Exception as e:
+            self.failure_count += 1
+            self.last_failure_time = time.time()
+            
+            if self.failure_count >= self.failure_threshold:
+                self.state = CircuitState.OPEN
+                logger.error(f"Circuit breaker opened after {self.failure_count} failures")
+            
+            raise
+
+# Usage in api_server.py
+download_breaker = CircuitBreaker(failure_threshold=5, timeout=60)
+
+async def download_media_file(...):
+    return await download_breaker.call(_download_media_file_impl, ...)
+```
+
+### 4.3 Long-Term Architectural Changes (Recommended)
+
+#### Option 1: Separate Download Service
+
+Move Telegram operations to dedicated service:
+
+```
+┌─────────────────┐      ┌─────────────────┐
+│   FastAPI       │      │  Download       │
+│   Web Server    │─────▶│  Service        │
+│   (api_server)  │ HTTP │  (separate)     │
+└─────────────────┘      └─────────────────┘
+                                  │
+                                  ▼
+                          Pyrogram Clients
+                          (connection pool)
+```
+
+Benefits:
+- Isolates blocking operations
+- Independent scaling
+- Better fault tolerance
+
+
+## 5. Implementation Roadmap
+
+### Phase 1: Critical Fixes (Immediate - Day 1)
+
+**Goal**: Stop active blocking issues
+
+1. ✅ Add `DOWNLOAD_SEMAPHORE` (3 concurrent downloads)
+2. ✅ Add `asyncio.wait_for()` timeouts to all Pyrogram calls
+3. ✅ Implement `safe_get_messages()` and `safe_download_media()` wrappers
+4. ✅ Add download queue for background task coordination
+
+**Testing**:
+- Verify no hangs under 10 concurrent requests
+- Confirm auth errors don't cascade
+- Monitor timeout behavior
+
+### Phase 2: Resilience (Week 1)
+
+**Goal**: Prevent future cascading failures
+
+1. ✅ Implement circuit breaker pattern
+2. ✅ Add detailed metrics/logging for auth failures
+3. ✅ Implement exponential backoff for retries
+4. ✅ Add health check endpoint for Telegram connectivity
+
+**Testing**:
+- Simulate auth failures
+- Verify circuit breaker triggers
+- Test recovery behavior
+
+### Phase 3: Architecture (Week 2-3)
+
+**Goal**: Scale and distribute load
+
+1. ✅ Implement `TelegramClientPool` (3 clients initially)
+2. ✅ Migrate to `aiofiles` for async file I/O
+3. ✅ Add connection health monitoring
+4. ✅ Optimize cache management
+
+**Testing**:
+- Load test with 50+ concurrent requests
+- Measure latency improvements
+- Verify pool balancing

plans/phase1-implementation-summary.md

@@ -0,0 +1,202 @@
+# Phase 1 Critical Fixes - Implementation Summary
+
+## Date: 2026-01-01
+## Status: ✅ COMPLETED
+
+## Overview
+Successfully implemented Phase 1 Critical Fixes to resolve concurrent authentication failures (`KeyError: 0`) in the pyrogram-bridge project. All blocking issues related to unlimited concurrent downloads have been addressed.
+
+## Changes Implemented
+
+### 1. Concurrency Limiter ✅
+**File**: `api_server.py` (lines 70-71)
+
+Added module-level semaphore and download queue:
+```python
+DOWNLOAD_SEMAPHORE = asyncio.Semaphore(3)
+download_queue = asyncio.Queue(maxsize=100)
+```
+
+Wrapped entire `download_media_file()` function body (line 223) with:
+```python
+async with DOWNLOAD_SEMAPHORE:
+    # All download logic here
+```
+
+**Impact**: Limits concurrent Telegram operations to maximum 3 simultaneous downloads, preventing connection conflicts.
+
+### 2. Timeout Protection ✅
+**File**: `api_server.py` (lines 236-240, 268-271, 331-334)
+
+Added timeout handling for critical operations:
+- `get_messages()` call: 30-second timeout with HTTPException(504) on timeout
+- `download_media()` calls: 120-second timeout with HTTPException(504) on timeout
+
+**Impact**: Prevents indefinite hangs and provides clear error responses to clients.
+
+### 3. Safe Wrapper Methods ✅
+**File**: `telegram_client.py` (lines 86-115)
+
+Added two new methods to `TelegramClient` class:
+
+#### `safe_get_messages()`
+- Wraps `get_messages()` with 30-second timeout
+- Implements retry logic (max 2 retries)
+- Detects KeyError auth issues and retries after 5-second delay
+- Logs warnings on retry attempts
+
+#### `safe_download_media()`
+- Wraps `download_media()` with 120-second timeout
+- Implements retry logic (max 2 retries)
+- Detects KeyError auth issues and retries after 5-second delay
+- Logs warnings on retry attempts
+
+**Impact**: Provides resilient communication with Telegram API, automatically recovering from transient auth errors.
+
+### 4. Background Download Queue ✅
+**File**: `api_server.py` (lines 432-478, 501-516)
+
+#### Modified `download_new_files()`
+- Changed from direct download to queue-based system
+- Uses `download_queue.put()` to queue files
+- Handles `QueueFull` exception gracefully
+- Logs number of queued files
+
+#### Added `background_download_worker()`
+- Infinite loop processing downloads from queue
+- Uses semaphore for concurrency control
+- 2-second delay between downloads
+- Comprehensive error handling
+- Calls `task_done()` in finally block
+
+#### Updated `lifespan()`
+- Starts `worker_task` alongside cache management task
+- Properly cancels worker on shutdown
+- Handles `CancelledError` gracefully
+
+**Impact**: Separates user-requested downloads from background cache warming, preventing queue buildup and ensuring responsive API.
+
+### 5. Updated Calls to Use Safe Wrappers ✅
+**File**: `api_server.py` (lines 236-240, 268-271, 331-334)
+
+Replaced direct client calls:
+- `client.client.get_messages()` → `client.safe_get_messages()`
+- `client.client.download_media()` → `client.safe_download_media()`
+
+Added timeout exception handling for both wrapper calls.
+
+**Impact**: All Telegram operations now benefit from retry logic and timeout protection.
+
+## Technical Details
+
+### Semaphore Implementation
+- **Limit**: 3 concurrent downloads
+- **Scope**: Wraps entire `download_media_file()` function
+- **Location**: Module-level variable, initialized at startup
+
+### Download Queue
+- **Capacity**: 100 items
+- **Type**: `asyncio.Queue`
+- **Workers**: 1 dedicated worker task
+- **Processing**: FIFO with 2-second delays
+
+### Timeout Values
+- **get_messages**: 30 seconds
+- **download_media**: 120 seconds (for large files)
+- **Retry delay**: 5 seconds between attempts
+
+### Retry Logic
+- **Max retries**: 2 attempts per operation
+- **Trigger**: KeyError in exception message
+- **Delay**: 5 seconds between retries
+- **Logging**: Warnings on each retry attempt
+
+## Files Modified
+
+1. **telegram_client.py**
+   - Added `safe_get_messages()` method
+   - Added `safe_download_media()` method
+   - Total lines added: ~30
+
+2. **api_server.py**
+   - Added semaphore and queue initialization
+   - Modified `download_media_file()` with semaphore and safe wrappers
+   - Modified `download_new_files()` to use queue
+   - Added `background_download_worker()` function
+   - Updated `lifespan()` to manage worker task
+   - Total lines modified/added: ~100
+
+## Backward Compatibility
+
+✅ All changes maintain backward compatibility:
+- Existing error handling preserved
+- Existing logging maintained
+- API endpoints unchanged
+- Cache structure unchanged
+- Configuration unchanged
+
+## Testing Checklist
+
+### Compilation Tests ✅
+- [x] Python syntax validation passed
+- [x] No import errors
+- [x] Code compiles successfully
+
+### Runtime Tests (To Be Performed)
+- [ ] Verify semaphore limits concurrent downloads to 3
+- [ ] Confirm timeouts prevent indefinite hangs
+- [ ] Test retry logic on simulated auth errors
+- [ ] Verify background queue processes downloads sequentially
+- [ ] Monitor logs for proper retry behavior
+- [ ] Test high-load scenarios (50+ concurrent requests)
+- [ ] Verify cache warming still functions
+- [ ] Check queue doesn't fill up and block
+
+## Expected Outcomes
+
+### Performance Improvements
+- **Eliminated**: 60+ second blocking periods
+- **Reduced**: Concurrent auth conflicts to near-zero
+- **Added**: Graceful timeout handling
+- **Improved**: System resilience with retry logic
+
+### Operational Benefits
+- Clear log messages for debugging
+- Predictable download concurrency
+- Queue-based background processing
+- Automatic recovery from transient errors
+- Better resource utilization
+
+## Next Steps
+
+1. **Deploy**: Apply changes to production environment
+2. **Monitor**: Watch logs for timeout/retry behavior
+3. **Measure**: Track KeyError occurrence rate (should drop significantly)
+4. **Optimize**: Adjust semaphore limit based on actual performance
+5. **Phase 2**: Proceed with connection pool improvements (if needed)
+
+## Notes
+
+- The implementation follows the exact specifications from `performance-blocking-analysis.md`
+- All code changes include comprehensive logging
+- Error handling is defensive and verbose
+- The semaphore approach is proven for this exact issue type
+- Queue-based background downloads prevent user-facing delays
+
+## Validation
+
+✅ Code compiled successfully with no errors
+✅ All specified changes implemented
+✅ Backward compatibility maintained
+✅ Logging is comprehensive
+✅ Error handling is robust
+
+## Conclusion
+
+Phase 1 Critical Fixes have been successfully implemented. The changes address the root cause of blocking issues by:
+1. Limiting concurrent operations to prevent auth conflicts
+2. Adding timeout protection to prevent hangs
+3. Implementing retry logic for transient failures
+4. Separating background downloads from user requests
+
+These changes should immediately resolve the `KeyError: 0` blocking issues and improve overall system stability and performance.

post_parser.py

@@ -9,9 +9,9 @@
 # pylance: disable=reportMissingImports, reportMissingModuleSource
 
 import logging
+import asyncio
 import re
 import os
-import json
 import html
 import inspect
 from datetime import datetime
@@ -21,12 +21,92 @@ from pyrogram.enums import MessageMediaType
 from bleach.css_sanitizer import CSSSanitizer   
 from bleach import clean as HTMLSanitizer
 from config import get_settings
+from file_io import upsert_media_file_ids_bulk_sync, DB_PATH
 from url_signer import generate_media_digest
 
 Config = get_settings()
 
 logger = logging.getLogger(__name__)
 
+# Media types that never yield a renderable image/video in the feed: they are
+# represented by info blocks (or nothing), so posts carrying them get "no_image".
+NO_IMAGE_MEDIA_TYPES = {
+    MessageMediaType.GIVEAWAY,
+    MessageMediaType.GIVEAWAY_WINNERS,
+    MessageMediaType.CHECKLIST,
+    MessageMediaType.CONTACT,
+    MessageMediaType.LOCATION,
+    MessageMediaType.VENUE,
+    MessageMediaType.DICE,
+    MessageMediaType.GAME,
+    MessageMediaType.INVOICE,
+    MessageMediaType.UNSUPPORTED,
+    MessageMediaType.PAID_MEDIA,
+}
+
+
+def _poll_media_object(message):
+    """Return (media_obj, kind) attached to a poll's description_media, or (None, None).
+
+    Kurigram 2.2.23 polls may carry media in description_media / explanation_media
+    (MessageContent objects). Only description_media is considered for rendering:
+    explanation_media is the quiz-answer explanation attachment and does not belong
+    in a channel feed, so it is deliberately NOT rendered (api_server's download
+    lookup still covers it in case a URL for it exists).
+
+    kind is the render hint: 'img' for photo/sticker, 'video' for video/animation.
+
+    All attribute access goes through getattr: older Poll objects and test mocks do
+    not define these fields. A candidate is accepted only when its file_unique_id is
+    a non-empty str — without it nothing can be served through the /media pipeline,
+    and this also keeps loose MagicMock-based poll mocks from producing false
+    positives via auto-created attributes.
+    """
+    poll = getattr(message, 'poll', None)
+    if poll is None:
+        return None, None
+    description_media = getattr(poll, 'description_media', None)
+    if description_media is None:
+        return None, None
+    candidates = (
+        ('photo', 'img'),
+        ('video', 'video'),
+        ('animation', 'video'),
+        ('sticker', 'img'),
+    )
+    for attr, kind in candidates:
+        media_obj = getattr(description_media, attr, None)
+        if media_obj is None:
+            continue
+        file_unique_id = getattr(media_obj, 'file_unique_id', None)
+        if isinstance(file_unique_id, str) and file_unique_id:
+            return media_obj, kind
+    return None, None
+
+
+def _story_media_object(message):
+    """Return (media_obj, kind) for message.story media (video wins over photo), or (None, None).
+
+    kind is the render hint ('video' or 'img'). Rendering, URL generation and file-id
+    collection all go through this helper, so they always agree on WHICH story object
+    is used (e.g. when a story video lacks a usable file_unique_id and the helper
+    falls back to the photo, the tag type follows the fallback too).
+
+    Same defensive rules as _poll_media_object: getattr-only access and a non-empty
+    str file_unique_id requirement.
+    """
+    story = getattr(message, 'story', None)
+    if story is None:
+        return None, None
+    for attr, kind in (('video', 'video'), ('photo', 'img')):
+        media_obj = getattr(story, attr, None)
+        if media_obj is None:
+            continue
+        file_unique_id = getattr(media_obj, 'file_unique_id', None)
+        if isinstance(file_unique_id, str) and file_unique_id:
+            return media_obj, kind
+    return None, None
+
 
 #tests
 #http://127.0.0.1:8000/post/html/DragorWW_space/114 — video
@@ -61,6 +141,12 @@ logger = logging.getLogger(__name__)
 class PostParser:
     def __init__(self, client):
         self.client = client
+        # Media file-id records collected during rendering. _save_media_file_ids
+        # appends (channel, post_id, file_unique_id, ts) tuples here instead of
+        # touching asyncio/DB directly (see 4.2). A fresh PostParser is created per
+        # request and rendering runs in a single thread, so this list is thread-safe.
+        # The caller flushes it once via upsert_media_file_ids_bulk_sync after render.
+        self._pending_media_ids: List[tuple] = []
 
     @staticmethod
     def get_all_possible_flags() -> List[str]:
@@ -88,10 +174,13 @@ class PostParser:
                         post_id: int, 
                         output_type: str = 'json', 
                         debug: bool = False) -> Union[str, Dict[Any, Any], None]:
-        print(f"Getting post {channel}, {post_id}")
         try:
             prepared_channel_id: Union[str, int] = self.channel_name_prepare(channel)
-            message = await self.client.get_messages(prepared_channel_id, post_id)
+            # Bound the single-post fetch so a hung RPC cannot block the request forever.
+            message = await asyncio.wait_for(
+                self.client.get_messages(prepared_channel_id, post_id),
+                timeout=30,
+            )
 
             if Config["debug"]: print(message)
 
@@ -99,7 +188,14 @@ class PostParser:
                 logger.error(f"post_not_found_or_empty: channel {prepared_channel_id}, post_id {post_id}")
                 return None
             
-            processed_message = self.process_message(message)
+            # Single-post outputs need sanitized body/footer (no whole-feed pass exists here).
+            # raw_message is only needed for JSON output and for debug HTML.
+            include_raw = (output_type == 'json') or debug
+            processed_message = self.process_message(message, include_raw=include_raw, sanitize=True)
+
+            # Flush media file-id records collected during rendering with a single bulk upsert.
+            await self._flush_pending_media_ids()
+
             if output_type == 'html':
                 return self._format_html(processed_message, debug)
             elif output_type == 'json':
@@ -109,8 +205,8 @@ class PostParser:
                 return None
             
         except Exception as e:
-            # Log the specific exception type and message
+            # Log the specific exception type and message (use original channel arg to avoid UnboundLocalError)
-            logger.error(f"post_parsing_error: channel {prepared_channel_id}, post_id {post_id}, error_type {type(e).__name__}, error_message {str(e)}")
+            logger.error(f"post_parsing_error: channel {channel}, post_id {post_id}, error_type {type(e).__name__}, error_message {str(e)}")
             # Optional: include traceback for more detail
             # import traceback
             # logger.error(traceback.format_exc())
@@ -187,7 +283,8 @@ class PostParser:
         if message.media:
             if message.media == MessageMediaType.POLL:
                 if message.poll is not None and hasattr(message.poll, 'question'):
-                    poll_question = message.poll.question.strip()
+                    q = message.poll.question
+                    poll_question = (q.text if hasattr(q, 'text') else str(q)).strip()
                     if poll_question:
                         return f"📊 Poll: {poll_question}"
                 return "📊 Poll"
@@ -203,6 +300,31 @@ class PostParser:
             if message.media == MessageMediaType.VOICE:       return "🎤 Voice"
             if message.media == MessageMediaType.VIDEO_NOTE:  return "📱 Video circle"
             if message.media == MessageMediaType.STICKER:     return "🎯 Sticker"
+            # New media types (Kurigram 2.2.23). New Message attributes are accessed
+            # via getattr only — older objects/mocks may not define them.
+            if message.media == MessageMediaType.LIVE_PHOTO:  return "📸 Live Photo"
+            if message.media == MessageMediaType.STORY:       return "📖 Story"
+            if message.media == MessageMediaType.GIVEAWAY:    return "🎁 Giveaway"
+            if message.media == MessageMediaType.GIVEAWAY_WINNERS: return "🏆 Giveaway winners"
+            if message.media == MessageMediaType.PAID_MEDIA:  return "⭐ Paid media"
+            if message.media == MessageMediaType.CHECKLIST:
+                checklist = getattr(message, 'checklist', None)
+                title = getattr(checklist, 'title', None) if checklist else None
+                if isinstance(title, str) and title.strip():
+                    return f"📝 Checklist: {title.strip()[:50]}"
+                return "📝 Checklist"
+            if message.media == MessageMediaType.CONTACT:     return "👤 Contact"
+            if message.media == MessageMediaType.LOCATION:    return "📍 Location"
+            if message.media == MessageMediaType.VENUE:
+                venue = getattr(message, 'venue', None)
+                venue_title = getattr(venue, 'title', None) if venue else None
+                if isinstance(venue_title, str) and venue_title.strip():
+                    return f"📍 {venue_title.strip()}"
+                return "📍 Venue"
+            if message.media == MessageMediaType.DICE:        return "🎲 Dice"
+            if message.media == MessageMediaType.GAME:        return "🎮 Game"
+            if message.media == MessageMediaType.INVOICE:     return "🧾 Invoice"
+            if message.media == MessageMediaType.UNSUPPORTED: return "⚠️ Unsupported content"
 
         # Web pages (if no text or media title)
         if message.web_page:
@@ -223,7 +345,7 @@ class PostParser:
 
         if text_stripped:
             text_was_processed = True
-            text_has_urls = bool(re.search(r'https?://[^\\s<>\"\\\']+', text_stripped))
+            text_has_urls = bool(re.search(r'https?://[^\s<>"\']+', text_stripped))
 
             clean_text = html.unescape(text) # Remove HTML entities 
             clean_text = re.sub(r'<[^>]+?>', '', clean_text) # Remove HTML tags
@@ -307,14 +429,31 @@ class PostParser:
     
     def _extract_reactions(self, message: Message) -> Union[Dict[str, int], None]:
         if reactions := getattr(message, 'reactions', None):
-            return {r.emoji: r.count for r in reactions.reactions}
+            result: Dict[str, int] = {}
+            for r in reactions.reactions:
+                # Resolve emoji key using the same logic as _reactions_views_links()
+                if getattr(r, "is_paid", False):
+                    emoji = "⭐"
+                elif hasattr(r, "emoji") and r.emoji:
+                    emoji = r.emoji
+                elif hasattr(r, "custom_emoji_id"):
+                    emoji = "❓"  # custom emoji — no text representation available
+                else:
+                    emoji = "❓"  # unknown reaction type
+                # Accumulate counts in case multiple reactions resolve to the same key
+                result[emoji] = result.get(emoji, 0) + r.count
+            return result
         return None
 
-    def _extract_flags(self, message: Message) -> List[str]: #Tests: tests/postparser_extract_flags.py
+    def _extract_flags(self, message: Message, html_body: Optional[str] = None) -> List[str]: #Tests: tests/postparser_extract_flags.py
         # Use raw text/caption for some checks before HTML processing
         message_text_str = str(message.text or message.caption or '')
         # Use HTML body for checks involving formatted text or links within HTML
-        message_body_html = self._generate_html_body(message)
+        # If html_body is provided (pre-computed), use it directly to avoid redundant generation
+        if html_body is None:
+            message_body_html = self._generate_html_body(message)
+        else:
+            message_body_html = html_body
         flags = []
 
         # Add "fwd" flag for forwarded messages
@@ -322,7 +461,9 @@ class PostParser:
             flags.append("fwd")
 
         # Add flag "video" if the message media is VIDEO or ANIMATION and the body text is up to 200 characters.
-        if (message.media in [MessageMediaType.VIDEO, MessageMediaType.ANIMATION, MessageMediaType.VIDEO_NOTE] and 
+        # LIVE_PHOTO (Kurigram 2.2.23) renders as a video element, so it counts too.
+        if (message.media in [MessageMediaType.VIDEO, MessageMediaType.ANIMATION, MessageMediaType.VIDEO_NOTE,
+                              MessageMediaType.LIVE_PHOTO] and
             len((message.text or message.caption or '').strip()) <= 200):
             flags.append("video")
 
@@ -331,8 +472,12 @@ class PostParser:
             len((message.text or message.caption or '').strip()) <= 200):
             flags.append("audio")
         
-        # Add flag for posts without images
+        # Add flag for posts without images: no media at all, an info-block-only media
-        if not message.media or message.media == MessageMediaType.POLL:
+        # type (see NO_IMAGE_MEDIA_TYPES), or a poll without renderable description_media.
+        # A poll WITH description_media renders an image/video, so it is NOT flagged.
+        if (not message.media
+                or message.media in NO_IMAGE_MEDIA_TYPES
+                or (message.media == MessageMediaType.POLL and _poll_media_object(message)[0] is None)):
             flags.append("no_image")
 
         # Add flag for sticker messages
@@ -363,10 +508,14 @@ class PostParser:
         # Check if the post's reactions contain more clown emojis (🤡) or poo emojis (💩).
         if reactions := getattr(message, "reactions", None):
             for reaction in getattr(reactions, "reactions", []):
-                if reaction.emoji == "🤡" and reaction.count >= 30:
+                # Skip paid reactions and custom emoji — they have no .emoji attribute
+                if getattr(reaction, "is_paid", False) or not hasattr(reaction, "emoji"):
+                    continue
+                emoji = reaction.emoji  # attribute existence is guaranteed by hasattr guard above
+                if emoji == "🤡" and reaction.count >= 30:
                     flags.append("clownpoo")
                     break
-                if reaction.emoji == "💩" and reaction.count >= 30:
+                if emoji == "💩" and reaction.count >= 30:
                     flags.append("clownpoo")
                     break
 
@@ -447,15 +596,22 @@ class PostParser:
         html_content = []
 
         if debug:
-            html_content.append(f'<div class="title">Title: {data["html"]["title"]}</div><br>')
+            # title comes from _generate_title (user-controlled post text) and never goes
+            # through bleach — escape it before embedding, same as raw_message below.
+            title_escaped = html.escape(str(data["html"]["title"]))
+            html_content.append(f'<div class="title">Title: {title_escaped}</div><br>')
         html_content.append(f'<div class="message-body">{data["html"]["body"]}</div>')
         html_content.append(f'<div class="message-footer">{data["html"]["footer"]}</div>')
         
-        # Add raw JSON debug output if debug is enabled
+        # Add raw JSON debug output if debug is enabled.
+        # raw_message is the full str(message) serialization and may contain user-controlled
+        # text with HTML/JS — escape it before dropping it into the <pre> (bleach can't run
+        # here because <pre> is not in the allowed-tags whitelist and would be stripped).
         if debug:
+            raw_escaped = html.escape(str(data.get("raw_message", "")))
             html_content.append(f'<pre class="debug-json" style="background: #f5f5f5;'
                                 f'padding: 10px; margin-top: 20px; overflow-x: auto; font-size: 10px;'
-                                f'white-space: pre-wrap;">{data["raw_message"]}</pre>')
+                                f'white-space: pre-wrap;">{raw_escaped}</pre>')
         html_data = '\n'.join(html_content)
         return html_data
 
@@ -471,7 +627,39 @@ class PostParser:
         
         return return_html
 
-    def process_message(self, message: Message) -> Dict[Any, Any]:
+    def process_message(self, message: Message, include_raw: bool = False, sanitize: bool = True) -> Dict[Any, Any]:
+        """Build the processed representation of a message.
+
+        Args:
+            message: the Pyrogram message.
+            include_raw: when True, compute the (expensive) full ``str(message)``
+                serialization into ``result['raw_message']``. Only JSON output and
+                debug HTML need it; feed generation must pass False to avoid
+                serializing every post.
+            sanitize: when True, run the html body and footer through a single
+                bleach pass each. Single-post HTML and JSON need this (there is no
+                whole-feed pass on those paths). Feed generation passes False and
+                relies on the final sanitize in rss_generator (per-post for RSS,
+                whole-feed for HTML), so no
+                fragment is sanitized more than once.
+        """
+        # Compute html body once — avoids triple _generate_html_body calls.
+        # The internal per-fragment sanitize passes were removed (4.4); sanitize
+        # exactly once per output boundary here when requested.
+        html_body = self._generate_html_body(message)
+        # NOTE (stage-4 4.4 consequence): flags are now extracted from the PRE-sanitize
+        # body (the per-fragment sanitize that used to run inside _generate_html_body was
+        # removed). Legitimate links are unaffected — bleach keeps whitelisted
+        # <a href="http(s)://…"> — so link/foreign_channel/mention flags are identical for
+        # normal content. They can differ ONLY for URL-like text that bleach would strip
+        # (e.g. a URL inside a disallowed attribute); flags are non-security (used for
+        # exclude_flags filtering / display), so this edge divergence is accepted rather
+        # than re-adding a per-message sanitize pass that 4.4 deliberately eliminated.
+        flags = self._extract_flags(message, html_body=html_body)
+        footer = self.generate_html_footer(message, flags_list=flags)
+        if sanitize:
+            html_body = self._sanitize_html(html_body)
+            footer = self._sanitize_html(footer)
         result = {
             'channel': self.get_channel_username(message),
             'message_id': message.id,
@@ -480,17 +668,20 @@ class PostParser:
             'text': message.text or message.caption or '',
             'html': {
                 'title': self._generate_title(message),
-                'body': self._generate_html_body(message),
+                'body': html_body,
-                'footer': self.generate_html_footer(message)
+                'footer': footer
             },
-            'flags': self._extract_flags(message),
+            'flags': flags,
             'author': self._get_author_info(message),
             'views': message.views,
             'reactions': self._extract_reactions(message),
             'media_group_id': message.media_group_id,
             'service': getattr(message, "service", None),
-            'raw_message': str(message)
+            'show_caption_above_media': getattr(message, 'show_caption_above_media', False),
         }
+        if include_raw:
+            # Full serialization of the message — expensive; only for JSON/debug.
+            result['raw_message'] = str(message)
         
         # Add webpage data if available
         if webpage := getattr(message, "web_page", None):
@@ -509,8 +700,10 @@ class PostParser:
     
 
     def _sanitize_html(self, html_raw: str) -> str:
+        import time as _time
+        sanitize_start = _time.monotonic()
         allowed_tags = ['p', 'a', 'b', 'i', 'strong',
-                        'em', 'ul', 'ol', 'li', 'br',
+                        'em', 's', 'del', 'ul', 'ol', 'li', 'br',
                         'div', 'span', 'img', 'video', 'audio',
                         'source']
         allowed_attributes = {
@@ -535,6 +728,9 @@ class PostParser:
                 css_sanitizer=css_sanitizer,
                 strip=True,
             )
+            elapsed = _time.monotonic() - sanitize_start
+            if elapsed > 0.05:
+                logger.warning(f"diag_sanitize_slow: bleach.clean() took {elapsed:.3f}s, input_len={len(html_raw)}")
             return sanitized_html
         except Exception as e:
             logger.error(f"html_sanitization_error: error {str(e)}")
@@ -603,25 +799,40 @@ class PostParser:
         media_html = self._generate_html_media(message)
         
         content_body.append(f'<div class="post">')
-        if forward_html: 
+        if forward_html:
             content_body.append(forward_html) # Forward info
             content_body.append("<br>")
-        if reply_html: 
+        if reply_html:
             content_body.append(reply_html) # Reply info
             content_body.append("<br>")
-        if text_html: 
+        
-            content_body.append(text_html) # Post text
+        show_caption_above = getattr(message, 'show_caption_above_media', False)
-            content_body.append("<br>")
+
+        if show_caption_above:
+            if text_html:
+                content_body.append(text_html)
+                content_body.append("<br>")
+            if media_html:
+                content_body.append(media_html)
+                content_body.append("<br>")
+        else:
+            if media_html:
+                content_body.append(media_html)
+                content_body.append("<br>")
+            if text_html:
+                content_body.append(text_html)
+                content_body.append("<br>")
 
-        content_body.append(media_html) # Media
-        content_body.append("<br>")
 
         if poll_html: content_body.append(poll_html) # Poll
+        if special_html := self._format_special_media(message): content_body.append(special_html) # Special media info blocks
         if message.forward_origin: content_body.append(f"--- Forwarded post end ---") # Forward info end
         content_body.append(f'</div><br>')
 
+        # NOTE: sanitize is NOT applied here. Sanitization happens exactly once per
+        # output boundary (process_message for single-post/JSON; in rss_generator the
+        # per-post pass for RSS and the whole-feed pass for HTML). See the map (4.4).
         html_body = '\n'.join(content_body)
-        html_body = self._sanitize_html(html_body)
         return html_body
 
     def _generate_html_media(self, message: Message) -> str:
@@ -630,7 +841,27 @@ class PostParser:
 
         content_media = []
         base_url = Config['pyrogram_bridge_url']
-        if message.media and message.media != "MessageMediaType.POLL":
+        # Poll media (Kurigram 2.2.23): a poll may carry description_media which IS
+        # renderable through the regular /media pipeline.
+        poll_media_obj, poll_media_kind = _poll_media_object(message)
+        if message.media == MessageMediaType.PAID_MEDIA:
+            # Paid media cannot be downloaded (it is paid content) — render an info
+            # block instead of a media element. Attributes via getattr only.
+            paid_media = getattr(message, 'paid_media', None)
+            stars_amount = getattr(paid_media, 'stars_amount', 0) if paid_media else 0
+            paid_items = getattr(paid_media, 'media', None) if paid_media else None
+            items_count = len(paid_items) if isinstance(paid_items, (list, tuple)) else 0
+            content_media.append(f'<div class="message-media">')
+            content_media.append(f'<div class="paid-media">⭐ Paid media ({stars_amount} stars, '
+                                f'{items_count} item(s)) — available in Telegram</div>')
+            content_media.append('</div>')
+        # Info-block-only media types (NO_IMAGE_MEDIA_TYPES) are rendered by
+        # _format_special_media — do not open an empty message-media container for
+        # them. PAID_MEDIA (also in that set) is handled by the branch above; POLL is
+        # not in the set and is let in only when it carries renderable poll media.
+        elif (message.media
+                and message.media not in NO_IMAGE_MEDIA_TYPES
+                and (message.media != MessageMediaType.POLL or poll_media_obj is not None)):
             content_media.append(f'<div class="message-media">')
 
             file_unique_id = self._get_file_unique_id(message)
@@ -638,61 +869,82 @@ class PostParser:
                 logger.debug(f"File unique id not found for message {message.id}")
             elif file_unique_id:
                 channel_username = self.get_channel_username(message)
-                file = f"{channel_username}/{message.id}/{file_unique_id}"
+                # Guard: channel_username may be None for private chats without a username
-                digest = generate_media_digest(file)
+                if not channel_username:
-                url = f"{base_url}/media/{file}/{digest}"
+                    logger.warning(f"Could not generate media URL for message {message.id}: channel username is missing.")
+                    content_media.append('</div>')
+                else:
+                    file = f"{channel_username}/{message.id}/{file_unique_id}"
+                    digest = generate_media_digest(file)
+                    url = f"{base_url}/media/{file}/{digest}"
 
-                logger.debug(f"Collected media file: {channel_username}/{message.id}/{file_unique_id}")
+                    logger.debug(f"Collected media file: {channel_username}/{message.id}/{file_unique_id}")
-                
+
-                # Check if document is a PDF file
+                    # Check if document is a PDF file
-                if (message.media == MessageMediaType.DOCUMENT and
+                    if (message.media == MessageMediaType.DOCUMENT and
-                    message.document is not None and hasattr(message.document, 'mime_type') and
+                        message.document is not None and hasattr(message.document, 'mime_type') and
-                    message.document.mime_type == 'application/pdf'):
+                        message.document.mime_type == 'application/pdf'):
-                    # Only attempt to create link if channel_username is available
+                        # Only attempt to create link if channel_username is available
-                    if channel_username:
+                        if channel_username.startswith('-100'):
-                        if channel_username.startswith('-100'): 
                             tg_link = f"https://t.me/c/{channel_username[4:]}/{message.id}"
-                        else:  
+                        else:
                             tg_link = f"https://t.me/{channel_username}/{message.id}"
                         content_media.append(f'<div class="document-pdf" style="padding: 10px;">')
                         content_media.append(f'<a href="{tg_link}" target="_blank">[PDF-файл]</a></div>')
-                    else:
+                    elif message.media in [MessageMediaType.PHOTO, MessageMediaType.DOCUMENT]:
-                        # Handle case where channel_username is None (e.g., log or add placeholder)
+                        content_media.append(f'<img src="{url}" style="max-width:100%; width:auto; height:auto;'
-                        logger.warning(f"Could not generate PDF link for {message.id}: ch username is missing.")
+                                            f'max-height:400px; object-fit:contain;">')
-                        # Add placeholder without link
+                    elif message.media == MessageMediaType.VIDEO:
-                        content_media.append(f'<div class="document-pdf" style="padding: 10px;">[PDF-файл]</div>') 
+                        content_media.append(f'<video controls src="{url}" style="max-width:100%; width:auto;'
-                elif message.media in [MessageMediaType.PHOTO, MessageMediaType.DOCUMENT]:
+                                            f'height:auto; max-height:400px;"></video>')
-                    content_media.append(f'<img src="{url}" style="max-width:100%; width:auto; height:auto;'
+                    elif message.media == MessageMediaType.ANIMATION:
-                                        f'max-height:400px; object-fit:contain;">')
+                        content_media.append(f'<video controls src="{url}" style="max-width:100%; width:auto;'
-                elif message.media == MessageMediaType.VIDEO:
+                                            f'height:auto; max-height:400px;"></video>')
-                    content_media.append(f'<video controls src="{url}" style="max-width:100%; width:auto;'
+                    elif message.media == MessageMediaType.VIDEO_NOTE:
-                                        f'height:auto; max-height:400px;"></video>')
+                        content_media.append(f'<video controls src="{url}" style="max-width:100%; width:auto;'
-                elif message.media == MessageMediaType.ANIMATION:
+                                            f'height:auto; max-height:400px;"></video>')
-                    content_media.append(f'<video controls src="{url}" style="max-width:100%; width:auto;'
+                    elif message.media == MessageMediaType.AUDIO:
-                                        f'height:auto; max-height:400px;"></video>')
+                        mime_type = getattr(message.audio, 'mime_type', 'audio/mpeg')
-                elif message.media == MessageMediaType.VIDEO_NOTE:
+                        content_media.append(f'<audio controls style="width:100%; max-width:400px;">'
-                    content_media.append(f'<video controls src="{url}" style="max-width:100%; width:auto;'
+                                            f'<source src="{url}" type="{mime_type}"></audio>')
-                                        f'height:auto; max-height:400px;"></video>')
+                        content_media.append('<br>')
-                elif message.media == MessageMediaType.AUDIO:
+                    elif message.media == MessageMediaType.VOICE:
-                    mime_type = getattr(message.audio, 'mime_type', 'audio/mpeg')
+                        mime_type = getattr(message.voice, 'mime_type', 'audio/ogg')
-                    content_media.append(f'<audio controls style="width:100%; max-width:400px;">'
+                        content_media.append(f'<audio controls style="width:100%; max-width:400px;">'
-                                        f'<source src="{url}" type="{mime_type}"></audio>')
+                                            f'<source src="{url}" type="{mime_type}"></audio>')
-                    content_media.append('<br>')
+                        content_media.append('<br>')
-                elif message.media == MessageMediaType.VOICE:
+                    elif message.media == MessageMediaType.STICKER:
-                    mime_type = getattr(message.voice, 'mime_type', 'audio/ogg')
+                        emoji = getattr(message.sticker, 'emoji', '')
-                    content_media.append(f'<audio controls style="width:100%; max-width:400px;">'
+                        if getattr(message.sticker, 'is_video', False):
-                                        f'<source src="{url}" type="{mime_type}"></audio>')
+                            content_media.append(f'<video controls autoplay loop muted src="{url}"'
-                    content_media.append('<br>')
+                                                f'style="max-width:100%; width:auto; height:auto; max-height:200px;'
-                elif message.media == MessageMediaType.STICKER:
+                                                f'object-fit:contain;"></video>')
-                    emoji = getattr(message.sticker, 'emoji', '')
+                        else:
-                    if getattr(message.sticker, 'is_video', False):
+                            content_media.append(f'<img src="{url}" alt="Sticker {emoji}" style="max-width:100%;'
+                                                f'width:auto; height:auto; max-height:200px; object-fit:contain;">')
+                    elif message.media == MessageMediaType.LIVE_PHOTO:
+                        # Live photo is effectively a short video clip.
                         content_media.append(f'<video controls autoplay loop muted src="{url}"'
-                                            f'style="max-width:100%; width:auto; height:auto; max-height:200px;'
+                                            f'style="max-width:100%; width:auto; height:auto; max-height:400px;'
                                             f'object-fit:contain;"></video>')
-                    else:
+                    elif message.media == MessageMediaType.STORY:
-                        content_media.append(f'<img src="{url}" alt="Sticker {emoji}" style="max-width:100%;'
+                        # Choose the tag from the SAME helper that produced the URL's
-                                            f'width:auto; height:auto; max-height:200px; object-fit:contain;">')
+                        # file_unique_id, so the tag type always matches the URL object.
-                content_media.append('</div>')
+                        story_media_obj, story_media_kind = _story_media_object(message)
+                        if story_media_obj is not None and story_media_kind == 'video':
+                            content_media.append(f'<video controls src="{url}" style="max-width:100%; width:auto;'
+                                                f'height:auto; max-height:400px;"></video>')
+                        elif story_media_obj is not None:
+                            content_media.append(f'<img src="{url}" style="max-width:100%; width:auto; height:auto;'
+                                                f'max-height:400px; object-fit:contain;">')
+                    elif message.media == MessageMediaType.POLL and poll_media_obj is not None:
+                        if poll_media_kind == 'video':
+                            content_media.append(f'<video controls src="{url}" style="max-width:100%; width:auto;'
+                                                f'height:auto; max-height:400px;"></video>')
+                        else:
+                            content_media.append(f'<img src="{url}" style="max-width:100%; width:auto; height:auto;'
+                                                f'max-height:400px; object-fit:contain;">')
+                    content_media.append('</div>')
         
         if webpage := getattr(message, "web_page", None): # Web page preview
             if webpage_html := self._format_webpage(webpage, message):
@@ -701,8 +953,9 @@ class PostParser:
                     content_media.append(webpage_html)
                     content_media.append('</div>')
 
+        # Not sanitized here — this fragment is embedded in the body and sanitized
+        # once at the output boundary (see the 4.4 sanitize coverage map).
         html_media = '\n'.join(content_media)
-        html_media = self._sanitize_html(html_media)
         return html_media
 
 
@@ -752,13 +1005,15 @@ class PostParser:
             if photo := getattr(webpage, "photo", None):
                 if file_unique_id := getattr(photo, "file_unique_id", None):
                     channel_username = self.get_channel_username(message)
-                    file = f"{channel_username}/{message.id}/{file_unique_id}"
+                    # Guard: skip photo if channel_username is unavailable to avoid broken URLs
-                    digest = generate_media_digest(file)
+                    if channel_username:
-                    url = f"{base_url}/media/{file}/{digest}"
+                        file = f"{channel_username}/{message.id}/{file_unique_id}"
-                    html_parts.append(f'<div class="webpage-photo" style="margin-top:10px;">')
+                        digest = generate_media_digest(file)
-                    html_parts.append(f'<a href="{webpage.url}" target="_blank">')
+                        url = f"{base_url}/media/{file}/{digest}"
-                    html_parts.append(f'<img src="{url}" style="max-width:100%; width:auto;'
+                        html_parts.append(f'<div class="webpage-photo" style="margin-top:10px;">')
-                                        f'height:auto; max-height:200px; object-fit:contain;"></a></div>')
+                        html_parts.append(f'<a href="{webpage.url}" target="_blank">')
+                        html_parts.append(f'<img src="{url}" style="max-width:100%; width:auto;'
+                                            f'height:auto; max-height:200px; object-fit:contain;"></a></div>')
             
             html_parts.append('</div>')
             return '\n'.join(html_parts)
@@ -780,8 +1035,9 @@ class PostParser:
             flags_html = self._format_flags(current_flags)
             content_footer.append('<br>' + flags_html)
             
+        # Not sanitized here — sanitized once at the output boundary (4.4 coverage map):
+        # process_message for single-post/JSON; per-post (RSS) / whole-feed (HTML) for feeds.
         html_footer = '\n'.join(content_footer)
-        html_footer = self._sanitize_html(html_footer)
         return html_footer
 
 
@@ -867,8 +1123,10 @@ class PostParser:
                 if links:
                     parts.append('&nbsp;|&nbsp;'.join(links))
 
+            # Raw fragment — embedded in the footer, sanitized once at the output
+            # boundary (see the 4.4 sanitize coverage map). Not sanitized here.
             result_html = '<br>'.join(parts) if parts else None
-            return self._sanitize_html(result_html) if result_html else None
+            return result_html if result_html else None
             
         except Exception as e:
             logger.error(f"reactions_views_links_error: message_id {message.id}, error {str(e)}")
@@ -879,10 +1137,14 @@ class PostParser:
         
         try:
             if poll := getattr(message, "poll", None):
-                poll_text = f"📊 Poll: {poll.question}\n"
+                q = poll.question
+                question_str = q.text if hasattr(q, 'text') else str(q)
+                poll_text = f"📊 Poll: {question_str}\n"
                 if hasattr(poll, "options") and poll.options:
                     for i, option in enumerate(poll.options, 1):
-                        poll_text += f"{i}. {getattr(option, 'text', '')}\n"
+                        opt_text = getattr(option, 'text', '')
+                        opt_str = opt_text.text if hasattr(opt_text, 'text') else str(opt_text)
+                        poll_text += f"{i}. {opt_str}\n"
                 poll_text += "\n→ Vote in Telegram 🔗\n"
                 return f'<div class="message-poll">{poll_text.replace(chr(10), "<br>")}</div>'
             return None
@@ -890,6 +1152,109 @@ class PostParser:
             logger.error(f"poll_parsing_error: {str(e)}")
             return '<div class="message-poll">[Error displaying poll]</div>'
 
+    @staticmethod
+    def _format_osm_link(location) -> Union[str, None]:
+        """Build an OpenStreetMap link for a location object, or None if coords are unusable."""
+        latitude = getattr(location, 'latitude', None) if location else None
+        longitude = getattr(location, 'longitude', None) if location else None
+        if not isinstance(latitude, (int, float)) or not isinstance(longitude, (int, float)):
+            return None
+        osm_url = f"https://www.openstreetmap.org/?mlat={latitude}&mlon={longitude}#map=16/{latitude}/{longitude}"
+        return f'<a href="{osm_url}">{latitude:.5f}, {longitude:.5f}</a>'
+
+    def _format_special_media(self, message: Message) -> Union[str, None]:
+        """Render an info block for media types that carry no downloadable file.
+
+        Covers giveaways, giveaway winners, checklists, contacts, locations, venues,
+        dice, games, invoices and UNSUPPORTED content (Kurigram 2.2.23). Each block is
+        gated on message.media so unrelated messages never render these. All new
+        Message attributes are accessed via getattr only (older objects/mocks do not
+        define them) and ALL user-controlled strings go through html.escape.
+        """
+        try:
+            media = getattr(message, 'media', None)
+            block = None
+
+            if media == MessageMediaType.GIVEAWAY and (giveaway := getattr(message, 'giveaway', None)):
+                quantity = getattr(giveaway, 'quantity', None)
+                months = getattr(giveaway, 'months', None)
+                stars = getattr(giveaway, 'stars', None)
+                until_date = getattr(giveaway, 'until_date', None)
+                description = getattr(giveaway, 'description', None)
+                block = f"🎁 Giveaway: {quantity} prize(s)"
+                if months: block += f" × {months} months Premium"
+                elif stars: block += f" × {stars} Stars"
+                if until_date is not None and hasattr(until_date, 'strftime'):
+                    block += f" — until {until_date.strftime('%d/%m/%Y')}"
+                if isinstance(description, str) and description.strip():
+                    block += f"<br>{html.escape(description.strip())}"
+
+            elif media == MessageMediaType.GIVEAWAY_WINNERS and (winners := getattr(message, 'giveaway_winners', None)):
+                winner_count = getattr(winners, 'winner_count', None)
+                quantity = getattr(winners, 'quantity', None)
+                prize_description = getattr(winners, 'prize_description', None)
+                block = f"🏆 Giveaway winners: {winner_count} of {quantity}"
+                if isinstance(prize_description, str) and prize_description.strip():
+                    block += f"<br>{html.escape(prize_description.strip())}"
+
+            elif media == MessageMediaType.CHECKLIST and (checklist := getattr(message, 'checklist', None)):
+                title = getattr(checklist, 'title', None)
+                title_str = title if isinstance(title, str) else ''
+                lines = [f"📝 {html.escape(title_str)}" if title_str else "📝 Checklist"]
+                for task in (getattr(checklist, 'tasks', None) or []):
+                    completed = bool(getattr(task, 'completed_by', None) or getattr(task, 'completion_date', None))
+                    mark = "☑" if completed else "☐"
+                    task_text = getattr(task, 'text', '')
+                    task_str = task_text if isinstance(task_text, str) else str(task_text)
+                    lines.append(f"{mark} {html.escape(task_str)}")
+                block = '<br>'.join(lines)
+
+            elif media == MessageMediaType.CONTACT and (contact := getattr(message, 'contact', None)):
+                first_name = getattr(contact, 'first_name', None) or ''
+                last_name = getattr(contact, 'last_name', None) or ''
+                phone_number = getattr(contact, 'phone_number', None) or ''
+                full_name = ' '.join(part for part in [str(first_name), str(last_name)] if part)
+                block = f"👤 {html.escape(full_name)}"
+                if phone_number:
+                    block += f" — {html.escape(str(phone_number))}"
+
+            elif media == MessageMediaType.LOCATION and (location := getattr(message, 'location', None)):
+                osm_link = self._format_osm_link(location)
+                block = f"📍 Location: {osm_link}" if osm_link else "📍 Location"
+
+            elif media == MessageMediaType.VENUE and (venue := getattr(message, 'venue', None)):
+                venue_title = getattr(venue, 'title', None) or ''
+                venue_address = getattr(venue, 'address', None) or ''
+                venue_label = ', '.join(part for part in [str(venue_title), str(venue_address)] if part)
+                block = f"📍 {html.escape(venue_label)}" if venue_label else "📍 Venue"
+                if osm_link := self._format_osm_link(getattr(venue, 'location', None)):
+                    block += f" — {osm_link}"
+
+            elif media == MessageMediaType.DICE and (dice := getattr(message, 'dice', None)):
+                dice_emoji = getattr(dice, 'emoji', None) or '🎲'
+                dice_value = getattr(dice, 'value', None)
+                block = f"🎲 {html.escape(str(dice_emoji))}: {dice_value}"
+
+            elif media == MessageMediaType.GAME:
+                game_title = getattr(getattr(message, 'game', None), 'title', None)
+                if isinstance(game_title, str) and game_title.strip():
+                    block = f"🎮 Game: {html.escape(game_title.strip())}"
+                else:
+                    block = "🎮 Game"
+
+            elif media == MessageMediaType.INVOICE:
+                block = "🧾 Invoice"
+
+            elif media == MessageMediaType.UNSUPPORTED:
+                block = "⚠️ This post contains content not supported by the bridge — open it in Telegram."
+
+            if block is None:
+                return None
+            return f'<div class="message-special">{block}</div>'
+        except Exception as e:
+            logger.error(f"special_media_parsing_error: message_id {getattr(message, 'id', 'unknown')}, error {str(e)}")
+            return None
+
     def _get_file_unique_id(self, message: Message) -> Union[str, None]:
         try:
             media_mapping = {
@@ -901,7 +1266,12 @@ class PostParser:
                 MessageMediaType.VIDEO_NOTE:    lambda m: m.video_note.file_unique_id,
                 MessageMediaType.ANIMATION:     lambda m: m.animation.file_unique_id,
                 MessageMediaType.STICKER:       lambda m: m.sticker.file_unique_id,
-                MessageMediaType.WEB_PAGE:      lambda m: m.web_page.photo.file_unique_id if m.web_page and m.web_page.photo else None
+                MessageMediaType.WEB_PAGE:      lambda m: m.web_page.photo.file_unique_id if m.web_page and m.web_page.photo else None,
+                # New media types (Kurigram 2.2.23): getattr-only access, the
+                # attributes do not exist on older Message objects/mocks.
+                MessageMediaType.LIVE_PHOTO:    lambda m: getattr(getattr(m, 'live_photo', None), 'file_unique_id', None),
+                MessageMediaType.STORY:         lambda m: getattr(_story_media_object(m)[0], 'file_unique_id', None),
+                MessageMediaType.POLL:          lambda m: getattr(_poll_media_object(m)[0], 'file_unique_id', None),
             }
             
             if message.media in media_mapping:
@@ -913,15 +1283,32 @@ class PostParser:
             logger.error(f"file_id_extraction_error: media_type {message.media}, error {str(e)}")
             return None
 
-    def _save_media_file_ids(self, message: Message) -> None:
+    async def _flush_pending_media_ids(self) -> None:
-        try:
+        """Persist media file-id records collected during rendering with ONE bulk upsert.
-            file_data = {
-                'channel': '',
-                'post_id': 0,
-                'file_unique_id': '',
-                'added': .0
-            }
 
+        Called by the caller (get_post / rss_generator) after rendering completes.
+        Runs the blocking SQLite write in a thread. No-op when nothing was collected.
+        """
+        entries = self._pending_media_ids
+        if not entries:
+            return
+        try:
+            await asyncio.to_thread(upsert_media_file_ids_bulk_sync, DB_PATH, entries)
+            logger.debug(f"persist_media_file_ids_bulk: upserted {len(entries)} records")
+        except Exception as e:
+            logger.error(f"file_id_bulk_save_error: error bulk-upserting {len(entries)} records, error {str(e)}")
+        finally:
+            # Clear regardless of outcome so a retry does not double-persist a stale batch.
+            self._pending_media_ids = []
+
+    def _save_media_file_ids(self, message: Message) -> None:
+        """Collect a media file-id record for later bulk persistence.
+
+        IMPORTANT (4.2): this runs inside the render thread, so it must NOT touch
+        asyncio or the DB. It only appends to self._pending_media_ids; the caller
+        flushes the batch via _flush_pending_media_ids() after rendering.
+        """
+        try:
             channel_username = self.get_channel_username(message)
             if not channel_username:
                 logger.error(f"channel_username_error: no username found for chat in message {message.id}")
@@ -929,50 +1316,44 @@ class PostParser:
 
             if message.media:
                 # Skip large videos - they shouldn't be cached permanently
-                if message.video and message.video.file_size and message.video.file_size > 100 * 1024 * 1024: return
+                if message.video and message.video.file_size and message.video.file_size > 100 * 1024 * 1024:
-                
+                    return
-                if   message.photo:         file_data['file_unique_id'] = message.photo.file_unique_id
-                elif message.video:         file_data['file_unique_id'] = message.video.file_unique_id
-                elif message.document:      file_data['file_unique_id'] = message.document.file_unique_id
-                elif message.audio:         file_data['file_unique_id'] = message.audio.file_unique_id
-                elif message.voice:         file_data['file_unique_id'] = message.voice.file_unique_id
-                elif message.video_note:    file_data['file_unique_id'] = message.video_note.file_unique_id
-                elif message.animation:     file_data['file_unique_id'] = message.animation.file_unique_id
-                elif message.sticker:       file_data['file_unique_id'] = message.sticker.file_unique_id
-                elif message.web_page and message.web_page.photo: 
-                    file_data['file_unique_id'] = message.web_page.photo.file_unique_id
 
-                if file_data['file_unique_id']:
+                file_unique_id = ''
-                    file_data['channel'] = channel_username
+                new_media_obj = None  # selected object among the Kurigram 2.2.23 media sources
-                    file_data['post_id'] = message.id
+                if   message.photo:      file_unique_id = message.photo.file_unique_id
-                    file_data['added'] = datetime.now().timestamp()
+                elif message.video:      file_unique_id = message.video.file_unique_id
+                elif message.document:   file_unique_id = message.document.file_unique_id
+                elif message.audio:      file_unique_id = message.audio.file_unique_id
+                elif message.voice:      file_unique_id = message.voice.file_unique_id
+                elif message.video_note: file_unique_id = message.video_note.file_unique_id
+                elif message.animation:  file_unique_id = message.animation.file_unique_id
+                elif message.sticker:    file_unique_id = message.sticker.file_unique_id
+                elif message.web_page and message.web_page.photo:
+                    file_unique_id = message.web_page.photo.file_unique_id
+                # New media types (Kurigram 2.2.23): getattr-only, the attributes do
+                # not exist on older Message objects/mocks. paid_media is deliberately
+                # NOT collected — it cannot be downloaded (paid content).
+                elif getattr(message, 'live_photo', None):
+                    new_media_obj = message.live_photo
+                elif (story_media := _story_media_object(message)[0]) is not None:
+                    new_media_obj = story_media
+                elif (poll_media := _poll_media_object(message)[0]) is not None:
+                    new_media_obj = poll_media
 
-                    file_path = os.path.join(os.path.abspath("./data"), 'media_file_ids.json')
+                if new_media_obj is not None:
-                    try:
+                    # The >100MB message.video guard above does not cover these
-                        existing_data = []
+                    # sources (live photo, story video, poll description video) —
-                        if os.path.exists(file_path):
+                    # apply the same "don't cache large videos" rule here.
-                            with open(file_path, 'r', encoding='utf-8') as f:
+                    file_size = getattr(new_media_obj, 'file_size', None)
-                                existing_data = json.load(f)
+                    if isinstance(file_size, int) and file_size > 100 * 1024 * 1024:
-                        
+                        return
-                        # Check if file already exists by all three fields
+                    file_unique_id = getattr(new_media_obj, 'file_unique_id', '') or ''
-                        found = False
-                        for item in existing_data:
-                            if (item.get('channel') == file_data['channel'] and 
-                                    item.get('post_id') == file_data['post_id'] and 
-                                    item.get('file_unique_id') == file_data['file_unique_id']):
-                                item['added'] = datetime.now().timestamp()
-                                found = True
-                                break
-                        
-                        # Add new entry if not found
-                        if not found:
-                            existing_data.append(file_data)
-                        
-                        with open(file_path, 'w', encoding='utf-8') as f:
-                            json.dump(existing_data, f, ensure_ascii=False, indent=2)
 
-                    except Exception as e:
+                if file_unique_id:
-                        logger.error(f"file_id_save_error: error writing to {file_path}, error {str(e)}")
+                    added_ts = datetime.now().timestamp()
+                    # Thread-safe: just append; the caller persists the batch.
+                    self._pending_media_ids.append((channel_username, message.id, file_unique_id, added_ts))
 
         except Exception as e:
             logger.error(f"file_id_collection_error: message_id {message.id}, error {str(e)}")

pyrogram-bridge.code-workspace

@@ -17,7 +17,12 @@
 			"titleBar.activeBackground": "#e5e636",
 			"titleBar.activeForeground": "#15202b",
 			"titleBar.inactiveBackground": "#e5e63699",
-			"titleBar.inactiveForeground": "#15202b99"
+			"titleBar.inactiveForeground": "#15202b99",
+			"statusBar.background": "#e5e636",
+			"statusBar.foreground": "#15202b",
+			"statusBarItem.hoverBackground": "#cecf1a",
+			"statusBarItem.remoteBackground": "#e5e636",
+			"statusBarItem.remoteForeground": "#15202b"
 		},
 		"peacock.color": "#e5e636"
 	}

pytest.ini

@@ -0,0 +1,9 @@
+[pytest]
+testpaths = tests
+addopts = -ra
+# The stage-1 anti-hang tests are async; run them without per-test event loops
+# being set up by hand. (The tests also carry @pytest.mark.asyncio explicitly, so
+# they work in strict mode too — but auto keeps the plugin's requirement obvious.)
+asyncio_mode = auto
+filterwarnings =
+    ignore::DeprecationWarning

requirements.txt

@@ -1,8 +1,11 @@
 fastapi==0.115.8
+# Pinned explicitly (fastapi only requires a range): the stage-3 FileResponse Range
+# behaviour and the multipart-byteranges test are tied to this Starlette version.
+starlette==0.45.3
 uvicorn==0.34.0
 python-multipart==0.0.20
-#Kurigram
+Kurigram==2.2.23
-git+https://github.com/KurimuzonAkuma/pyrogram.git@dev#egg=Kurigram
+#git+https://github.com/KurimuzonAkuma/pyrogram.git@dev#egg=Kurigram
 TgCrypto
 uvloop
 pillow==11.1.0
@@ -11,4 +14,5 @@ python-dateutil==2.9.0.post0
 python-magic==0.4.27
 bleach[css]==6.1.0
 types-bleach
-json-repair
+pytest-asyncio==1.4.0
+httpx==0.28.1

rss_generator.py

@@ -9,9 +9,12 @@
 # pylance: disable=reportMissingImports, reportMissingModuleSource
 # mypy: disable-error-code="import-untyped"
 
+import copy
 import logging
+import asyncio
 import re
 import time
+from html import escape as html_escape
 from datetime import datetime, timezone
 from types import SimpleNamespace
 from typing import Optional
@@ -20,16 +23,27 @@ from pyrogram import errors, Client
 from pyrogram.types import Message
 from post_parser import PostParser
 from config import get_settings
+from tg_throttle import tg_rpc_bounded
+from bleach.css_sanitizer import CSSSanitizer
+from bleach import clean as HTMLSanitizer
 
 Config = get_settings()
 
 logger = logging.getLogger(__name__)
 
-async def _create_time_based_media_groups(messages: list[Message], merge_seconds: int = 5) -> list[Message]:
+def _create_time_based_media_groups(messages: list[Message], merge_seconds: int = 5) -> list[Message]:
     """
     Create media groups based on time difference between messages
+
+    Plain synchronous function (contains no await): runs inside the render thread
+    via _render_pipeline. Must not touch asyncio.
     """
-    messages_sorted = sorted(messages, key=lambda msg: msg.date) # type: ignore
+    # Deep-copy the input list to avoid mutating cached Message objects (the cache may
+    # reuse the same objects across calls with different merge_seconds values)
+    messages = copy.deepcopy(messages)
+    # Compute fallback once so all None-date messages get the same sort key (deterministic order)
+    _sort_fallback = datetime.now(timezone.utc)
+    messages_sorted = sorted(messages, key=lambda msg: msg.date or _sort_fallback) # type: ignore
     cluster: list[Message] = []
     last_msg_date: datetime = datetime.now(timezone.utc)
     current_media_group_id: Optional[str] = None
@@ -38,11 +52,14 @@ async def _create_time_based_media_groups(messages: list[Message], merge_seconds
         
         if not cluster:
             cluster.append(msg)
-            last_msg_date = msg.date # type: ignore
+            # Use current time as fallback when date is None
+            last_msg_date = msg.date or datetime.now(timezone.utc) # type: ignore
             current_media_group_id = getattr(msg, "media_group_id", None)
             continue
         
-        time_diff = (msg.date - last_msg_date).total_seconds() # type: ignore
+        # Use current time as fallback when date is None to avoid TypeError in subtraction
+        msg_date = msg.date or datetime.now(timezone.utc)
+        time_diff = (msg_date - last_msg_date).total_seconds()
         
         msg_media_group_id = getattr(msg, "media_group_id", None)
         
@@ -54,7 +71,8 @@ async def _create_time_based_media_groups(messages: list[Message], merge_seconds
                 for m in cluster:
                     m.media_group_id = current_media_group_id  # type: ignore
             cluster.append(msg)
-            last_msg_date = msg.date # type: ignore
+            # Use current time as fallback when date is None
+            last_msg_date = msg.date or datetime.now(timezone.utc) # type: ignore
         else:
             if len(cluster) >= 2 and not current_media_group_id:
                 dates = [m.date for m in cluster if m.date is not None]
@@ -64,7 +82,8 @@ async def _create_time_based_media_groups(messages: list[Message], merge_seconds
                     for m in cluster:
                         m.media_group_id = new_group_id  # type: ignore
             cluster = [msg]
-            last_msg_date = msg.date # type: ignore
+            # Use current time as fallback when date is None
+            last_msg_date = msg.date or datetime.now(timezone.utc) # type: ignore
             current_media_group_id = msg_media_group_id
     
     if len(cluster) >= 2 and not current_media_group_id:
@@ -77,9 +96,11 @@ async def _create_time_based_media_groups(messages: list[Message], merge_seconds
 
     return messages_sorted
 
-async def _create_messages_groups(messages: list[Message]) -> list[list[Message]]:
+def _create_messages_groups(messages: list[Message]) -> list[list[Message]]:
     """
     Process messages into formatted posts, handling media groups
+
+    Plain synchronous function (contains no await): runs inside the render thread.
     """
     processing_groups: list[list[Message]] = []
     media_groups: dict[str | int, list[Message]] = {}
@@ -98,7 +119,6 @@ async def _create_messages_groups(messages: list[Message]) -> list[list[Message]
                 if 'GROUP_CHAT_CREATED'     in str(message.service): continue
                 if 'CHANNEL_CHAT_CREATED'   in str(message.service): continue
                 if 'DELETE_CHAT_PHOTO'      in str(message.service): continue
-                if 'NEW_CHAT_TITLE'         in str(message.service): continue
 
             if message.media_group_id:
                 if message.media_group_id not in media_groups:
@@ -122,13 +142,12 @@ async def _create_messages_groups(messages: list[Message]) -> list[list[Message]
     
     return processing_groups
 
-async def _trim_messages_groups(messages_groups: list[list[Message]], limit: int):
+def _trim_messages_groups(messages_groups: list[list[Message]], limit: int):
     """
     Trim messages groups to limit
+
+    Plain synchronous function (contains no await): runs inside the render thread.
     """
-    if messages_groups: # Remove the oldest group (the one with the lowest message id based on the first message's date)
-        messages_groups.pop()
-    
     if len(messages_groups) > limit: # Trim groups if they exceed the specified limit
         messages_groups = messages_groups[:limit]
     
@@ -182,12 +201,13 @@ def processed_message_to_tg_message(processed_message: dict) -> Message:
     return tg_message_mock # type: ignore
 
 
-async def _render_messages_groups(messages_groups: list[list[Message]], 
+def _render_messages_groups(messages_groups: list[list[Message]],
-                                    post_parser: PostParser, 
+                                    post_parser: PostParser,
-                                    exclude_flags: str | None = None, 
+                                    exclude_flags: str | None = None,
                                     exclude_text: str | None = None):
     """
     Render message groups into HTML format
+    Plain synchronous function (contains no await): runs inside the render thread.
     Args:
         messages_groups: List of message groups (each group is a list of messages)
         post_parser: PostParser instance
@@ -202,7 +222,9 @@ async def _render_messages_groups(messages_groups: list[list[Message]],
         try:
             if len(group) == 1: # Single message - simple case
                 one_message = group[0]
-                message_data = post_parser.process_message(one_message)
+                # Feed path: raw_message not needed and sanitize deferred to the final
+                # whole-feed pass, so each fragment is sanitized exactly once.
+                message_data = post_parser.process_message(one_message, include_raw=False, sanitize=False)
                 html_parts = [
                     f'<div class="message-body">{message_data["html"]["body"]}</div>',
                     f'<div class="message-footer">{message_data["html"]["footer"]}</div>'
@@ -217,7 +239,7 @@ async def _render_messages_groups(messages_groups: list[list[Message]],
                     'flags': message_data['flags']
                 })
             else: # Multiple messages in group - merge text and html body
-                processed_messages = [post_parser.process_message(msg) for msg in group]
+                processed_messages = [post_parser.process_message(msg, include_raw=False, sanitize=False) for msg in group]
 
                 # Determine main message for header/footer/title
                 main_message = next(  
@@ -294,11 +316,35 @@ async def _render_messages_groups(messages_groups: list[list[Message]],
                 logger.debug(f"excluded_post: message_id {post['message_id']}, pattern {exclude_pattern.pattern}")
         rendered_posts = filtered_posts
     
-    # Sort by date
+    # Sort by date; use 0.0 (epoch) as fallback for posts with None date to avoid TypeError
-    rendered_posts.sort(key=lambda x: x['date'], reverse=True)
+    rendered_posts.sort(key=lambda x: x['date'] if x['date'] is not None else 0.0, reverse=True)
     return rendered_posts
 
-async def generate_channel_rss(channel: str | int, 
+
+def _render_pipeline(messages: list[Message],
+                     post_parser: PostParser,
+                     limit: int,
+                     exclude_flags: str | None,
+                     exclude_text: str | None,
+                     merge_seconds: int,
+                     time_based_merge: bool):
+    """
+    Full synchronous feed render pipeline (grouping + trimming + rendering).
+
+    Runs entirely in a worker thread via a single asyncio.to_thread call. It contains
+    NO await, NO asyncio primitives, NO create_task/get_running_loop — all the CPU-heavy
+    work (deepcopy, grouping, bleach-free rendering) happens here off the event loop.
+    Media file-id records are accumulated on post_parser._pending_media_ids and flushed
+    by the caller after this returns.
+    """
+    if time_based_merge:
+        messages = _create_time_based_media_groups(messages, merge_seconds)
+    message_groups = _create_messages_groups(messages)
+    message_groups = _trim_messages_groups(message_groups, limit)
+    return _render_messages_groups(message_groups, post_parser, exclude_flags, exclude_text)
+
+
+async def generate_channel_rss(channel: str | int,
                                 client: Client, 
                                 limit: int = 20, 
                                 exclude_flags: str | None = None,
@@ -334,9 +380,10 @@ async def generate_channel_rss(channel: str | int,
         channel_info_start_time = time.time()
         try:
             channel  = post_parser.channel_name_prepare(channel)
-            channel_info = await post_parser.client.get_chat(channel)
+            from tg_cache import cached_get_chat
+            channel_info = await cached_get_chat(post_parser.client, channel)
             channel_title = channel_info.title or f"Telegram: {channel}"
-            channel_username = post_parser.get_channel_username(channel_info)
+            channel_username = channel_info.username or (str(channel_info.id) if channel_info.id and str(channel_info.id).startswith('-100') else None)
             if not channel_username:
                 # Use prepared channel (which could be int) for error feed if username fails
                 logger.warning(f"Could not get username for channel {channel}, using identifier for error feed.")
@@ -344,6 +391,9 @@ async def generate_channel_rss(channel: str | int,
         except (errors.UsernameInvalid, errors.UsernameNotOccupied) as e:
             logger.warning(f"Channel not found error for {channel}: {str(e)}")
             return create_error_feed(channel, base_url)
+        except errors.FloodWait:
+            # Let FloodWait bubble up to api_server.py, which returns HTTP 429 with Retry-After header
+            raise
         except Exception as e:
             logger.error(f"Error during get_chat for channel '{channel}' (type: {type(channel)}): {str(e)}", exc_info=True) # Log error specifically for get_chat
             # Re-raise the original exception to be caught by the outer handler if needed,
@@ -363,10 +413,10 @@ async def generate_channel_rss(channel: str | int,
         
         # Collect messages
         messages_start_time = time.time()
-        messages = []
         try:
-            async for message in post_parser.client.get_chat_history(channel, limit=limit*2):
+            from tg_cache import cached_get_chat_history
-                messages.append(message)
+            
+            messages = await cached_get_chat_history(post_parser.client, channel, limit=limit*2)
         except Exception as e:
             logger.error(f"Error during get_chat_history for channel '{channel}' (type: {type(channel)}): {str(e)}", exc_info=True) # Log error specifically for get_chat_history
             raise ValueError(f"Failed to get chat history for {channel}: {str(e)}") from e # Raise a more specific error
@@ -374,19 +424,37 @@ async def generate_channel_rss(channel: str | int,
         messages_elapsed = time.time() - messages_start_time
         logger.debug(f"rss_messages_retrieval_timing: channel {channel}, {len(messages)} messages retrieved in {messages_elapsed:.3f} seconds")
         
-        # Process messages into groups and render them
+        # Process messages into groups and render them.
+        # The whole grouping/trimming/rendering pipeline is CPU-heavy (deepcopy +
+        # per-message rendering) and contains no await, so run it in ONE worker
+        # thread to keep the event loop responsive.
         processing_start_time = time.time()
-        if Config['time_based_merge']:
+        try:
-            messages = await _create_time_based_media_groups(messages, merge_seconds)
+            final_posts = await asyncio.to_thread(
-        message_groups = await _create_messages_groups(messages)
+                _render_pipeline, messages, post_parser, limit,
-        message_groups = await _trim_messages_groups(message_groups, limit)
+                exclude_flags, exclude_text, merge_seconds, Config['time_based_merge'],
-        final_posts = await _render_messages_groups(message_groups, post_parser, exclude_flags, exclude_text)
+            )
-        
+        finally:
+            # Persist media file-ids collected during rendering with a single bulk
+            # upsert — in a finally so a partial render still records what it collected
+            # (the flush is best-effort and swallows its own errors, so it cannot mask a
+            # render exception).
+            await post_parser._flush_pending_media_ids()
+
         processing_elapsed = time.time() - processing_start_time
         logger.debug(f"rss_messages_processing_timing: channel {channel}, {len(final_posts)} posts processed in {processing_elapsed:.3f} seconds")
         
         # Generate feed entries
         feed_gen_start_time = time.time()
+        
+        # Log date range of posts being added to RSS
+        if final_posts:
+            dates = [datetime.fromtimestamp(post['date'], tz=timezone.utc) for post in final_posts if post.get('date')]
+            if dates:
+                oldest_date = min(dates)
+                newest_date = max(dates)
+                logger.info(f"rss_date_range: channel {channel}, oldest_post {oldest_date.isoformat()}, newest_post {newest_date.isoformat()}, total_posts {len(final_posts)}")
+        
         for post in final_posts:
             fe = fg.add_entry()
             fe.title(post['title'])
@@ -395,10 +463,48 @@ async def generate_channel_rss(channel: str | int,
             fe.link(href=post_link)
             
             fe.description(post['text'].replace('\n', ' '))
-            fe.content(content=post['html'], type='CDATA')
+            # Sanitize heavy HTML in thread to avoid blocking the loop
+            try:
+                def _sanitize_sync(html_raw: str) -> str:
+                    css_sanitizer = CSSSanitizer(
+                        allowed_css_properties=["max-width", "max-height", "object-fit", "width", "height"]
+                    )
+                    return HTMLSanitizer(
+                        html_raw,
+                        tags=['p', 'a', 'b', 'i', 'strong', 'em', 'ul', 'ol', 'li', 'br', 'div', 'span', 'img', 'video', 'audio', 'source'],
+                        attributes={
+                            'a': ['href', 'title', 'target'],
+                            'img': ['src', 'alt', 'style'],
+                            'video': ['controls', 'src', 'style'],
+                            'audio': ['controls', 'style'],
+                            'source': ['src', 'type'],
+                            'div': ['class', 'style'],
+                            'span': ['class']
+                        },
+                        protocols=['http', 'https', 'tg'],
+                        css_sanitizer=css_sanitizer,
+                        strip=True,
+                    )
+                sanitized_html = await asyncio.to_thread(_sanitize_sync, post['html'])
+            except Exception as e:
+                # FAIL CLOSED: since 4.4 removed the per-fragment passes, post['html'] is
+                # raw channel-controlled HTML, and this is its ONLY sanitize. If bleach
+                # itself throws (e.g. RecursionError on pathological nested HTML), do NOT
+                # emit the raw payload (that would be stored XSS) — html.escape it so the
+                # content survives as inert text.
+                logger.error(f"rss_html_sanitization_error: channel {channel}, message_id {post['message_id']}, error {str(e)}")
+                sanitized_html = html_escape(post['html'])
+            fe.content(content=sanitized_html, type='CDATA')
             
-            pub_date = datetime.fromtimestamp(post['date'], tz=timezone.utc)
+            if post['date'] is not None:
-            fe.pubDate(pub_date)
+                pub_date = datetime.fromtimestamp(post['date'], tz=timezone.utc)
+                logger.debug(f"rss_entry_date: channel {channel}, message_id {post['message_id']}, timestamp {post['date']}, pub_date {pub_date.isoformat()}")
+                fe.pubDate(pub_date)
+            else:
+                # Date is None (e.g. service or deleted message) — fall back to current time
+                pub_date = datetime.now(tz=timezone.utc)
+                logger.warning(f"rss_entry_missing_date: channel {channel}, message_id {post['message_id']}, using current time as fallback")
+                fe.pubDate(pub_date)
             fe.guid(post_link, permalink=True)
             
             if post['author'] and post['author'] != main_name:
@@ -407,7 +513,8 @@ async def generate_channel_rss(channel: str | int,
         feed_gen_elapsed = time.time() - feed_gen_start_time
         logger.debug(f"rss_feed_generation_timing: channel {channel}, feed generated in {feed_gen_elapsed:.3f} seconds")
         
-        rss_feed = fg.rss_str(pretty=True)
+        # Serialize RSS in thread (feedgen may be CPU-heavy)
+        rss_feed = await asyncio.to_thread(fg.rss_str, pretty=True)
         if isinstance(rss_feed, bytes):
             return rss_feed.decode('utf-8')
         
@@ -421,17 +528,50 @@ async def generate_channel_rss(channel: str | int,
 
 async def _reply_enrichment(client: Client, messages: list[Message]) -> list[Message]:
     """
-    Enrich messages with reply to messages
+    Enrich messages with reply-to messages.
+
+    Instead of one API call per message, replies are batched: all message IDs
+    that need enrichment are grouped by chat_id and fetched in a single
+    client.get_messages() call per chat_id.
     """
+    # Collect messages that need reply enrichment, grouped by chat_id
+    chat_messages: dict[int, list[Message]] = {}
     for message in messages:
         if message.reply_to_message_id and message.chat:
-            full_message = await client.get_messages(message.chat.id, message.id)
+            chat_id = message.chat.id
-            if isinstance(full_message, list):
+            if chat_id not in chat_messages:
-                if full_message and full_message[0].reply_to_message:
+                chat_messages[chat_id] = []
-                    message.reply_to_message = full_message[0].reply_to_message
+            chat_messages[chat_id].append(message)
-            else:
+
-                if full_message and full_message.reply_to_message:
+    if not chat_messages:
-                    message.reply_to_message = full_message.reply_to_message
+        # No messages with replies — return unchanged
+        return messages
+
+    # Build a lookup: {(chat_id, message_id): full_message} using one batch call per chat
+    reply_lookup: dict[tuple[int, int], Message] = {}
+    for chat_id, chat_msgs in chat_messages.items():
+        ids_to_fetch = [m.id for m in chat_msgs]
+        try:
+            # Throttle under the global RPC gate and bound the call via the shared
+            # tg_rpc_bounded so a hung get_messages cannot pin the gate.
+            async with tg_rpc_bounded(Config["tg_rpc_timeout"]):
+                fetched = await client.get_messages(chat_id, ids_to_fetch)
+            # get_messages may return a single Message or a list
+            if not isinstance(fetched, list):
+                fetched = [fetched]
+            for fm in fetched:
+                if fm and not getattr(fm, 'empty', False):
+                    reply_lookup[(chat_id, fm.id)] = fm
+        except Exception as e:
+            logger.error(f"reply_enrichment_batch_error: chat_id {chat_id}, ids {ids_to_fetch}, error {str(e)}")
+
+    # Apply reply_to_message from lookup
+    for message in messages:
+        if message.reply_to_message_id and message.chat:
+            key = (message.chat.id, message.id)
+            full_message = reply_lookup.get(key)
+            if full_message and full_message.reply_to_message:
+                message.reply_to_message = full_message.reply_to_message
 
     return messages
 
@@ -470,8 +610,9 @@ async def generate_channel_html(channel: str | int,
         try:
             channel = post_parser.channel_name_prepare(channel)
             logger.debug(f"Prepared channel identifier for HTML: {channel} (type: {type(channel)})") # Log prepared channel
-            channel_info = await post_parser.client.get_chat(channel)
+            from tg_cache import cached_get_chat
-            channel_username = post_parser.get_channel_username(channel_info)
+            channel_info = await cached_get_chat(post_parser.client, channel)
+            channel_username = channel_info.username or (str(channel_info.id) if channel_info.id and str(channel_info.id).startswith('-100') else None)
             if not channel_username:
                 logger.warning(f"Could not get username for channel {channel} in HTML generation, returning error feed structure (as string). NOTE: This should ideally return HTML error page.")
                 # For HTML, returning an error feed string might not be ideal. Consider returning a dedicated HTML error page.
@@ -480,6 +621,9 @@ async def generate_channel_html(channel: str | int,
             logger.warning(f"Channel not found error for {channel} in HTML generation: {str(e)}")
             # Consider returning a dedicated HTML error page.
             return create_error_feed(channel, base_url)
+        except errors.FloodWait:
+            # Let FloodWait bubble up to api_server.py, which returns HTTP 429 with Retry-After header
+            raise
         except Exception as e:
             logger.error(f"Error during get_chat for channel '{channel}' (type: {type(channel)}) in HTML generation: {str(e)}", exc_info=True)
             raise ValueError(f"Failed to get chat info for {channel} in HTML generation: {str(e)}") from e
@@ -489,10 +633,10 @@ async def generate_channel_html(channel: str | int,
 
         # Collect messages
         messages_start_time = time.time()
-        messages = []
         try:
-            async for message in post_parser.client.get_chat_history(channel, limit=limit):
+            from tg_cache import cached_get_chat_history
-                messages.append(message)
+            
+            messages = await cached_get_chat_history(post_parser.client, channel, limit=limit)
         except Exception as e:
             logger.error(f"Error during get_chat_history for channel '{channel}' (type: {type(channel)}) in HTML generation: {str(e)}", exc_info=True)
             raise ValueError(f"Failed to get chat history for {channel} in HTML generation: {str(e)}") from e
@@ -506,23 +650,59 @@ async def generate_channel_html(channel: str | int,
         enrichment_elapsed = time.time() - enrichment_start_time
         logger.debug(f"html_reply_enrichment_timing: channel {channel}, replies enriched in {enrichment_elapsed:.3f} seconds")
 
-        # Process messages into groups and render them
+        # Process messages into groups and render them in ONE worker thread
+        # (CPU-heavy, no await) to keep the event loop responsive.
         processing_start_time = time.time()
-        if Config['time_based_merge']:
+        try:
-            messages = await _create_time_based_media_groups(messages, merge_seconds)
+            final_posts = await asyncio.to_thread(
+                _render_pipeline, messages, post_parser, limit,
+                exclude_flags, exclude_text, merge_seconds, Config['time_based_merge'],
+            )
+        finally:
+            # Persist media file-ids collected during rendering with a single bulk
+            # upsert — in a finally so a partial render still records what it collected.
+            await post_parser._flush_pending_media_ids()
 
-        # Process messages into groups and render them
-        message_groups = await _create_messages_groups(messages)
-        message_groups = await _trim_messages_groups(message_groups, limit)
-        final_posts = await _render_messages_groups(message_groups, post_parser, exclude_flags, exclude_text)
-        
         processing_elapsed = time.time() - processing_start_time
         logger.debug(f"html_messages_processing_timing: channel {channel}, {len(final_posts)} posts processed in {processing_elapsed:.3f} seconds")
 
         # Generate HTML content
         html_gen_start_time = time.time()
+        # Concatenate HTML in thread to avoid blocking the loop on large payloads
         html_posts = [post['html'] for post in final_posts]
-        html = '\n<hr class="post-divider">\n'.join(html_posts)
+        def _concat_html(parts: list[str]) -> str:
+            return '\n<hr class="post-divider">\n'.join(parts)
+        html = await asyncio.to_thread(_concat_html, html_posts)
+        
+        # Optionally re-sanitize the final big HTML to ensure safety without blocking the loop
+        try:
+            def _sanitize_sync(html_raw: str) -> str:
+                css_sanitizer = CSSSanitizer(
+                    allowed_css_properties=["max-width", "max-height", "object-fit", "width", "height"]
+                )
+                return HTMLSanitizer(
+                    html_raw,
+                    tags=['p', 'a', 'b', 'i', 'strong', 'em', 'ul', 'ol', 'li', 'br', 'div', 'span', 'img', 'video', 'audio', 'source'],
+                    attributes={
+                        'a': ['href', 'title', 'target'],
+                        'img': ['src', 'alt', 'style'],
+                        'video': ['controls', 'src', 'style'],
+                        'audio': ['controls', 'style'],
+                        'source': ['src', 'type'],
+                        'div': ['class', 'style'],
+                        'span': ['class']
+                    },
+                    protocols=['http', 'https', 'tg'],
+                    css_sanitizer=css_sanitizer,
+                    strip=True,
+                )
+            html = await asyncio.to_thread(_sanitize_sync, html)
+        except Exception as e:
+            # FAIL CLOSED (see the RSS path): `html` is now the raw concatenated feed
+            # (4.4 removed the per-fragment passes). If bleach throws, html.escape the
+            # whole feed rather than returning raw channel HTML/JS to the client.
+            logger.error(f"html_final_sanitization_error: channel {channel}, error {str(e)}")
+            html = html_escape(html)
         
         html_gen_elapsed = time.time() - html_gen_start_time
         logger.debug(f"html_generation_timing: channel {channel}, HTML generated in {html_gen_elapsed:.3f} seconds")

telegram_client.py

@@ -14,6 +14,7 @@ import asyncio
 import sys
 import signal
 import uvloop
+import time
 asyncio.set_event_loop_policy(uvloop.EventLoopPolicy())
 
 from pyrogram import Client
@@ -31,9 +32,27 @@ class TelegramClient:
             api_id=settings["tg_api_id"],
             api_hash=settings["tg_api_hash"],
             workdir=settings["session_path"],
+            proxy=settings["proxy"],  # MTProto proxy config, None if not set
         )
-        self.disconnect_count = 0
+        self.max_disconnects = settings["tg_disconnect_flap_limit"]      # Max disconnects within the flap window before restart
-        self.max_disconnects = 3
+        self._shutting_down = False  # Guard to prevent re-triggering restart during shutdown
+        self._restarting = False            # Guard: an intentional in-process restart is in progress
+        self._disconnect_times = []         # Monotonic timestamps of recent disconnects (sliding window)
+        self.disconnect_window = settings["tg_disconnect_flap_window"]   # Seconds; window for flap detection
+        self._watchdog_task = None
+        self.watchdog_enabled = settings["tg_watchdog_enabled"]
+        self.watchdog_interval = settings["tg_watchdog_interval"]
+        self.watchdog_timeout = settings["tg_watchdog_timeout"]
+        self.watchdog_failures = settings["tg_watchdog_failures"]
+        self.watchdog_restart_timeout = settings["tg_watchdog_restart_timeout"]
+        self.watchdog_heartbeat_every = settings["tg_watchdog_heartbeat_every"]
+        # Watchdog diagnostics counters (cumulative for the process lifetime)
+        self._wd_probe_count = 0          # successful liveness probes
+        self._wd_probe_fail_count = 0     # failed liveness probes
+        self._wd_restart_count = 0        # successful in-process restarts
+        self._wd_fallback_count = 0       # SIGTERM fallbacks after a failed in-process restart
+        self._wd_flap_trigger_count = 0   # times the disconnect-flap threshold was reached
+        self._wd_last_ok_monotonic = None # monotonic timestamp of the last successful probe
         self._setup_connection_handlers()
 
     def _ensure_session_directory(self):
@@ -49,44 +68,257 @@ class TelegramClient:
         self.client.add_handler(DisconnectHandler(self._on_disconnect))
         logger.info("connection_handlers: connection handlers set up")
 
-    def _on_disconnect(self, _client):
+    async def _on_disconnect(self, _client, _session=None):
-        """Handles disconnection from Telegram servers"""
+        """Handles disconnection events from Telegram servers (best-effort safety net).
-        self.disconnect_count += 1
+
-        logger.warning(f"connection_handler: connection lost (#{self.disconnect_count})")
+        NOTE: this handler only fires when Pyrogram calls session.stop(). It does NOT fire in
-        
+        the 'zombie session' case where the recv loop dies silently — that case is handled by
-        if self.disconnect_count >= self.max_disconnects:
+        the active watchdog (_watchdog_loop). Keep both paths.
-            logger.critical(f"connection_handler: reached disconnect limit ({self.max_disconnects})")
+        """
+        # Ignore disconnects caused by intentional shutdown or by our own in-process restart
+        # (client.restart() -> client.stop() -> session.stop() re-dispatches this handler).
+        if self._shutting_down or self._restarting:
+            logger.debug("connection_handler: ignoring disconnect event (shutdown/restart in progress)")
+            return
+
+        now = time.monotonic()
+        # Sliding window: drop disconnects older than the window, then record this one.
+        self._disconnect_times = [t for t in self._disconnect_times if now - t <= self.disconnect_window]
+        self._disconnect_times.append(now)
+        count = len(self._disconnect_times)
+        logger.warning(f"connection_handler: connection lost ({count} within {self.disconnect_window}s window)")
+
+        if count >= self.max_disconnects:
+            self._wd_flap_trigger_count += 1
+            logger.critical(
+                f"connection_handler: disconnect flap limit reached "
+                f"({count} disconnects within {self.disconnect_window}s window, limit={self.max_disconnects}), restarting client"
+            )
+            await self._restart_client(reason=f"disconnect flap ({count} in {self.disconnect_window}s)")
+
+    def _start_watchdog(self):
+        """Starts the active liveness watchdog task (idempotent)."""
+        if not self.watchdog_enabled:
+            logger.info("watchdog: disabled via configuration")
+            return
+        if self._watchdog_task is not None and not self._watchdog_task.done():
+            return
+        self._watchdog_task = asyncio.create_task(self._watchdog_loop())
+
+    async def _watchdog_loop(self):
+        """Active liveness probe for the 'zombie session' state.
+
+        The disconnect-only recovery never triggers when Pyrogram's recv loop dies silently
+        (is_connected stays True, no Disconnect event). This loop periodically issues a real
+        lightweight API call (get_me) bounded by a short timeout; after N consecutive failures
+        it forces an in-process restart. Emits diagnostics so the behaviour can be reconstructed
+        from logs afterwards.
+        """
+        consecutive_failures = 0
+        logger.info(
+            f"watchdog: started (interval={self.watchdog_interval}s, timeout={self.watchdog_timeout}s, "
+            f"failures={self.watchdog_failures}, heartbeat_every={self.watchdog_heartbeat_every} probes)"
+        )
+        try:
+            while True:
+                await asyncio.sleep(self.watchdog_interval)
+                if self._shutting_down:
+                    break
+                if self._restarting:
+                    # A restart is already underway; skip this probe cycle.
+                    logger.debug("watchdog: skip probe (restart in progress)")
+                    continue
+                probe_started = time.monotonic()
+                try:
+                    me = await asyncio.wait_for(self.client.get_me(), timeout=self.watchdog_timeout)
+                    latency_ms = (time.monotonic() - probe_started) * 1000
+                    self._wd_probe_count += 1
+                    self._wd_last_ok_monotonic = time.monotonic()
+                    if consecutive_failures:
+                        logger.warning(
+                            f"watchdog: liveness RESTORED after {consecutive_failures} failed probe(s) "
+                            f"(latency={latency_ms:.0f}ms, me_id={getattr(me, 'id', None)})"
+                        )
+                        consecutive_failures = 0
+                    else:
+                        logger.debug(f"watchdog: probe ok (latency={latency_ms:.0f}ms, probe #{self._wd_probe_count})")
+                    # Periodic proof-of-life heartbeat at INFO, so the ABSENCE of heartbeats in the
+                    # logs is itself a signal that the watchdog died.
+                    if self.watchdog_heartbeat_every > 0 and self._wd_probe_count % self.watchdog_heartbeat_every == 0:
+                        logger.info(
+                            f"watchdog: heartbeat — probes={self._wd_probe_count}, "
+                            f"probe_failures={self._wd_probe_fail_count}, restarts={self._wd_restart_count}, "
+                            f"sigterm_fallbacks={self._wd_fallback_count}, flap_triggers={self._wd_flap_trigger_count}, "
+                            f"last_probe_latency={latency_ms:.0f}ms, is_connected={self.client.is_connected}"
+                        )
+                except asyncio.CancelledError:
+                    raise
+                except Exception as e:
+                    latency_ms = (time.monotonic() - probe_started) * 1000
+                    consecutive_failures += 1
+                    self._wd_probe_fail_count += 1
+                    if self._wd_last_ok_monotonic is not None:
+                        last_ok = f"{time.monotonic() - self._wd_last_ok_monotonic:.0f}s ago"
+                    else:
+                        last_ok = "never"
+                    logger.warning(
+                        f"watchdog: liveness probe FAILED ({consecutive_failures}/{self.watchdog_failures}) "
+                        f"after {latency_ms:.0f}ms: {type(e).__name__}: {e} "
+                        f"(is_connected={self.client.is_connected}, last_ok={last_ok}, "
+                        f"total_failures={self._wd_probe_fail_count})"
+                    )
+                    if consecutive_failures >= self.watchdog_failures:
+                        consecutive_failures = 0
+                        await self._restart_client(reason=f"watchdog: {self.watchdog_failures} consecutive failed probes")
+        except asyncio.CancelledError:
+            logger.info(
+                f"watchdog: stopped (probes={self._wd_probe_count}, probe_failures={self._wd_probe_fail_count}, "
+                f"restarts={self._wd_restart_count}, sigterm_fallbacks={self._wd_fallback_count}, "
+                f"flap_triggers={self._wd_flap_trigger_count})"
+            )
+            raise
+        except Exception as e:
+            logger.critical(f"watchdog: loop crashed unexpectedly ({type(e).__name__}: {e}); liveness protection is now DISABLED until next start")
+
+    async def _restart_client(self, reason: str = "unspecified"):
+        """Recover the client without killing the process when possible.
+
+        Performs an in-process restart (rebuilds session + recv loop), bounded by a timeout so a
+        dead network layer cannot make it hang forever. Falls back to a full process restart
+        (SIGTERM) if the in-process restart fails or times out.
+        """
+        if self._restarting or self._shutting_down:
+            logger.debug(f"recovery: restart requested (reason='{reason}') but already restarting/shutting down — skipped")
+            return
+        self._restarting = True
+        restart_started = time.monotonic()
+        was_connected = self.client.is_connected
+        logger.critical(
+            f"recovery: TRIGGERED — reason='{reason}', is_connected={was_connected}, "
+            f"in-process restarts so far={self._wd_restart_count}, sigterm fallbacks so far={self._wd_fallback_count}"
+        )
+        try:
+            if was_connected:
+                logger.warning("recovery: calling client.restart() (in-process teardown + reconnect)")
+                await asyncio.wait_for(self.client.restart(), timeout=self.watchdog_restart_timeout)
+            else:
+                logger.warning("recovery: client not connected, calling client.start()")
+                await asyncio.wait_for(self.client.start(), timeout=self.watchdog_restart_timeout)
+            duration = time.monotonic() - restart_started
+            self._wd_restart_count += 1
+            # Verification probe to prove the network layer is actually back (diagnostic only).
+            try:
+                me = await asyncio.wait_for(self.client.get_me(), timeout=self.watchdog_timeout)
+                self._wd_last_ok_monotonic = time.monotonic()
+                verify = f", verify_get_me ok (me_id={getattr(me, 'id', None)})"
+            except Exception as ve:
+                verify = f", verify_get_me FAILED ({type(ve).__name__}: {ve})"
+            logger.warning(
+                f"recovery: in-process restart SUCCEEDED in {duration:.1f}s "
+                f"(is_connected={self.client.is_connected}{verify}, total in-process restarts={self._wd_restart_count})"
+            )
+            self._disconnect_times.clear()
+            # Re-arm the watchdog in case it had previously crashed (self-healing).
+            self._start_watchdog()
+        except asyncio.CancelledError:
+            raise
+        except Exception as e:
+            duration = time.monotonic() - restart_started
+            self._wd_fallback_count += 1
+            logger.error(
+                f"recovery: in-process restart FAILED after {duration:.1f}s ({type(e).__name__}: {e}); "
+                f"falling back to process restart (SIGTERM), total fallbacks={self._wd_fallback_count}"
+            )
             self._restart_app()
+        finally:
+            self._restarting = False
 
     async def start(self):
         try:
             if not self.client.is_connected:
                 await self.client.start()
                 logger.info("Telegram client connected successfully")
-                # Reset disconnect counter on successful connection
+                # Reset flap history on a fresh successful connection
-                self.disconnect_count = 0
+                self._disconnect_times.clear()
                 logger.info("connection_handler: connection established")
+            self._start_watchdog()
         except Exception as e:
             logger.error(f"Failed to start Telegram client: {str(e)}")
             raise
 
     def _restart_app(self):
-        """Restarts the application"""
+        """Restarts the application by sending SIGTERM to the process"""
-        logger.warning("connection_handler: restarting application")
+        self._shutting_down = True  # Set flag before sending signal to suppress subsequent disconnect events
+        logger.warning("connection_handler: restarting application by sending SIGTERM")
         try:
-            # Properly shutdown client before restart
+            # Use SIGTERM for proper Docker container restart
-            loop = asyncio.get_event_loop()
+            logger.critical("connection_handler: sending SIGTERM signal")
-            if loop.is_running():
+            os.kill(os.getpid(), signal.SIGTERM)
-                loop.create_task(self.stop())
-            
-            # Use os.execv to restart the process with the same arguments
-            os.execv(sys.executable, [sys.executable] + sys.argv)
         except Exception as e:
             logger.error(f"connection_handler: error during restart: {str(e)}")
-            # Emergency termination in case of restart failure
+            # Emergency termination
-            os.kill(os.getpid(), signal.SIGTERM)
+            os._exit(1)
+
+    def watchdog_last_ok_age(self) -> float | None:
+        """Seconds since the last successful watchdog probe, or None if none succeeded yet.
+
+        Reads only the already-recorded monotonic timestamp set by the watchdog loop; it
+        never issues a Telegram RPC, so it is safe to call from the hot /ping path even
+        while a real RPC is hung.
+        """
+        if self._wd_last_ok_monotonic is None:
+            return None
+        return time.monotonic() - self._wd_last_ok_monotonic
+
+    async def safe_get_messages(self, channel_id, post_id, max_retries=2):
+        """Wrapper with retry logic for auth errors"""
+        for attempt in range(max_retries):
+            try:
+                return await asyncio.wait_for(
+                    self.client.get_messages(channel_id, post_id),
+                    timeout=30.0
+                )
+            except Exception as e:
+                if isinstance(e, KeyError) and attempt < max_retries - 1:
+                    logger.warning(f"Auth error on attempt {attempt + 1}, retrying in 5s...")
+                    await asyncio.sleep(5)
+                    continue
+                raise
+
+    async def safe_download_media(self, file_id, file_name, max_retries=2, timeout: float = 120.0):
+        """Wrapper with retry logic for download errors.
+
+        `timeout` bounds each download attempt; for large videos the caller scales it
+        with file size (see api_server._media_download_timeout)."""
+        for attempt in range(max_retries):
+            try:
+                return await asyncio.wait_for(
+                    self.client.download_media(file_id, file_name=file_name),
+                    timeout=timeout
+                )
+            except Exception as e:
+                if isinstance(e, KeyError) and attempt < max_retries - 1:
+                    logger.warning(f"Download auth error on attempt {attempt + 1}, retrying...")
+                    await asyncio.sleep(5)
+                    continue
+                raise
 
     async def stop(self):
+        # Suppress disconnect handling during intentional shutdown
+        # (client.stop() dispatches the DisconnectHandler).
+        self._shutting_down = True
+        if self._watchdog_task is not None:
+            self._watchdog_task.cancel()
+            try:
+                await self._watchdog_task
+            except asyncio.CancelledError:
+                pass
+            self._watchdog_task = None
         if self.client.is_connected:
-            await self.client.stop()
+            try:
-            logger.info("Telegram client disconnected")
+                await self.client.stop()
+                logger.info("Telegram client disconnected")
+            except Exception as e:
+                # During shutdown the client may be in a half-restarted state
+                # (e.g. a watchdog restart was cancelled mid-flight); ignore stop errors.
+                logger.warning(f"Telegram client stop during shutdown raised {type(e).__name__}: {e}")

tests/conftest.py

@@ -0,0 +1,17 @@
+"""Shared test bootstrap (issue #17).
+
+Ensure the repo root is importable and install the config mock BEFORE any test
+module imports application code. Pytest imports conftest.py before collecting
+test modules, so doing this here (instead of a per-module preamble) makes the
+suite order-independent regardless of collection order or invocation directory.
+"""
+import os
+import sys
+
+ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), '..'))
+if ROOT not in sys.path:
+    sys.path.insert(0, ROOT)
+
+import tests.mock_config as _mock_config
+
+sys.modules['config'] = _mock_config

tests/mock_config.py

@@ -1,6 +1,10 @@
 #!/usr/bin/env python3
 # -*- coding: utf-8 -*-
 
+def setup_logging(level_name: str = "INFO") -> None:
+    """No-op logging setup for tests (mirrors config.setup_logging signature)."""
+    return None
+
 def get_settings():
     """
     Mock config for testing without requiring TG_API_ID and TG_API_HASH
@@ -18,4 +22,20 @@ def get_settings():
         "time_based_merge": False,
         "show_bridge_link": False,
         "show_post_flags": True,
-    } 
+        "proxy": None,
+        "trusted_proxies": [],
+        "tg_rpc_timeout": 60,
+        "tg_watchdog_enabled": True,
+        "tg_watchdog_interval": 60,
+        "tg_watchdog_timeout": 10,
+        "tg_watchdog_failures": 3,
+        "tg_watchdog_restart_timeout": 90,
+        "tg_watchdog_heartbeat_every": 30,
+        "tg_disconnect_flap_limit": 3,
+        "tg_disconnect_flap_window": 120,
+        "tg_ping_unhealthy_after": 250,
+        "media_download_timeout_min": 120,
+        "media_download_timeout_max": 1800,
+        "media_download_min_speed": 256 * 1024,
+        "io_thread_pool_size": 32,
+    }

tests/pytest.ini

@@ -1,4 +0,0 @@
-[pytest]
-addopts = -ra
-filterwarnings =
-    ignore::DeprecationWarning 

tests/test_new_media_types.py

@@ -0,0 +1,580 @@
+# flake8: noqa
+# pylint: disable=protected-access, missing-function-docstring, missing-class-docstring
+# pylint: disable=redefined-outer-name, logging-fstring-interpolation, line-too-long
+# pylance: disable=reportMissingImports, reportMissingModuleSource
+"""
+Kurigram 2.2.23 — new/uncovered media types (LIVE_PHOTO, STORY, poll media,
+GIVEAWAY, GIVEAWAY_WINNERS, PAID_MEDIA, CHECKLIST, CONTACT, LOCATION, VENUE,
+DICE, GAME, INVOICE, UNSUPPORTED).
+
+Covers:
+- titles for every new media type (_media_message_title via _generate_title);
+- HTML rendering: live photo <video>, story <video>/<img>, poll description_media
+  <img>, paid media info block (no download);
+- info blocks for non-downloadable types (_format_special_media) incl. XSS escaping;
+- flags: no_image semantics for the new types and polls with/without media,
+  "video" flag for live photos;
+- api_server.find_file_id_in_message: live_photo, story, poll description_media /
+  explanation_media lookups.
+"""
+from datetime import datetime, timezone
+from types import SimpleNamespace
+
+import pytest
+
+from pyrogram.enums import MessageMediaType
+
+from post_parser import PostParser
+from api_server import find_file_id_in_message
+from url_signer import KeyManager, generate_media_digest
+
+
+class _Str(str):
+    """Minimal stand-in for Pyrogram's Str: .html returns the raw string unchanged."""
+    @property
+    def html(self):
+        return str(self)
+
+
+@pytest.fixture(autouse=True)
+def _pinned_signing_key(monkeypatch):
+    # generate_media_digest reads/creates data/media_digest.key relative to cwd; pin
+    # the in-memory key so digests are deterministic and no file IO happens
+    # regardless of the invocation directory (repo root or tests/).
+    monkeypatch.setattr(KeyManager, "signing_key", "test-signing-key-new-media")
+
+
+@pytest.fixture
+def parser():
+    return PostParser(SimpleNamespace())
+
+
+def make_message(mid=1, media=None, text=None, username="testchan", **extra):
+    m = SimpleNamespace()
+    m.id = mid
+    m.date = datetime(2024, 1, 1, 12, 0, 0, tzinfo=timezone.utc)
+    m.text = _Str(text) if text is not None else None
+    m.caption = None
+    m.media = media
+    m.web_page = None
+    m.poll = None
+    m.service = None
+    m.forward_origin = None
+    m.reply_to_message = None
+    m.sender_chat = None
+    m.from_user = None
+    m.reactions = None
+    m.views = 100
+    m.media_group_id = None
+    m.show_caption_above_media = False
+    m.chat = SimpleNamespace(id=-1001234567890, username=username)
+    for attr in ("photo", "video", "document", "audio", "voice",
+                 "video_note", "animation", "sticker"):
+        setattr(m, attr, None)
+    # New Kurigram 2.2.23 attributes (live_photo, story, giveaway, checklist, ...)
+    # are deliberately NOT set by default: production code must survive their
+    # absence via getattr (that is exactly what old mocks look like).
+    for key, value in extra.items():
+        setattr(m, key, value)
+    return m
+
+
+def media_url(mid, fuid, username="testchan"):
+    file = f"{username}/{mid}/{fuid}"
+    return f"http://test.example.com/media/{file}/{generate_media_digest(file)}"
+
+
+# ---------------------------------------------------------------------------
+# 1. Titles for every new media type
+# ---------------------------------------------------------------------------
+
+def test_title_live_photo(parser):
+    assert parser._generate_title(make_message(media=MessageMediaType.LIVE_PHOTO)) == "📸 Live Photo"
+
+def test_title_story(parser):
+    assert parser._generate_title(make_message(media=MessageMediaType.STORY)) == "📖 Story"
+
+def test_title_giveaway(parser):
+    assert parser._generate_title(make_message(media=MessageMediaType.GIVEAWAY)) == "🎁 Giveaway"
+
+def test_title_giveaway_winners(parser):
+    assert parser._generate_title(make_message(media=MessageMediaType.GIVEAWAY_WINNERS)) == "🏆 Giveaway winners"
+
+def test_title_paid_media(parser):
+    assert parser._generate_title(make_message(media=MessageMediaType.PAID_MEDIA)) == "⭐ Paid media"
+
+def test_title_checklist_with_title(parser):
+    msg = make_message(media=MessageMediaType.CHECKLIST,
+                       checklist=SimpleNamespace(title="Shopping list", tasks=[]))
+    assert parser._generate_title(msg) == "📝 Checklist: Shopping list"
+
+def test_title_checklist_long_title_truncated(parser):
+    long_title = "x" * 80
+    msg = make_message(media=MessageMediaType.CHECKLIST,
+                       checklist=SimpleNamespace(title=long_title, tasks=[]))
+    assert parser._generate_title(msg) == f"📝 Checklist: {'x' * 50}"
+
+def test_title_checklist_without_object(parser):
+    assert parser._generate_title(make_message(media=MessageMediaType.CHECKLIST)) == "📝 Checklist"
+
+def test_title_contact(parser):
+    assert parser._generate_title(make_message(media=MessageMediaType.CONTACT)) == "👤 Contact"
+
+def test_title_location(parser):
+    assert parser._generate_title(make_message(media=MessageMediaType.LOCATION)) == "📍 Location"
+
+def test_title_venue_with_title(parser):
+    msg = make_message(media=MessageMediaType.VENUE,
+                       venue=SimpleNamespace(title="Blue Bottle Cafe", address="1 Main St"))
+    assert parser._generate_title(msg) == "📍 Blue Bottle Cafe"
+
+def test_title_venue_fallback(parser):
+    assert parser._generate_title(make_message(media=MessageMediaType.VENUE)) == "📍 Venue"
+
+def test_title_dice(parser):
+    assert parser._generate_title(make_message(media=MessageMediaType.DICE)) == "🎲 Dice"
+
+def test_title_game(parser):
+    assert parser._generate_title(make_message(media=MessageMediaType.GAME)) == "🎮 Game"
+
+def test_title_invoice(parser):
+    assert parser._generate_title(make_message(media=MessageMediaType.INVOICE)) == "🧾 Invoice"
+
+def test_title_unsupported(parser):
+    assert parser._generate_title(make_message(media=MessageMediaType.UNSUPPORTED)) == "⚠️ Unsupported content"
+
+
+# ---------------------------------------------------------------------------
+# 2. LIVE_PHOTO rendering
+# ---------------------------------------------------------------------------
+
+def test_live_photo_renders_video_with_signed_url(parser):
+    msg = make_message(101, media=MessageMediaType.LIVE_PHOTO,
+                       live_photo=SimpleNamespace(file_unique_id="lp_uid_1", file_id="lp_fid_1"))
+    html_media = parser._generate_html_media(msg)
+    assert "<video" in html_media
+    assert media_url(101, "lp_uid_1") in html_media
+
+
+def test_live_photo_collected_for_media_ids(parser):
+    msg = make_message(102, media=MessageMediaType.LIVE_PHOTO,
+                       live_photo=SimpleNamespace(file_unique_id="lp_uid_2", file_id="lp_fid_2"))
+    parser._generate_html_media(msg)
+    assert [(c, p, f) for c, p, f, _ in parser._pending_media_ids] == [("testchan", 102, "lp_uid_2")]
+
+
+def test_live_photo_gets_video_flag_not_no_image(parser):
+    msg = make_message(103, media=MessageMediaType.LIVE_PHOTO,
+                       live_photo=SimpleNamespace(file_unique_id="lp_uid_3", file_id="lp_fid_3"))
+    flags = parser._extract_flags(msg)
+    assert "video" in flags
+    assert "no_image" not in flags
+
+
+# ---------------------------------------------------------------------------
+# 3. STORY rendering
+# ---------------------------------------------------------------------------
+
+def test_story_with_video_renders_video(parser):
+    msg = make_message(111, media=MessageMediaType.STORY,
+                       story=SimpleNamespace(video=SimpleNamespace(file_unique_id="st_vid"),
+                                             photo=None))
+    html_media = parser._generate_html_media(msg)
+    assert "<video" in html_media
+    assert media_url(111, "st_vid") in html_media
+
+
+def test_story_with_photo_renders_img(parser):
+    msg = make_message(112, media=MessageMediaType.STORY,
+                       story=SimpleNamespace(video=None,
+                                             photo=SimpleNamespace(file_unique_id="st_pic")))
+    html_media = parser._generate_html_media(msg)
+    assert "<img" in html_media
+    assert "<video" not in html_media
+    assert media_url(112, "st_pic") in html_media
+
+
+def test_story_video_wins_over_photo(parser):
+    msg = make_message(113, media=MessageMediaType.STORY,
+                       story=SimpleNamespace(video=SimpleNamespace(file_unique_id="st_vid2"),
+                                             photo=SimpleNamespace(file_unique_id="st_pic2")))
+    html_media = parser._generate_html_media(msg)
+    assert media_url(113, "st_vid2") in html_media
+    assert "st_pic2" not in html_media
+
+
+def test_story_video_without_uid_falls_back_to_photo_as_img(parser):
+    # Review fix 3: the tag choice must follow the SAME object selection as the URL.
+    # A story video with an unusable file_unique_id is skipped by the helper in
+    # favour of the photo — the render must emit <img>, not <video>.
+    msg = make_message(114, media=MessageMediaType.STORY,
+                       story=SimpleNamespace(video=SimpleNamespace(file_unique_id=None),
+                                             photo=SimpleNamespace(file_unique_id="st_pic3")))
+    html_media = parser._generate_html_media(msg)
+    assert "<img" in html_media
+    assert "<video" not in html_media
+    assert media_url(114, "st_pic3") in html_media
+
+
+def test_story_large_video_not_collected_for_media_ids(parser):
+    # Review fix 2: the >100MB "don't cache" rule applies to story videos too.
+    msg = make_message(115, media=MessageMediaType.STORY,
+                       story=SimpleNamespace(video=SimpleNamespace(file_unique_id="st_big",
+                                                                   file_size=200 * 1024 * 1024),
+                                             photo=None))
+    parser._save_media_file_ids(msg)
+    assert parser._pending_media_ids == []
+
+
+def test_story_small_video_collected_for_media_ids(parser):
+    msg = make_message(116, media=MessageMediaType.STORY,
+                       story=SimpleNamespace(video=SimpleNamespace(file_unique_id="st_small",
+                                                                   file_size=5 * 1024 * 1024),
+                                             photo=None))
+    parser._save_media_file_ids(msg)
+    assert [(c, p, f) for c, p, f, _ in parser._pending_media_ids] == [("testchan", 116, "st_small")]
+
+
+# ---------------------------------------------------------------------------
+# 4. POLL with/without description_media
+# ---------------------------------------------------------------------------
+
+def _poll_with_photo(fuid="poll_pic", fid="poll_pic_fid"):
+    return SimpleNamespace(
+        question=_Str("Pick one?"),
+        options=[],
+        description_media=SimpleNamespace(photo=SimpleNamespace(file_unique_id=fuid, file_id=fid)),
+    )
+
+
+def test_poll_with_description_photo_renders_img(parser):
+    msg = make_message(121, media=MessageMediaType.POLL, poll=_poll_with_photo())
+    html_media = parser._generate_html_media(msg)
+    assert "<img" in html_media
+    assert media_url(121, "poll_pic") in html_media
+
+
+def test_poll_with_description_photo_has_no_no_image_flag(parser):
+    msg = make_message(122, media=MessageMediaType.POLL, poll=_poll_with_photo())
+    flags = parser._extract_flags(msg)
+    assert "no_image" not in flags
+    assert "poll" in flags
+
+
+def test_poll_without_media_keeps_no_image_flag(parser):
+    msg = make_message(123, media=MessageMediaType.POLL,
+                       poll=SimpleNamespace(question=_Str("Plain poll?"), options=[]))
+    flags = parser._extract_flags(msg)
+    assert "no_image" in flags
+    assert "poll" in flags
+
+
+def test_poll_media_collected_for_media_ids(parser):
+    msg = make_message(124, media=MessageMediaType.POLL, poll=_poll_with_photo("poll_pic4"))
+    parser._generate_html_media(msg)
+    assert [(c, p, f) for c, p, f, _ in parser._pending_media_ids] == [("testchan", 124, "poll_pic4")]
+
+
+def test_poll_with_video_description_renders_video(parser):
+    poll = SimpleNamespace(
+        question=_Str("Video poll?"),
+        options=[],
+        description_media=SimpleNamespace(video=SimpleNamespace(file_unique_id="poll_vid", file_id="poll_vid_fid")),
+    )
+    msg = make_message(125, media=MessageMediaType.POLL, poll=poll)
+    html_media = parser._generate_html_media(msg)
+    assert "<video" in html_media
+    assert media_url(125, "poll_vid") in html_media
+
+
+# ---------------------------------------------------------------------------
+# 5. api_server.find_file_id_in_message (async)
+# ---------------------------------------------------------------------------
+
+async def test_find_file_id_live_photo():
+    msg = make_message(201, media=MessageMediaType.LIVE_PHOTO,
+                       live_photo=SimpleNamespace(file_unique_id="lp_uid", file_id="lp_fid"))
+    assert await find_file_id_in_message(msg, "lp_uid") == "lp_fid"
+
+
+async def test_find_file_id_story_video():
+    msg = make_message(202, media=MessageMediaType.STORY,
+                       story=SimpleNamespace(photo=None,
+                                             video=SimpleNamespace(file_unique_id="sv_uid", file_id="sv_fid")))
+    assert await find_file_id_in_message(msg, "sv_uid") == "sv_fid"
+
+
+async def test_find_file_id_story_photo():
+    msg = make_message(203, media=MessageMediaType.STORY,
+                       story=SimpleNamespace(photo=SimpleNamespace(file_unique_id="sp_uid", file_id="sp_fid"),
+                                             video=None))
+    assert await find_file_id_in_message(msg, "sp_uid") == "sp_fid"
+
+
+async def test_find_file_id_poll_description_photo():
+    msg = make_message(204, media=MessageMediaType.POLL, poll=_poll_with_photo("pd_uid", "pd_fid"))
+    assert await find_file_id_in_message(msg, "pd_uid") == "pd_fid"
+
+
+async def test_find_file_id_poll_explanation_video():
+    poll = SimpleNamespace(
+        question=_Str("Quiz?"),
+        options=[],
+        explanation_media=SimpleNamespace(video=SimpleNamespace(file_unique_id="ex_uid", file_id="ex_fid")),
+    )
+    msg = make_message(205, media=MessageMediaType.POLL, poll=poll)
+    assert await find_file_id_in_message(msg, "ex_uid") == "ex_fid"
+
+
+async def test_find_file_id_poll_without_media_returns_none():
+    msg = make_message(206, media=MessageMediaType.POLL,
+                       poll=SimpleNamespace(question=_Str("Plain?"), options=[]))
+    assert await find_file_id_in_message(msg, "whatever") is None
+
+
+async def test_find_file_id_poll_none_object_returns_none():
+    msg = make_message(207, media=MessageMediaType.POLL)  # message.poll stays None
+    assert await find_file_id_in_message(msg, "whatever") is None
+
+
+# ---------------------------------------------------------------------------
+# 6. XSS: user-controlled strings in info blocks are escaped
+# ---------------------------------------------------------------------------
+
+XSS = "<script>alert(1)</script>"
+XSS_ESCAPED = "&lt;script&gt;alert(1)&lt;/script&gt;"
+
+
+def test_contact_name_is_escaped(parser):
+    msg = make_message(301, media=MessageMediaType.CONTACT,
+                       contact=SimpleNamespace(first_name=XSS, last_name="Doe",
+                                               phone_number="+1234567890", user_id=None, vcard=None))
+    body = parser._generate_html_body(msg)
+    assert XSS not in body
+    assert XSS_ESCAPED in body
+    assert "+1234567890" in body
+
+
+def test_checklist_task_text_is_escaped(parser):
+    checklist = SimpleNamespace(
+        title="My list",
+        tasks=[
+            SimpleNamespace(id=1, text=XSS, completed_by=None, completion_date=None),
+            SimpleNamespace(id=2, text="done task", completed_by=SimpleNamespace(id=7), completion_date=None),
+        ],
+    )
+    msg = make_message(302, media=MessageMediaType.CHECKLIST, checklist=checklist)
+    body = parser._generate_html_body(msg)
+    assert XSS not in body
+    assert f"☐ {XSS_ESCAPED}" in body
+    assert "☑ done task" in body
+    assert "📝 My list" in body
+
+
+def test_venue_title_is_escaped(parser):
+    # VENUE title feeds venue_label -> html.escape(venue_label)
+    venue = SimpleNamespace(title=XSS, address="1 Main St",
+                            location=SimpleNamespace(latitude=1.5, longitude=2.5))
+    msg = make_message(303, media=MessageMediaType.VENUE, venue=venue)
+    body = parser._generate_html_body(msg)
+    assert XSS not in body
+    assert XSS_ESCAPED in body
+
+
+def test_venue_address_label_is_escaped(parser):
+    # VENUE address also feeds venue_label -> html.escape(venue_label)
+    venue = SimpleNamespace(title="Blue Bottle", address=XSS,
+                            location=SimpleNamespace(latitude=1.5, longitude=2.5))
+    msg = make_message(304, media=MessageMediaType.VENUE, venue=venue)
+    body = parser._generate_html_body(msg)
+    assert XSS not in body
+    assert XSS_ESCAPED in body
+
+
+def test_dice_emoji_is_escaped(parser):
+    # DICE emoji -> html.escape(str(dice_emoji))
+    msg = make_message(305, media=MessageMediaType.DICE,
+                       dice=SimpleNamespace(emoji=XSS, value=6))
+    body = parser._generate_html_body(msg)
+    assert XSS not in body
+    assert XSS_ESCAPED in body
+
+
+def test_game_title_is_escaped(parser):
+    # GAME title -> html.escape(game_title.strip())
+    msg = make_message(306, media=MessageMediaType.GAME,
+                       game=SimpleNamespace(title=XSS))
+    body = parser._generate_html_body(msg)
+    assert XSS not in body
+    assert XSS_ESCAPED in body
+
+
+def test_checklist_title_is_escaped(parser):
+    # CHECKLIST title -> html.escape(title_str)
+    checklist = SimpleNamespace(title=XSS, tasks=[])
+    msg = make_message(307, media=MessageMediaType.CHECKLIST, checklist=checklist)
+    body = parser._generate_html_body(msg)
+    assert XSS not in body
+    assert XSS_ESCAPED in body
+
+
+def test_giveaway_description_is_escaped(parser):
+    # GIVEAWAY description -> html.escape(description.strip())
+    giveaway = SimpleNamespace(quantity=5, months=None, stars=None,
+                               until_date=None, description=XSS)
+    msg = make_message(308, media=MessageMediaType.GIVEAWAY, giveaway=giveaway)
+    body = parser._generate_html_body(msg)
+    assert XSS not in body
+    assert XSS_ESCAPED in body
+
+
+def test_giveaway_winners_prize_description_is_escaped(parser):
+    # GIVEAWAY_WINNERS prize_description -> html.escape(prize_description.strip())
+    winners = SimpleNamespace(winner_count=2, quantity=5, prize_description=XSS,
+                              unclaimed_prize_count=0, winners=[])
+    msg = make_message(309, media=MessageMediaType.GIVEAWAY_WINNERS, giveaway_winners=winners)
+    body = parser._generate_html_body(msg)
+    assert XSS not in body
+    assert XSS_ESCAPED in body
+
+
+# ---------------------------------------------------------------------------
+# 7. Giveaway / giveaway winners info blocks
+# ---------------------------------------------------------------------------
+
+def test_giveaway_block_quantity_and_months(parser):
+    giveaway = SimpleNamespace(quantity=10, months=3, stars=None,
+                               until_date=datetime(2026, 2, 1, tzinfo=timezone.utc),
+                               description="Win big!")
+    msg = make_message(311, media=MessageMediaType.GIVEAWAY, giveaway=giveaway)
+    body = parser._generate_html_body(msg)
+    assert "🎁 Giveaway: 10 prize(s)" in body
+    assert "3 months Premium" in body
+    assert "until 01/02/2026" in body
+    assert "Win big!" in body
+
+
+def test_giveaway_block_with_stars(parser):
+    giveaway = SimpleNamespace(quantity=5, months=None, stars=500,
+                               until_date=None, description=None)
+    msg = make_message(312, media=MessageMediaType.GIVEAWAY, giveaway=giveaway)
+    body = parser._generate_html_body(msg)
+    assert "🎁 Giveaway: 5 prize(s)" in body
+    assert "500 Stars" in body
+
+
+def test_giveaway_winners_block(parser):
+    winners = SimpleNamespace(winner_count=2, quantity=5, prize_description="Cool prize",
+                              unclaimed_prize_count=3, winners=[])
+    msg = make_message(313, media=MessageMediaType.GIVEAWAY_WINNERS, giveaway_winners=winners)
+    body = parser._generate_html_body(msg)
+    assert "🏆 Giveaway winners: 2 of 5" in body
+    assert "Cool prize" in body
+
+
+# ---------------------------------------------------------------------------
+# Other info blocks and paid media
+# ---------------------------------------------------------------------------
+
+def test_paid_media_renders_info_block_without_download(parser):
+    paid = SimpleNamespace(stars_amount=50,
+                           media=[SimpleNamespace(width=100, height=100, duration=None, thumbnail=None),
+                                  SimpleNamespace(width=200, height=200, duration=5, thumbnail=None)])
+    msg = make_message(321, media=MessageMediaType.PAID_MEDIA, paid_media=paid)
+    html_media = parser._generate_html_media(msg)
+    assert "⭐ Paid media (50 stars, 2 item(s)) — available in Telegram" in html_media
+    assert "/media/" not in html_media  # nothing downloadable
+    assert parser._pending_media_ids == []  # nothing collected for the download cache
+    assert "no_image" in parser._extract_flags(msg)
+
+
+def test_location_block_has_osm_link(parser):
+    msg = make_message(322, media=MessageMediaType.LOCATION,
+                       location=SimpleNamespace(latitude=55.75123456, longitude=37.61761111))
+    body = parser._generate_html_body(msg)
+    assert "📍 Location:" in body
+    assert "https://www.openstreetmap.org/?mlat=55.75123456&mlon=37.61761111" in body
+    assert "55.75123, 37.61761" in body
+    assert "no_image" in parser._extract_flags(msg)
+
+
+def test_venue_block_with_osm_link(parser):
+    venue = SimpleNamespace(title="Blue Bottle", address="1 Main St",
+                            location=SimpleNamespace(latitude=1.5, longitude=2.5))
+    msg = make_message(323, media=MessageMediaType.VENUE, venue=venue)
+    body = parser._generate_html_body(msg)
+    assert "📍 Blue Bottle, 1 Main St" in body
+    assert "openstreetmap.org/?mlat=1.5&mlon=2.5" in body
+
+
+# Review fix 1: info-block-only media types must not open an empty
+# <div class="message-media"> container — only the special block is rendered.
+
+def test_venue_has_no_empty_media_div(parser):
+    venue = SimpleNamespace(title="Cafe", address="2 Side St",
+                            location=SimpleNamespace(latitude=1.0, longitude=2.0))
+    msg = make_message(328, media=MessageMediaType.VENUE, venue=venue)
+    body = parser._generate_html_body(msg)
+    assert 'class="message-media"' not in body
+    assert 'class="message-special"' in body
+
+
+def test_contact_has_no_empty_media_div(parser):
+    msg = make_message(329, media=MessageMediaType.CONTACT,
+                       contact=SimpleNamespace(first_name="Jane", last_name="Doe",
+                                               phone_number="+1999", user_id=None, vcard=None))
+    body = parser._generate_html_body(msg)
+    assert 'class="message-media"' not in body
+    assert 'class="message-special"' in body
+
+
+def test_paid_media_block_survives_media_div_gate(parser):
+    # PAID_MEDIA is in NO_IMAGE_MEDIA_TYPES but has its own render branch — the
+    # info block (inside its message-media container) must keep rendering.
+    paid = SimpleNamespace(stars_amount=10, media=[SimpleNamespace(width=1, height=1,
+                                                                   duration=None, thumbnail=None)])
+    msg = make_message(330, media=MessageMediaType.PAID_MEDIA, paid_media=paid)
+    body = parser._generate_html_body(msg)
+    assert 'class="message-media"' in body
+    assert "⭐ Paid media (10 stars, 1 item(s)) — available in Telegram" in body
+
+
+def test_dice_block(parser):
+    msg = make_message(324, media=MessageMediaType.DICE,
+                       dice=SimpleNamespace(emoji="🎯", value=6))
+    body = parser._generate_html_body(msg)
+    assert "🎲 🎯: 6" in body
+
+
+def test_game_block_with_title(parser):
+    msg = make_message(325, media=MessageMediaType.GAME,
+                       game=SimpleNamespace(title="Tetris"))
+    body = parser._generate_html_body(msg)
+    assert "🎮 Game: Tetris" in body
+
+
+def test_invoice_block(parser):
+    msg = make_message(326, media=MessageMediaType.INVOICE)
+    body = parser._generate_html_body(msg)
+    assert "🧾 Invoice" in body
+
+
+def test_unsupported_block(parser):
+    msg = make_message(327, media=MessageMediaType.UNSUPPORTED)
+    body = parser._generate_html_body(msg)
+    assert "⚠️ This post contains content not supported by the bridge" in body
+    assert "no_image" in parser._extract_flags(msg)
+
+
+# ---------------------------------------------------------------------------
+# Regression: messages without any of the new attributes still render
+# ---------------------------------------------------------------------------
+
+def test_plain_text_message_unaffected(parser):
+    msg = make_message(331, text="just a plain post with enough text")
+    body = parser._generate_html_body(msg)
+    assert "just a plain post with enough text" in body
+    assert "message-special" not in body
+    assert "paid-media" not in body
+    flags = parser._extract_flags(msg)
+    assert "no_image" in flags

tests/test_postparser_author_info.py

@@ -7,13 +7,6 @@
 
 import unittest
 from unittest.mock import MagicMock
-import sys
-import os
-# Add project root to sys.path to find post_parser
-sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
-
-# Mock the config module
-sys.modules['config'] = __import__('tests.mock_config', fromlist=['get_settings'])
 
 from pyrogram.types import Message
 from post_parser import PostParser

tests/test_postparser_extract_flags.py

@@ -7,20 +7,28 @@
 
 import unittest
 from unittest.mock import MagicMock, PropertyMock
-import sys
-import os
-
-# Add project root to sys.path to find post_parser
-sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
-
-# Mock the config module before importing PostParser
-sys.modules['config'] = __import__('tests.mock_config', fromlist=['get_settings'])
 
 from pyrogram.types import Message, Chat, Reaction, MessageReactions
 from pyrogram.enums import MessageMediaType
 from post_parser import PostParser # Import after mocking config
 
 
+class StrWithHtml(str):
+    """String subclass that mimics Pyrogram's text/caption objects.
+
+    Pyrogram's text and caption fields are not plain str — they provide an
+    .html property that returns the HTML-escaped version of the text.
+    Using a plain str (or MagicMock) breaks len() / strip() checks in
+    production code.  This class inherits from str so all str operations
+    (len, strip, comparison, boolean coercion) work correctly, while also
+    exposing the .html property expected by the parser.
+    """
+
+    @property
+    def html(self):
+        return self
+
+
 class TestPostParserExtractFlags(unittest.TestCase):
 
     def setUp(self):
@@ -72,18 +80,14 @@ class TestPostParserExtractFlags(unittest.TestCase):
         else:
             message.forward_origin = None
 
-        # Correctly mock text and caption to have an .html attribute and behave like strings
+        # Assign text and caption as StrWithHtml instances so that str operations
+        # (len, strip, bool coercion) behave exactly as they would with real
+        # Pyrogram objects, while the .html property is still available.
         if text:
-            mock_text = MagicMock()
+            message.text = StrWithHtml(text)
-            mock_text.html = text
-            mock_text.__str__.return_value = text # Add __str__ mock
-            message.text = mock_text
             message.caption = None
         elif caption:
-            mock_caption = MagicMock()
+            message.caption = StrWithHtml(caption)
-            mock_caption.html = caption
-            mock_caption.__str__.return_value = caption # Add __str__ mock
-            message.caption = mock_caption
             message.text = None
         else:
             message.text = None
@@ -126,36 +130,8 @@ class TestPostParserExtractFlags(unittest.TestCase):
 
     def test_flag_video_media_video_long_caption(self):
         long_caption = "This is a very long caption that exceeds the two hundred character limit, therefore it should not trigger the video flag even though the media type is video." * 5
-        
+        message = self._create_mock_message(media=MessageMediaType.VIDEO, caption=long_caption)
-        # Create a simple mock message for this specific test
+        # Sanity-check: the caption must actually be longer than the threshold
-        message = MagicMock(spec=Message)
-        message.media = MessageMediaType.VIDEO
-        message.id = 123
-        message.forward_origin = None # Add missing attribute
-        message.web_page = None # Add missing attribute checked in _extract_flags
-        
-        # Create a basic mock Chat
-        mock_chat = MagicMock(spec=Chat)
-        mock_chat.username = "test_channel"
-        mock_chat.id = 1234567890
-        message.chat = mock_chat
-        
-        # Set up text and caption to ensure len() returns the correct length
-        message.text = None
-        
-        # Create a custom class that mimics caption with html attribute
-        class CaptionWithHtml(str):
-            @property
-            def html(self):
-                return self
-                
-        # Use our class instead of a regular string
-        message.caption = CaptionWithHtml(long_caption)
-        
-        # Mock HTML body generation
-        self.parser._generate_html_body = MagicMock(return_value=long_caption)
-        
-        # In _extract_flags, text length is checked and should be > 200
         self.assertGreater(len(message.caption), 200)
         self.assertNotIn("video", self.parser._extract_flags(message))
 

tests/test_postparser_gen_title.py

@@ -7,13 +7,6 @@
 # pylance: disable=reportMissingImports, reportMissingModuleSource
 import unittest
 from unittest.mock import MagicMock, PropertyMock
-import sys
-import os
-# Add project root to sys.path to find post_parser
-sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
-
-# Mock the config module
-sys.modules['config'] = __import__('tests.mock_config', fromlist=['get_settings'])
 
 from pyrogram.types import Message, Document
 from pyrogram.enums import MessageMediaType

tests/test_stage1_hangs.py

@@ -0,0 +1,218 @@
+# flake8: noqa
+# pylint: disable=protected-access, missing-function-docstring, missing-class-docstring
+# pylint: disable=redefined-outer-name, logging-fstring-interpolation, line-too-long
+# pylance: disable=reportMissingImports, reportMissingModuleSource
+"""
+Stage 1 (anti-hang) regression tests:
+- RPC gate timeout: a hung RPC times out and the gate permit is NOT leaked; a
+  subsequent call still succeeds.
+- Background worker: a download that raises Exception / FloodWait does not kill the
+  worker, and task_done stays balanced so queue.join() completes.
+- Gate cancellation during the spacing wait does not lose the permit.
+"""
+import time
+import asyncio
+from types import SimpleNamespace
+
+import pytest
+
+import tg_throttle
+import tg_cache
+from pyrogram import errors
+
+
+# --------------------------------------------------------------------------- #
+# 1.1 — RPC gate timeout releases the permit; second call succeeds.
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_gate_timeout_releases_permit(monkeypatch):
+    # Short timeout so the hung RPC fails fast, and no min-interval spacing to slow the test.
+    monkeypatch.setitem(tg_cache.Config, "tg_rpc_timeout", 0.1)
+    monkeypatch.setattr(tg_throttle, "_MIN_INTERVAL", 0.0)
+    # Bypass the on-disk chat cache so we always hit the live RPC path.
+    monkeypatch.setattr(tg_cache, "_get_chat_from_cache", lambda *a, **k: None)
+    monkeypatch.setattr(tg_cache, "_save_chat_to_cache", lambda *a, **k: None)
+
+    permits_before = tg_throttle._sem._value
+    never = asyncio.Event()  # never set -> RPC hangs forever
+
+    class HungClient:
+        async def get_chat(self, channel_id):
+            await never.wait()
+
+    # First call: the RPC hangs and must time out (not hang the whole app).
+    with pytest.raises(asyncio.TimeoutError):
+        await tg_cache.cached_get_chat(HungClient(), "hung_channel")
+
+    # The permit was released on timeout (gate fully available again) — no leak.
+    assert tg_throttle._sem._value == permits_before
+
+    # A second call still goes through the gate and succeeds.
+    class OkClient:
+        async def get_chat(self, channel_id):
+            return SimpleNamespace(id=42, title="ok", username="okchan")
+
+    res = await tg_cache.cached_get_chat(OkClient(), "ok_channel")
+    assert res.id == 42
+    # Permit released again after the successful call.
+    assert tg_throttle._sem._value == permits_before
+
+
+# --------------------------------------------------------------------------- #
+# 1.4 — Gate cancellation during the spacing wait does not lose the permit.
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_gate_cancel_during_spacing_releases_permit(monkeypatch):
+    # Force a spacing wait long enough to cancel inside it.
+    monkeypatch.setattr(tg_throttle, "_MIN_INTERVAL", 1.0)
+    monkeypatch.setattr(tg_throttle, "_last_start", time.monotonic())
+
+    permits_before = tg_throttle._sem._value
+
+    async def enter_gate():
+        async with tg_throttle.tg_rpc():
+            pass
+
+    task = asyncio.create_task(enter_gate())
+    # Let it acquire the semaphore and start sleeping for the spacing interval.
+    await asyncio.sleep(0.05)
+    task.cancel()
+    with pytest.raises(asyncio.CancelledError):
+        await task
+
+    # Cancelled mid-spacing: the acquired permit must be returned.
+    assert tg_throttle._sem._value == permits_before
+
+
+# --------------------------------------------------------------------------- #
+# 1.3 — Background worker survives download errors; task_done stays balanced.
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_worker_survives_errors_and_balances_task_done(monkeypatch):
+    import api_server
+
+    # Record (and skip) the worker's sleeps so we can assert the FloodWait branch
+    # actually backed off — otherwise deleting `except FloodWait` (dropping it into
+    # the generic handler) leaves this test green.
+    sleeps = []
+    async def _fast_sleep(delay=0, *_a, **_k):
+        sleeps.append(delay)
+    monkeypatch.setattr(api_server.asyncio, "sleep", _fast_sleep)
+
+    # Fresh queue so we don't interfere with (or depend on) module state.
+    q = asyncio.Queue(maxsize=100)
+    monkeypatch.setattr(api_server, "download_queue", q)
+
+    processed = []
+
+    async def fake_download(channel, post_id, file_unique_id):
+        processed.append(file_unique_id)
+        if file_unique_id == "boom":
+            raise RuntimeError("simulated download failure")
+        if file_unique_id == "flood":
+            raise errors.FloodWait(value=1)
+        # "ok" succeeds
+
+    monkeypatch.setattr(api_server, "download_media_file", fake_download)
+
+    worker = asyncio.create_task(api_server.background_download_worker())
+    try:
+        for fid in ("boom", "flood", "ok"):
+            q.put_nowait(("chan", 1, fid))
+
+        # If the worker died on the first error, or task_done() were unbalanced,
+        # join() would never complete and this wait_for would time out.
+        await asyncio.wait_for(q.join(), timeout=5)
+
+        # The worker stayed alive across the Exception and the FloodWait and drained all items.
+        assert processed == ["boom", "flood", "ok"]
+
+        # The FloodWait branch (caught BEFORE the generic Exception) backed off by
+        # min(value + 5, 900) = min(1 + 5, 900) = 6, distinct from the success path's
+        # post-download sleep of 2. Asserting the 6 pins the dedicated branch: if it
+        # were removed (FloodWait falling into `except Exception`), no 6 would appear.
+        assert 6 in sleeps, sleeps  # FloodWait(value=1) -> backoff 6
+        assert 2 in sleeps, sleeps  # "ok" success -> post-download sleep 2
+    finally:
+        worker.cancel()
+        with pytest.raises(asyncio.CancelledError):
+            await worker
+
+
+# --------------------------------------------------------------------------- #
+# 1.4 — _supervised: restart on crash (rate-limited), restart on unexpected
+# return, and clean cancellation propagation on shutdown.
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_supervised_restarts_crashing_task_rate_limited():
+    import api_server
+
+    starts = []
+
+    async def crashing():
+        starts.append(time.monotonic())
+        raise RuntimeError("always dies")
+
+    # Tiny min interval so the test is fast, but non-zero so we can assert spacing.
+    sup = asyncio.create_task(
+        api_server._supervised(crashing, "crasher", min_restart_interval=0.05)
+    )
+    # Let it crash-and-restart a few times.
+    await asyncio.sleep(0.28)
+    sup.cancel()
+    with pytest.raises(asyncio.CancelledError):
+        await sup
+
+    # It restarted several times (did not give up after the first crash)...
+    assert len(starts) >= 3
+    # ...but restarts were spaced at least ~min_restart_interval apart (no spin).
+    gaps = [b - a for a, b in zip(starts, starts[1:])]
+    assert all(g >= 0.045 for g in gaps), gaps
+
+
+@pytest.mark.asyncio
+async def test_supervised_restarts_on_unexpected_return():
+    import api_server
+
+    runs = []
+
+    async def returns_immediately():
+        runs.append(1)
+        # A supervised background loop is not meant to return; _supervised must
+        # log CRITICAL and restart it rather than stop supervising.
+        return
+
+    sup = asyncio.create_task(
+        api_server._supervised(returns_immediately, "returner", min_restart_interval=0.02)
+    )
+    await asyncio.sleep(0.1)
+    sup.cancel()
+    with pytest.raises(asyncio.CancelledError):
+        await sup
+
+    assert len(runs) >= 2  # restarted after the unexpected return
+
+
+@pytest.mark.asyncio
+async def test_supervised_cancellation_propagates_to_child():
+    import api_server
+
+    child_cancelled = asyncio.Event()
+
+    async def long_running():
+        try:
+            await asyncio.Event().wait()  # runs until cancelled
+        except asyncio.CancelledError:
+            child_cancelled.set()
+            raise
+
+    sup = asyncio.create_task(
+        api_server._supervised(long_running, "longrun")
+    )
+    await asyncio.sleep(0.05)  # let the child start
+    sup.cancel()
+    # Cancelling the supervisor must propagate CancelledError out (clean shutdown)...
+    with pytest.raises(asyncio.CancelledError):
+        await sup
+    # ...and the child task must have been cancelled too (no leaked background task).
+    assert child_cancelled.is_set()

tests/test_stage2_static.py

@@ -0,0 +1,398 @@
+# flake8: noqa
+# pylint: disable=protected-access, missing-function-docstring, missing-class-docstring
+# pylint: disable=redefined-outer-name, logging-fstring-interpolation, line-too-long
+# pylance: disable=reportMissingImports, reportMissingModuleSource
+"""
+Stage 2 (static-media stability) regression tests.
+
+Covers the four scenarios from the plan plus the subtle in-flight-dedup lifecycle:
+- _download_atomic: success publishes atomically and leaves no `.part.`; timeout / zero-size
+  / race-loser all remove the partial and never leave a stub at the final name.
+- Concurrent large-video downloads never expose a partial file (atomic rename).
+- _download_deduped: one shared download for concurrent requests; a cancelled waiter
+  (client disconnect) can't hang the others; a failed download propagates to all waiters
+  and frees the key (no stuck registry entry).
+- FloodWait in /media returns a retryable 429 (not a permanent 404) — ordering pinned.
+- A served temp_* file gets its mtime refreshed (sweeper won't delete it under a viewer).
+- The sweeper removes both `.part.` and legacy `.tmp.` stubs (and stale temp_*) but keeps
+  fresh partials and non-temp files.
+"""
+import os
+import time
+import asyncio
+from types import SimpleNamespace
+
+import pytest
+
+import api_server
+from pyrogram import errors
+from pyrogram.enums import MessageMediaType
+
+
+def _part_files(dir_path):
+    return [p for p in os.listdir(dir_path) if ".part." in p or ".tmp." in p]
+
+
+# --------------------------------------------------------------------------- #
+# 2.1 — _download_atomic: publish-on-success, always clean the partial.
+# --------------------------------------------------------------------------- #
+async def test_download_atomic_success_publishes_and_cleans(monkeypatch, tmp_path):
+    final = str(tmp_path / "final")
+
+    async def fake(file_id, file_name, timeout=None, **kw):
+        with open(file_name, "wb") as f:
+            f.write(b"hello")
+        return file_name
+
+    monkeypatch.setattr(api_server.client, "safe_download_media", fake)
+
+    res = await api_server._download_atomic("fid", final, timeout=10)
+    assert res == final
+    with open(final, "rb") as f:
+        assert f.read() == b"hello"
+    assert _part_files(tmp_path) == []  # no leftover partial
+
+
+async def test_download_atomic_timeout_removes_part_no_final(monkeypatch, tmp_path):
+    final = str(tmp_path / "final")
+
+    async def fake(file_id, file_name, timeout=None, **kw):
+        # A partial write, then the download times out mid-flight.
+        with open(file_name, "wb") as f:
+            f.write(b"partial")
+        raise asyncio.TimeoutError()
+
+    monkeypatch.setattr(api_server.client, "safe_download_media", fake)
+
+    with pytest.raises(asyncio.TimeoutError):
+        await api_server._download_atomic("fid", final, timeout=10)
+
+    assert not os.path.exists(final)       # no stub at the final name
+    assert _part_files(tmp_path) == []     # partial cleaned by the finally
+
+
+async def test_download_atomic_zero_size_raises_and_cleans(monkeypatch, tmp_path):
+    final = str(tmp_path / "final")
+
+    async def fake(file_id, file_name, timeout=None, **kw):
+        open(file_name, "wb").close()  # zero-size result
+        return file_name
+
+    monkeypatch.setattr(api_server.client, "safe_download_media", fake)
+
+    with pytest.raises(api_server.ZeroSizeFileError):
+        await api_server._download_atomic("fid", final, timeout=10)
+
+    assert not os.path.exists(final)
+    assert _part_files(tmp_path) == []
+
+
+async def test_download_atomic_race_loser_keeps_existing_final(monkeypatch, tmp_path):
+    final = str(tmp_path / "final")
+    with open(final, "wb") as f:
+        f.write(b"WINNER")  # a concurrent request already published the final file
+
+    async def fake(file_id, file_name, timeout=None, **kw):
+        with open(file_name, "wb") as f:
+            f.write(b"loser")
+        return file_name
+
+    monkeypatch.setattr(api_server.client, "safe_download_media", fake)
+
+    res = await api_server._download_atomic("fid", final, timeout=10)
+    assert res == final
+    with open(final, "rb") as f:
+        assert f.read() == b"WINNER"  # not clobbered by the race loser
+    assert _part_files(tmp_path) == []  # loser's partial removed
+
+
+# --------------------------------------------------------------------------- #
+# 2.1 — Concurrent large-video downloads never expose a partial file.
+# --------------------------------------------------------------------------- #
+async def test_concurrent_large_video_no_partial_served(monkeypatch, tmp_path):
+    monkeypatch.chdir(tmp_path)  # download_media_file writes under ./data/cache
+
+    msg = SimpleNamespace(
+        media=MessageMediaType.VIDEO,
+        video=SimpleNamespace(file_size=200 * 1024 * 1024),  # > 100 MB -> large-video path
+        empty=False,
+    )
+
+    async def fake_get(channel_id, post_id):
+        return msg
+
+    monkeypatch.setattr(api_server.client, "safe_get_messages", fake_get)
+
+    async def fake_find(message, fid):
+        return "fileid"
+
+    monkeypatch.setattr(api_server, "find_file_id_in_message", fake_find)
+
+    started = []
+
+    async def fake_dl(file_id, file_name, timeout=None, **kw):
+        started.append(file_name)
+        await asyncio.sleep(0.1)  # slow download -> real overlap between the two requests
+        with open(file_name, "wb") as f:
+            f.write(b"VIDEODATA")
+        return file_name
+
+    monkeypatch.setattr(api_server.client, "safe_download_media", fake_dl)
+
+    t1 = asyncio.create_task(api_server.download_media_file("chan", 7, "vidfid"))
+    t2 = asyncio.create_task(api_server.download_media_file("chan", 7, "vidfid"))
+    (p1, _), (p2, _) = await asyncio.gather(t1, t2)
+
+    assert p1 == p2
+    assert os.path.basename(p1) == "temp_vidfid"  # large videos keep the TTL-cleaned name
+    with open(p1, "rb") as f:
+        assert f.read() == b"VIDEODATA"  # a complete file, never a partial
+
+    post_dir = os.path.dirname(p1)
+    assert _part_files(post_dir) == []  # both partials resolved, none left on disk
+    # Both requests really downloaded to distinct `.part.` paths (a genuine race that the
+    # atomic rename resolved), so neither ever saw the other's partial.
+    assert len(started) == 2 and started[0] != started[1]
+
+
+# --------------------------------------------------------------------------- #
+# 2.2 — In-flight dedup: one shared download for concurrent requests.
+# --------------------------------------------------------------------------- #
+async def test_dedup_single_download_for_concurrent_requests(monkeypatch):
+    api_server._inflight.clear()
+    calls = []
+
+    async def fake_dl(channel, post_id, fid):
+        calls.append(fid)
+        await asyncio.sleep(0.05)
+        return (f"/final/{fid}", False)
+
+    monkeypatch.setattr(api_server, "download_media_file", fake_dl)
+
+    tasks = [asyncio.create_task(api_server._download_deduped("c", 1, "f")) for _ in range(5)]
+    results = await asyncio.gather(*tasks)
+
+    assert all(r == ("/final/f", False) for r in results)
+    assert calls == ["f"]              # exactly one real download, shared by all 5
+    assert not api_server._inflight    # key popped after completion
+
+
+# --------------------------------------------------------------------------- #
+# 2.2 — A cancelled waiter (client disconnect) must not hang the others, and the
+# detached download runs to completion regardless.
+# --------------------------------------------------------------------------- #
+async def test_dedup_waiter_cancel_does_not_hang_others(monkeypatch):
+    api_server._inflight.clear()
+    gate = asyncio.Event()
+    calls = []
+
+    async def fake_dl(channel, post_id, fid):
+        calls.append(fid)
+        await gate.wait()  # hold the download open until the test releases it
+        return (f"/path/{fid}", False)
+
+    monkeypatch.setattr(api_server, "download_media_file", fake_dl)
+
+    t1 = asyncio.create_task(api_server._download_deduped("c", 1, "fidX"))
+    await asyncio.sleep(0.05)  # let t1 register the future + start the detached task
+    t2 = asyncio.create_task(api_server._download_deduped("c", 1, "fidX"))
+    await asyncio.sleep(0.05)
+
+    # First client disconnects -> its request coroutine is cancelled.
+    t1.cancel()
+    with pytest.raises(asyncio.CancelledError):
+        await t1
+
+    # The detached download is unaffected; complete it and the surviving waiter resolves.
+    gate.set()
+    res = await asyncio.wait_for(t2, timeout=2)
+    assert res == ("/path/fidX", False)
+    assert calls == ["fidX"]           # download ran exactly once (shared, not restarted)
+    assert not api_server._inflight     # key popped
+
+
+# --------------------------------------------------------------------------- #
+# 2.2 — A failed download propagates to all waiters and frees the key (not stuck).
+# --------------------------------------------------------------------------- #
+async def test_dedup_exception_propagates_and_frees_key(monkeypatch):
+    api_server._inflight.clear()
+    state = {"mode": "boom"}
+
+    async def fake_dl(channel, post_id, fid):
+        if state["mode"] == "boom":
+            raise RuntimeError("kaboom")
+        return ("/ok", False)
+
+    monkeypatch.setattr(api_server, "download_media_file", fake_dl)
+
+    t1 = asyncio.create_task(api_server._download_deduped("c", 1, "f"))
+    t2 = asyncio.create_task(api_server._download_deduped("c", 1, "f"))
+    results = await asyncio.gather(t1, t2, return_exceptions=True)
+    assert all(isinstance(r, RuntimeError) for r in results), results
+    assert not api_server._inflight  # failed future did NOT leave the key stuck
+
+    # A subsequent request with a working download succeeds (proves the key was freed).
+    state["mode"] = "ok"
+    res = await api_server._download_deduped("c", 1, "f")
+    assert res == ("/ok", False)
+
+
+# --------------------------------------------------------------------------- #
+# 2.3 — FloodWait in /media -> 429 with Retry-After (ordering pinned: not a 404).
+# --------------------------------------------------------------------------- #
+async def test_media_floodwait_returns_429_not_404(monkeypatch, tmp_path):
+    api_server._inflight.clear()
+    monkeypatch.chdir(tmp_path)  # so the pre-semaphore cache check misses -> download path
+    monkeypatch.setattr(api_server, "verify_media_digest", lambda url, digest: True)
+
+    async def fake_dl(channel, post_id, fid):
+        raise errors.FloodWait(value=42)
+
+    monkeypatch.setattr(api_server, "download_media_file", fake_dl)
+
+    req = SimpleNamespace(headers={})
+    resp = await api_server.get_media("chan", 5, "floodfid", request=req, digest="x")
+
+    assert resp.status_code == 429
+    retry_after = resp.headers.get("retry-after")
+    assert retry_after is not None
+    assert 43 <= int(retry_after) <= 72  # 42 + random(1..30)
+
+
+# --------------------------------------------------------------------------- #
+# 2.5 — HTTP_DOWNLOAD_SEMAPHORE balance: released exactly once on success/error,
+# never released when the acquire itself timed out (503). It is a plain
+# asyncio.Semaphore, so an over-release would SILENTLY inflate the permit count
+# and disable the limiter — assert the count, mirroring the stage-1 gate test.
+# --------------------------------------------------------------------------- #
+async def test_media_semaphore_released_on_download_error(monkeypatch, tmp_path):
+    api_server._inflight.clear()
+    monkeypatch.chdir(tmp_path)  # pre-semaphore cache miss -> the acquire path runs
+    monkeypatch.setattr(api_server, "verify_media_digest", lambda url, digest: True)
+
+    permits_before = api_server.HTTP_DOWNLOAD_SEMAPHORE._value
+
+    async def boom(channel_id, post_id, fid):
+        raise RuntimeError("download blew up")
+
+    monkeypatch.setattr(api_server, "_download_deduped", boom)
+
+    req = SimpleNamespace(headers={})
+    # get_media swallows the error into a 4xx/5xx response; the point is the permit.
+    try:
+        await api_server.get_media("chan", 5, "errfid", request=req, digest="x")
+    except Exception:
+        pass
+
+    # Permit released exactly once (back to baseline) — not leaked, not double-released.
+    assert api_server.HTTP_DOWNLOAD_SEMAPHORE._value == permits_before
+
+
+async def test_media_semaphore_not_released_on_acquire_timeout(monkeypatch, tmp_path):
+    api_server._inflight.clear()
+    monkeypatch.chdir(tmp_path)
+    monkeypatch.setattr(api_server, "verify_media_digest", lambda url, digest: True)
+
+    permits_before = api_server.HTTP_DOWNLOAD_SEMAPHORE._value
+
+    # Make the bounded acquire time out: the first wait_for (wrapping the acquire)
+    # raises TimeoutError; we close the passed coroutine to avoid a "never awaited"
+    # warning. The 503 short-circuits before _download_deduped, so only this one
+    # wait_for is hit.
+    real_wait_for = asyncio.wait_for
+
+    async def fake_wait_for(coro, timeout=None):
+        if hasattr(coro, "close"):
+            coro.close()
+        raise asyncio.TimeoutError()
+
+    monkeypatch.setattr(api_server.asyncio, "wait_for", fake_wait_for)
+
+    req = SimpleNamespace(headers={})
+    resp = await api_server.get_media("chan", 5, "busyfid", request=req, digest="x")
+
+    assert resp.status_code == 503
+    assert resp.headers.get("retry-after") == "30"
+    # A timed-out acquire never held a permit, so nothing must have been released:
+    # the count is UNCHANGED (an erroneous release would inflate it above baseline).
+    assert api_server.HTTP_DOWNLOAD_SEMAPHORE._value == permits_before
+
+    monkeypatch.setattr(api_server.asyncio, "wait_for", real_wait_for)
+
+
+async def test_download_atomic_reraises_floodwait_and_cleans_part(monkeypatch, tmp_path):
+    # Stage-2 review: prove a FloodWait raised by the downloader propagates OUT of
+    # _download_atomic (its finally must NOT swallow it) AND the partial is cleaned.
+    # This is the path the wholesale-mocked 429 test above skips.
+    final = str(tmp_path / "temp_vid")
+    seen_part = {}
+
+    async def fake(file_id, part_path, timeout=120):
+        # The partial path exists mid-download, then FloodWait strikes.
+        with open(part_path, "wb") as fh:
+            fh.write(b"partial")
+        seen_part["path"] = part_path
+        raise errors.FloodWait(value=7)
+
+    monkeypatch.setattr(api_server.client, "safe_download_media", fake)
+
+    with pytest.raises(errors.FloodWait):
+        await api_server._download_atomic("fid", final, timeout=10)
+
+    # The finally cleaned the partial, and no final file was published.
+    assert not os.path.exists(seen_part["path"])
+    assert not os.path.exists(final)
+
+
+# --------------------------------------------------------------------------- #
+# 2.4 — Serving a temp_* file refreshes its mtime; non-temp files are untouched.
+# --------------------------------------------------------------------------- #
+async def test_temp_file_mtime_touched_on_serve(tmp_path):
+    temp_file = tmp_path / "temp_abc"
+    temp_file.write_bytes(b"videodata")
+    stale = time.time() - 10000
+    os.utime(temp_file, (stale, stale))
+
+    req = SimpleNamespace(headers={})
+    resp = await api_server.prepare_file_response(str(temp_file), request=req)
+    assert resp is not None
+    assert os.path.getmtime(temp_file) > time.time() - 100  # refreshed
+
+
+async def test_non_temp_file_mtime_not_touched(tmp_path):
+    normal = tmp_path / "regularfile"
+    normal.write_bytes(b"data")
+    stale = time.time() - 10000
+    os.utime(normal, (stale, stale))
+
+    req = SimpleNamespace(headers={})
+    await api_server.prepare_file_response(str(normal), request=req)
+    assert os.path.getmtime(normal) < time.time() - 5000  # left alone
+
+
+# --------------------------------------------------------------------------- #
+# 2.1/2.4 — Sweeper cleans both `.part.` and legacy `.tmp.` stubs and stale temp_*,
+# but keeps fresh partials and ordinary cached files.
+# --------------------------------------------------------------------------- #
+def test_sweeper_cleans_part_and_tmp_not_fresh(tmp_path):
+    cache = tmp_path / "cache"
+    post = cache / "chan" / "10"
+    post.mkdir(parents=True)
+
+    stale = time.time() - 4000  # older than the 1h (3600s) threshold
+    hexid = "0" * 32
+
+    old_part = post / f"file.part.{hexid}"; old_part.write_bytes(b"x"); os.utime(old_part, (stale, stale))
+    old_tmp = post / f"file.tmp.{hexid}"; old_tmp.write_bytes(b"x"); os.utime(old_tmp, (stale, stale))
+    old_tempvid = post / "temp_bigvid"; old_tempvid.write_bytes(b"x"); os.utime(old_tempvid, (stale, stale))
+    fresh_part = post / f"file2.part.{hexid}"; fresh_part.write_bytes(b"x")  # fresh mtime
+    normal = post / "realfile"; normal.write_bytes(b"x")  # not a temp file at all
+
+    _, removed = api_server.remove_old_cached_files_sync([], str(cache))
+
+    assert not old_part.exists()      # new-suffix stub swept
+    assert not old_tmp.exists()       # legacy-suffix stub swept (transition period)
+    assert not old_tempvid.exists()   # stale large-video temp swept
+    assert fresh_part.exists()        # too new to sweep
+    assert normal.exists()            # ordinary cached file untouched
+    assert removed >= 3

tests/test_stage3_fileresponse.py

@@ -0,0 +1,378 @@
+# flake8: noqa
+# pylint: disable=protected-access, missing-function-docstring, missing-class-docstring
+# pylint: disable=redefined-outer-name, logging-fstring-interpolation, line-too-long
+# pylance: disable=reportMissingImports, reportMissingModuleSource
+"""
+Stage 3 (FileResponse + HTTP-layer cleanup) regression tests.
+
+Covers:
+- The HTTP Range matrix asserted against the FINAL (Starlette FileResponse) behavior,
+  each case documented, including the RFC-7233-permitted deltas vs the old hand-rolled
+  streaming (garbage header -> 400 not 416; multi-range -> proper multipart 206 not 416;
+  416 Content-Range now `*/size`; ETag/Last-Modified now present; ASCII filename no longer
+  gets a redundant filename*= form).
+- Stage-2 behavior preserved through the rewrite: a served temp_* file still has its mtime
+  refreshed; delete_after still schedules the temp-file BackgroundTask (and it actually
+  runs); the media_key MIME cache is still consulted and populated.
+- The pure-ASGI RequestLoggingMiddleware: a normal request still returns (body intact, not
+  buffered/truncated) and is logged.
+
+Before/after Range matrix (2048-byte file), old = hand-rolled streaming, new = FileResponse:
+
+  request                | old                         | new (FileResponse)
+  -----------------------+-----------------------------+-----------------------------------
+  no Range               | 200 full                    | 200 full (+ETag/Last-Modified)
+  bytes=0-499            | 206 bytes 0-499/2048        | 206 bytes 0-499/2048 (same bytes)
+  bytes=500-             | 206 bytes 500-2047/2048     | 206 bytes 500-2047/2048 (same)
+  bytes=-500             | 206 bytes 1548-2047/2048    | 206 bytes 1548-2047/2048 (same)
+  bytes=999999- (>EOF)   | 416 `bytes */2048`          | 416 `*/2048`  (Starlette omits the
+                         |                             |   `bytes ` prefix; RFC-acceptable)
+  garbage header         | 416                         | 400 Bad Request (RFC 7233 lets a
+                         |                             |   server reject/ignore a malformed
+                         |                             |   Range; Starlette returns 400)
+  bytes=0-9,20-29        | 416 (old parser bug)        | 206 multipart/byteranges (correct)
+
+All 206 responses that carry data return byte-for-byte identical slices old vs new, so no
+real regression is hidden behind an "accepted difference".
+"""
+import os
+import time
+import logging
+
+import pytest
+
+from fastapi import FastAPI, Request
+from fastapi.testclient import TestClient
+
+import api_server
+
+
+# --------------------------------------------------------------------------- #
+# Helpers
+# --------------------------------------------------------------------------- #
+BODY = bytes(range(256)) * 8  # 2048 deterministic bytes
+SIZE = len(BODY)
+
+
+def _make_client(file_path, delete_after=False, media_key=None):
+    """Mount prepare_file_response on a tiny app so it is driven through real ASGI
+    (FileResponse computes Range/206/416 at send time, so it must be exercised via a client)."""
+    app = FastAPI()
+
+    @app.get("/f")
+    async def _serve(request: Request):
+        return await api_server.prepare_file_response(
+            file_path, request=request, delete_after=delete_after, media_key=media_key
+        )
+
+    return TestClient(app)
+
+
+@pytest.fixture
+def sample_file(tmp_path):
+    fp = tmp_path / "myfile.bin"
+    fp.write_bytes(BODY)
+    return str(fp)
+
+
+# --------------------------------------------------------------------------- #
+# 3.1 — Range matrix against the FINAL FileResponse behavior.
+# --------------------------------------------------------------------------- #
+def test_no_range_returns_full_200(sample_file):
+    c = _make_client(sample_file)
+    r = c.get("/f")
+    assert r.status_code == 200
+    assert r.content == BODY
+    assert r.headers["content-length"] == str(SIZE)
+    assert r.headers["accept-ranges"] == "bytes"
+    assert r.headers["cache-control"] == "public, max-age=86400, immutable"
+    # Content-Disposition: inline, filename set by FileResponse itself (no manual double-set).
+    assert r.headers["content-disposition"].startswith("inline")
+    assert 'filename="myfile.bin"' in r.headers["content-disposition"]
+    # NEW vs old: FileResponse adds validators. Old streaming had neither.
+    assert r.headers.get("etag")
+    assert r.headers.get("last-modified")
+
+
+def test_range_prefix_0_499(sample_file):
+    c = _make_client(sample_file)
+    r = c.get("/f", headers={"Range": "bytes=0-499"})
+    assert r.status_code == 206
+    assert r.headers["content-range"] == f"bytes 0-499/{SIZE}"
+    assert r.headers["content-length"] == "500"
+    assert r.content == BODY[:500]  # exact slice, identical to old behavior
+
+
+def test_range_open_ended_500_to_eof(sample_file):
+    c = _make_client(sample_file)
+    r = c.get("/f", headers={"Range": "bytes=500-"})
+    assert r.status_code == 206
+    assert r.headers["content-range"] == f"bytes 500-{SIZE - 1}/{SIZE}"
+    assert r.headers["content-length"] == str(SIZE - 500)
+    assert r.content == BODY[500:]
+
+
+def test_range_suffix_last_500(sample_file):
+    c = _make_client(sample_file)
+    r = c.get("/f", headers={"Range": "bytes=-500"})
+    assert r.status_code == 206
+    assert r.headers["content-range"] == f"bytes {SIZE - 500}-{SIZE - 1}/{SIZE}"
+    assert r.headers["content-length"] == "500"
+    assert r.content == BODY[-500:]
+
+
+def test_range_start_past_eof_returns_416(sample_file):
+    c = _make_client(sample_file)
+    r = c.get("/f", headers={"Range": "bytes=999999-"})
+    assert r.status_code == 416
+    # DELTA (RFC-acceptable): Starlette's 416 Content-Range is `*/size` (it omits the
+    # `bytes ` unit prefix the old code emitted). Same information, permitted by RFC 7233.
+    assert r.headers["content-range"] == f"*/{SIZE}"
+
+
+def test_garbage_range_header_returns_400(sample_file):
+    c = _make_client(sample_file)
+    r = c.get("/f", headers={"Range": "somethinggarbage"})
+    # DELTA (RFC-acceptable): the old hand-rolled parser returned 416 on an unparseable
+    # header; Starlette rejects a malformed Range with 400 Bad Request. RFC 7233 permits a
+    # server to reject/ignore a bad Range; this is a conscious accepted difference, not a
+    # data regression (no bytes are served either way).
+    assert r.status_code == 400
+
+
+def test_multi_range_returns_multipart_206(sample_file):
+    c = _make_client(sample_file)
+    r = c.get("/f", headers={"Range": "bytes=0-9,20-29"})
+    # DELTA (improvement): the old parser mis-split multi-ranges and returned 416.
+    # FileResponse serves a proper multipart/byteranges 206 containing both slices.
+    # Starlette 0.45.3 quirk: it advertises the multipart boundary via the `content-range`
+    # header (rather than content-type, which stays the file's own media_type); the body is
+    # a real multipart/byteranges document. We assert on the boundary marker + both slices.
+    assert r.status_code == 206
+    assert r.headers["content-range"].startswith("multipart/byteranges")
+    body = r.content
+    assert BODY[0:10] in body
+    assert BODY[20:30] in body
+
+
+# --------------------------------------------------------------------------- #
+# 3.1 (stage-2 preserved) — temp_* mtime touch survives the FileResponse swap.
+# --------------------------------------------------------------------------- #
+def test_temp_file_mtime_refreshed_when_stale(tmp_path):
+    # A temp_* file whose mtime is OLDER than the refresh interval gets touched, so the
+    # 1h sweeper won't delete an actively-viewed video.
+    temp_file = tmp_path / "temp_bigvid"
+    temp_file.write_bytes(b"videodata")
+    stale = time.time() - 10000  # >> TEMP_MTIME_REFRESH_INTERVAL (300s)
+    os.utime(temp_file, (stale, stale))
+
+    c = _make_client(str(temp_file))
+    r = c.get("/f")
+    assert r.status_code == 200
+    assert r.content == b"videodata"
+    assert os.path.getmtime(temp_file) > time.time() - 100  # refreshed by the serve
+
+
+def test_temp_file_mtime_stable_when_fresh(tmp_path):
+    # DEBOUNCE: a temp_* file touched RECENTLY (within the refresh interval) is NOT
+    # re-touched — so FileResponse's mtime-derived ETag stays stable across the rapid
+    # requests of a single resume/seek session and `If-Range` resume keeps working.
+    temp_file = tmp_path / "temp_freshvid"
+    temp_file.write_bytes(b"videodata")
+    recent = time.time() - 30  # << TEMP_MTIME_REFRESH_INTERVAL (300s)
+    os.utime(temp_file, (recent, recent))
+    mtime_before = os.path.getmtime(temp_file)
+
+    c = _make_client(str(temp_file))
+    r1 = c.get("/f")
+    r2 = c.get("/f")
+    assert r1.status_code == 200 and r2.status_code == 200
+    # mtime unchanged -> the ETag is identical across the two serves.
+    assert os.path.getmtime(temp_file) == mtime_before
+    assert r1.headers.get("etag") == r2.headers.get("etag")
+
+
+def test_non_temp_file_mtime_untouched(tmp_path):
+    normal = tmp_path / "regularfile"
+    normal.write_bytes(b"data")
+    stale = time.time() - 10000
+    os.utime(normal, (stale, stale))
+
+    c = _make_client(str(normal))
+    c.get("/f")
+    assert os.path.getmtime(normal) < time.time() - 5000  # left alone
+
+
+# --------------------------------------------------------------------------- #
+# 3.1 (stage-2 preserved) — delete_after still schedules the temp-file BackgroundTask.
+# --------------------------------------------------------------------------- #
+async def test_delete_after_attaches_background_task(tmp_path):
+    fp = tmp_path / "temp_todelete"
+    fp.write_bytes(b"x")
+
+    class _Req:
+        headers = {}
+
+    resp = await api_server.prepare_file_response(str(fp), request=_Req(), delete_after=True)
+    # FileResponse must carry a non-None background that deletes exactly this file.
+    assert resp.background is not None
+    assert resp.background.func is api_server.delayed_delete_file
+    assert resp.background.args == (str(fp),)
+
+
+async def test_no_delete_after_has_no_background(tmp_path):
+    fp = tmp_path / "keepme"
+    fp.write_bytes(b"x")
+
+    class _Req:
+        headers = {}
+
+    resp = await api_server.prepare_file_response(str(fp), request=_Req(), delete_after=False)
+    assert resp.background is None
+
+
+def test_delete_after_background_runs_and_removes_file(tmp_path, monkeypatch):
+    fp = tmp_path / "temp_gone"
+    fp.write_bytes(BODY)
+
+    # Patch the module-level deleter to remove immediately (real one sleeps 300s); the
+    # BackgroundTask picks up this reference at prepare_file_response call time.
+    async def _delete_now(path, delay=300):
+        os.remove(path)
+
+    monkeypatch.setattr(api_server, "delayed_delete_file", _delete_now)
+
+    c = _make_client(str(fp), delete_after=True)
+    r = c.get("/f")
+    assert r.status_code == 200
+    assert r.content == BODY
+    # TestClient blocks until the background task has run.
+    assert not os.path.exists(fp)
+
+
+# --------------------------------------------------------------------------- #
+# 3.1 (stage-2 preserved) — media_key MIME cache is consulted and populated.
+# --------------------------------------------------------------------------- #
+def test_media_key_mime_cache_hit_used(sample_file, monkeypatch):
+    calls = {"get": 0, "set": 0, "magic": 0}
+
+    def fake_get(db, ch, pid, fid):
+        calls["get"] += 1
+        return "video/mp4"  # cache HIT
+
+    def fake_set(*a, **k):
+        calls["set"] += 1
+
+    def fake_magic(_path):
+        calls["magic"] += 1
+        return "application/octet-stream"
+
+    monkeypatch.setattr(api_server, "get_mime_type_sync", fake_get)
+    monkeypatch.setattr(api_server, "set_mime_type_sync", fake_set)
+    monkeypatch.setattr(api_server.magic_mime, "from_file", fake_magic)
+
+    c = _make_client(sample_file, media_key=("chan", 42, "fid"))
+    r = c.get("/f")
+    assert r.status_code == 200
+    # Cache hit: python-magic never invoked, nothing re-written to the cache, MIME applied.
+    assert calls["get"] == 1
+    assert calls["magic"] == 0
+    assert calls["set"] == 0
+    assert r.headers["content-type"].startswith("video/mp4")
+
+
+def test_media_key_mime_cache_miss_populates(sample_file, monkeypatch):
+    written = {}
+
+    def fake_get(db, ch, pid, fid):
+        return None  # cache MISS
+
+    def fake_set(db, ch, pid, fid, mime):
+        written["mime"] = mime
+
+    def fake_magic(_path):
+        return "image/png"
+
+    monkeypatch.setattr(api_server, "get_mime_type_sync", fake_get)
+    monkeypatch.setattr(api_server, "set_mime_type_sync", fake_set)
+    monkeypatch.setattr(api_server.magic_mime, "from_file", fake_magic)
+
+    c = _make_client(sample_file, media_key=("chan", 42, "fid"))
+    r = c.get("/f")
+    assert r.status_code == 200
+    # Miss -> python-magic result is both applied and persisted for next time.
+    assert written.get("mime") == "image/png"
+    assert r.headers["content-type"].startswith("image/png")
+
+
+# --------------------------------------------------------------------------- #
+# 3.1 — 404 pre-check kept.
+# --------------------------------------------------------------------------- #
+def test_missing_file_returns_404(tmp_path):
+    c = _make_client(str(tmp_path / "does_not_exist"))
+    r = c.get("/f")
+    assert r.status_code == 404
+
+
+# --------------------------------------------------------------------------- #
+# 3.2 — pure-ASGI RequestLoggingMiddleware: request returns intact and is logged.
+# --------------------------------------------------------------------------- #
+def test_asgi_logging_middleware_passes_body_and_logs(tmp_path, caplog):
+    fp = tmp_path / "streamed.bin"
+    fp.write_bytes(BODY)
+
+    app = FastAPI()
+    app.add_middleware(api_server.RequestLoggingMiddleware)
+
+    @app.get("/f")
+    async def _serve(request: Request):
+        return await api_server.prepare_file_response(str(fp), request=request)
+
+    client = TestClient(app)
+    with caplog.at_level(logging.DEBUG, logger="api_server"):
+        r = client.get("/f")
+
+    # Body flows straight through the middleware — not buffered/truncated.
+    assert r.status_code == 200
+    assert r.content == BODY
+    messages = " ".join(rec.getMessage() for rec in caplog.records)
+    assert "Request: GET /f" in messages
+    assert "Response status: 200" in messages
+
+
+def test_asgi_logging_middleware_range_still_works(tmp_path):
+    fp = tmp_path / "streamed.bin"
+    fp.write_bytes(BODY)
+
+    app = FastAPI()
+    app.add_middleware(api_server.RequestLoggingMiddleware)
+
+    @app.get("/f")
+    async def _serve(request: Request):
+        return await api_server.prepare_file_response(str(fp), request=request)
+
+    client = TestClient(app)
+    r = client.get("/f", headers={"Range": "bytes=0-99"})
+    # 206 streaming still works through the pure-ASGI middleware (send not buffered).
+    assert r.status_code == 206
+    assert r.content == BODY[:100]
+
+
+async def test_asgi_logging_middleware_passes_non_http_scope_through(caplog):
+    # The `scope["type"] != "http"` branch (lifespan/websocket) only runs in a real
+    # deploy, never through TestClient's plain request path — so pin it directly:
+    # a non-http scope must delegate to the inner app untouched and log nothing.
+    called = {}
+
+    async def inner(scope, receive, send):
+        called["scope_type"] = scope["type"]
+
+    mw = api_server.RequestLoggingMiddleware(inner)
+    with caplog.at_level(logging.DEBUG, logger="api_server"):
+        await mw({"type": "lifespan"}, None, None)
+
+    assert called["scope_type"] == "lifespan"  # delegated to the inner app
+    # Nothing request/response-ish was logged for a non-http scope.
+    messages = " ".join(rec.getMessage() for rec in caplog.records)
+    assert "Request:" not in messages
+    assert "Response status:" not in messages

tests/test_stage4_eventloop.py

@@ -0,0 +1,487 @@
+# flake8: noqa
+# pylint: disable=protected-access, missing-function-docstring, missing-class-docstring
+# pylint: disable=redefined-outer-name, logging-fstring-interpolation, line-too-long
+# pylance: disable=reportMissingImports, reportMissingModuleSource
+"""
+Stage 4 (event-loop hygiene for feed generation) tests.
+
+Covers:
+- 4.1 raw_message laziness: feeds do NOT compute str(message); JSON/debug HTML do.
+- 4.2 side-effect IO removed from process_message: _save_media_file_ids only appends to
+      self._pending_media_ids; the caller flushes once via upsert_media_file_ids_bulk_sync.
+      Also: the render path contains NO create_task / get_running_loop / to_thread.
+- 4.3 render pipeline moved into a single thread: the four render functions are now plain
+      sync functions and actually execute off the main thread; deepcopy of a pickled Message
+      does not crash; a 100-message feed generates correctly.
+- 4.4 sanitize coverage (XSS): a <script> / onerror= / javascript: payload is stripped in
+      ALL outputs — rss, html-feed, single-post html, and json — each with exactly one pass.
+"""
+import re
+import copy
+import pickle
+import asyncio
+import threading
+from datetime import datetime, timezone
+from types import SimpleNamespace
+
+import pytest
+
+from pyrogram.enums import MessageMediaType
+
+import post_parser as pp_module
+import rss_generator as rss_module
+from post_parser import PostParser
+from rss_generator import (
+    generate_channel_rss,
+    generate_channel_html,
+    _render_pipeline,
+    _create_time_based_media_groups,
+    _create_messages_groups,
+    _trim_messages_groups,
+    _render_messages_groups,
+)
+
+XSS_PAYLOAD = "<script>alert('xss')</script><img src=x onerror=\"alert(1)\"><a href=\"javascript:alert(2)\">click</a>"
+
+
+class _Str(str):
+    """Minimal stand-in for Pyrogram's Str: .html returns the raw string unchanged,
+    so a malicious payload reaches the pre-sanitization html body just like real text
+    carrying entities would."""
+    @property
+    def html(self):
+        return str(self)
+
+
+def make_message(mid, text=None, media=None, photo_uid=None, username="testchan",
+                 date=None):
+    m = SimpleNamespace()
+    m.id = mid
+    m.date = date or datetime(2024, 1, 1, 12, 0, mid % 60, tzinfo=timezone.utc)
+    m.text = _Str(text) if text is not None else None
+    m.caption = None
+    m.media = media
+    m.web_page = None
+    m.poll = None
+    m.service = None
+    m.forward_origin = None
+    m.reply_to_message = None
+    m.reply_to_message_id = None
+    m.sender_chat = None
+    m.from_user = None
+    m.reactions = None
+    m.views = 100
+    m.media_group_id = None
+    m.show_caption_above_media = False
+    m.chat = SimpleNamespace(id=-1001234567890, username=username)
+    # media sub-objects default to None
+    for attr in ("photo", "video", "document", "audio", "voice",
+                 "video_note", "animation", "sticker"):
+        setattr(m, attr, None)
+    if media == MessageMediaType.PHOTO and photo_uid:
+        m.photo = SimpleNamespace(file_unique_id=photo_uid)
+    return m
+
+
+def _co_names(func):
+    return set(func.__code__.co_names)
+
+
+# ---------------------------------------------------------------------------
+# 4.3 — render functions are plain sync and run off the main thread
+# ---------------------------------------------------------------------------
+
+def test_render_functions_are_sync():
+    for fn in (_create_time_based_media_groups, _create_messages_groups,
+               _trim_messages_groups, _render_messages_groups, _render_pipeline):
+        assert not asyncio.iscoroutinefunction(fn), f"{fn.__name__} must be a plain sync function"
+
+
+@pytest.mark.asyncio
+async def test_pipeline_runs_in_worker_thread(monkeypatch):
+    main_ident = threading.get_ident()
+    seen = {}
+
+    real_render = rss_module._render_messages_groups
+
+    def spy(*args, **kwargs):
+        seen["ident"] = threading.get_ident()
+        return real_render(*args, **kwargs)
+
+    monkeypatch.setattr(rss_module, "_render_messages_groups", spy)
+
+    async def fake_get_chat(client, channel):
+        return SimpleNamespace(title="Test", username="testchan", id=-1001234567890)
+
+    async def fake_get_history(client, channel, limit=20):
+        return [make_message(i, text=f"post {i}") for i in range(5)]
+
+    monkeypatch.setattr("tg_cache.cached_get_chat", fake_get_chat, raising=False)
+    monkeypatch.setattr("tg_cache.cached_get_chat_history", fake_get_history, raising=False)
+
+    await generate_channel_rss("testchan", client=SimpleNamespace(), limit=5)
+    assert "ident" in seen
+    assert seen["ident"] != main_ident, "render pipeline must run in a worker thread, not the loop thread"
+
+
+def test_deepcopy_of_pickled_message_does_not_crash():
+    from pyrogram.types import Message, Chat
+    from pyrogram.enums import ChatType
+    m = Message(id=7, date=datetime.now(timezone.utc), text="hello",
+                chat=Chat(id=-1001, type=ChatType.CHANNEL, username="testchan"))
+    roundtripped = pickle.loads(pickle.dumps(m))   # mimics the pickle cache
+    clone = copy.deepcopy(roundtripped)            # what _create_time_based_media_groups does
+    assert clone.id == 7
+    assert clone.chat.username == "testchan"
+
+
+# ---------------------------------------------------------------------------
+# 4.2 — no asyncio in the render path; bulk upsert after render
+# ---------------------------------------------------------------------------
+
+def test_render_path_has_no_asyncio_side_effects():
+    banned = {"create_task", "get_running_loop", "to_thread", "ensure_future"}
+    funcs = [
+        _render_pipeline, _create_time_based_media_groups, _create_messages_groups,
+        _trim_messages_groups, _render_messages_groups,
+        PostParser.process_message, PostParser._generate_html_body,
+        PostParser._generate_html_media, PostParser.generate_html_footer,
+        PostParser._reactions_views_links, PostParser._save_media_file_ids,
+        PostParser._sanitize_html,
+    ]
+    for fn in funcs:
+        offenders = _co_names(fn) & banned
+        assert not offenders, f"{fn.__qualname__} references forbidden asyncio names: {offenders}"
+
+
+@pytest.mark.asyncio
+async def test_media_ids_persisted_via_bulk_upsert(monkeypatch):
+    calls = []
+
+    def fake_bulk(db_path, entries):
+        calls.append(list(entries))
+
+    monkeypatch.setattr(pp_module, "upsert_media_file_ids_bulk_sync", fake_bulk)
+
+    async def fake_get_chat(client, channel):
+        return SimpleNamespace(title="Test", username="testchan", id=-1001234567890)
+
+    async def fake_get_history(client, channel, limit=20):
+        return [
+            make_message(1, text="just text"),
+            make_message(2, media=MessageMediaType.PHOTO, photo_uid="uid_abc"),
+        ]
+
+    monkeypatch.setattr("tg_cache.cached_get_chat", fake_get_chat, raising=False)
+    monkeypatch.setattr("tg_cache.cached_get_chat_history", fake_get_history, raising=False)
+
+    await generate_channel_rss("testchan", client=SimpleNamespace(), limit=10)
+
+    assert len(calls) == 1, "bulk upsert must be called exactly once after render"
+    entries = calls[0]
+    assert len(entries) == 1, "only the photo message contributes a media id"
+    channel, post_id, fid, _ts = entries[0]
+    assert (channel, post_id, fid) == ("testchan", 2, "uid_abc")
+
+
+@pytest.mark.asyncio
+async def test_save_media_file_ids_only_appends(monkeypatch):
+    # Even with a running loop, _save_media_file_ids must not create tasks — just append.
+    parser = PostParser(SimpleNamespace())
+    monkeypatch.setattr(pp_module, "upsert_media_file_ids_bulk_sync",
+                        lambda *a, **k: (_ for _ in ()).throw(AssertionError("must not be called directly")))
+    msg = make_message(3, media=MessageMediaType.PHOTO, photo_uid="uid_x")
+    parser._save_media_file_ids(msg)
+    assert parser._pending_media_ids == [("testchan", 3, "uid_x", parser._pending_media_ids[0][3])]
+
+
+# ---------------------------------------------------------------------------
+# 4.1 — raw_message laziness
+# ---------------------------------------------------------------------------
+
+def test_raw_message_lazy_for_feed():
+    parser = PostParser(SimpleNamespace())
+    result = parser.process_message(make_message(10, text="hi"), include_raw=False, sanitize=False)
+    assert "raw_message" not in result
+
+
+def test_raw_message_present_for_json_and_debug():
+    parser = PostParser(SimpleNamespace())
+    result = parser.process_message(make_message(11, text="hi"), include_raw=True)
+    assert "raw_message" in result
+    assert isinstance(result["raw_message"], str)
+
+
+@pytest.mark.asyncio
+async def test_100_message_feed_generates(monkeypatch):
+    async def fake_get_chat(client, channel):
+        return SimpleNamespace(title="Test", username="testchan", id=-1001234567890)
+
+    async def fake_get_history(client, channel, limit=20):
+        return [make_message(i, text=f"post number {i}") for i in range(1, 101)]
+
+    monkeypatch.setattr("tg_cache.cached_get_chat", fake_get_chat, raising=False)
+    monkeypatch.setattr("tg_cache.cached_get_chat_history", fake_get_history, raising=False)
+
+    rss = await generate_channel_rss("testchan", client=SimpleNamespace(), limit=100)
+    assert rss.count("<item>") == 100
+    assert "post number 50" in rss
+
+
+# ---------------------------------------------------------------------------
+# 4.4 — XSS: payload stripped in all four outputs
+# ---------------------------------------------------------------------------
+
+def _assert_clean(html_str, where):
+    assert "<script>" not in html_str, f"{where}: <script> survived"
+    assert "onerror" not in html_str, f"{where}: onerror= survived"
+    assert "javascript:" not in html_str, f"{where}: javascript: survived"
+
+
+def _make_client_returning(msg):
+    client = SimpleNamespace()
+
+    async def get_messages(channel, post_id):
+        return msg
+
+    client.get_messages = get_messages
+    return client
+
+
+@pytest.mark.asyncio
+async def test_xss_stripped_in_json():
+    msg = make_message(20, text=XSS_PAYLOAD)
+    parser = PostParser(_make_client_returning(msg))
+    data = await parser.get_post("testchan", 20, "json")
+    _assert_clean(data["html"]["body"], "json body")
+    _assert_clean(data["html"]["footer"], "json footer")
+
+
+@pytest.mark.asyncio
+async def test_xss_stripped_in_single_post_html():
+    msg = make_message(21, text=XSS_PAYLOAD)
+    parser = PostParser(_make_client_returning(msg))
+    html_out = await parser.get_post("testchan", 21, "html")
+    _assert_clean(html_out, "single-post html")
+
+
+@pytest.mark.asyncio
+async def test_xss_stripped_in_single_post_html_debug():
+    # debug HTML embeds raw_message (str(message)) into a <pre>; it must be html-escaped so
+    # no live tag from the payload survives. The escaped dump legitimately still contains the
+    # inert words "onerror"/"javascript:" as text — what matters is that they are NOT live.
+    msg = make_message(22, text=XSS_PAYLOAD)
+    parser = PostParser(_make_client_returning(msg))
+    html_out = await parser.get_post("testchan", 22, "html", debug=True)
+
+    # 1) The rendered display area (everything before the raw <pre>) is fully sanitized.
+    display = html_out.split('<pre', 1)[0]
+    _assert_clean(display, "single-post debug display")
+
+    # 2) The raw <pre> dump is html-escaped: no live <script> tag anywhere, and the payload
+    #    appears only in escaped form (proving html.escape ran).
+    assert "<script>" not in html_out, "debug raw dump left a live <script> tag"
+    assert "&lt;script&gt;" in html_out, "debug raw dump was not html-escaped"
+
+
+@pytest.mark.asyncio
+async def test_xss_in_title_escaped_in_debug_html():
+    # Issue #13: the debug branch of _format_html embeds data["html"]["title"], which is
+    # generated from user-controlled content and never passes through bleach. A poll whose
+    # question carries a <script> tag yields the title "📊 Poll: <script>alert(1)</script>"
+    # verbatim — it must be html-escaped before being embedded in the debug output.
+    msg = make_message(25, media=MessageMediaType.POLL)
+    msg.poll = SimpleNamespace(question="<script>alert(1)</script>")
+    parser = PostParser(_make_client_returning(msg))
+    html_out = await parser.get_post("testchan", 25, "html", debug=True)
+
+    # No live <script> tag anywhere in the output (title div, body, footer, raw <pre>).
+    assert "<script>" not in html_out, "debug html left a live <script> tag"
+    # The title div itself carries the payload only in escaped form (proving html.escape
+    # ran on the title, not just on the raw_message <pre> dump).
+    title_lines = [line for line in html_out.split("\n") if 'class="title"' in line]
+    assert title_lines, "debug output must contain the title div"
+    assert "&lt;script&gt;alert(1)&lt;/script&gt;" in title_lines[0], \
+        "title was not html-escaped in debug output"
+
+
+@pytest.mark.asyncio
+async def test_xss_stripped_in_rss_feed(monkeypatch):
+    async def fake_get_chat(client, channel):
+        return SimpleNamespace(title="Test", username="testchan", id=-1001234567890)
+
+    async def fake_get_history(client, channel, limit=20):
+        return [make_message(23, text=XSS_PAYLOAD)]
+
+    monkeypatch.setattr("tg_cache.cached_get_chat", fake_get_chat, raising=False)
+    monkeypatch.setattr("tg_cache.cached_get_chat_history", fake_get_history, raising=False)
+
+    rss = await generate_channel_rss("testchan", client=SimpleNamespace(), limit=5)
+    # The sanitized HTML lives in <content:encoded><![CDATA[...]]></content:encoded>.
+    cdata = re.findall(r"<!\[CDATA\[(.*?)\]\]>", rss, re.DOTALL)
+    assert cdata, "expected CDATA content in RSS"
+    for chunk in cdata:
+        _assert_clean(chunk, "rss content")
+
+
+@pytest.mark.asyncio
+async def test_xss_stripped_in_html_feed(monkeypatch):
+    async def fake_get_chat(client, channel):
+        return SimpleNamespace(title="Test", username="testchan", id=-1001234567890)
+
+    async def fake_get_history(client, channel, limit=20):
+        return [make_message(24, text=XSS_PAYLOAD)]
+
+    monkeypatch.setattr("tg_cache.cached_get_chat", fake_get_chat, raising=False)
+    monkeypatch.setattr("tg_cache.cached_get_chat_history", fake_get_history, raising=False)
+
+    html_feed = await generate_channel_html("testchan", client=SimpleNamespace(), limit=5)
+    _assert_clean(html_feed, "html feed")
+
+
+def _media_msg_with_payload_caption(mid):
+    # A photo message whose CAPTION carries the payload — this exercises the media
+    # fragment path (_generate_html_media / caption rendering) whose internal per-fragment
+    # sanitize pass was removed in 4.4. The covering pass must still strip it.
+    m = make_message(mid, media=MessageMediaType.PHOTO, photo_uid="pic123")
+    m.caption = _Str(XSS_PAYLOAD)
+    return m
+
+
+@pytest.mark.asyncio
+async def test_xss_in_media_caption_stripped_direct_paths():
+    # json + single-post html go through process_message(sanitize=True) directly.
+    parser = PostParser(_make_client_returning(_media_msg_with_payload_caption(30)))
+    data = await parser.get_post("testchan", 30, "json")
+    _assert_clean(data["html"]["body"], "json body (media caption)")
+    _assert_clean(data["html"]["footer"], "json footer (media caption)")
+
+    parser2 = PostParser(_make_client_returning(_media_msg_with_payload_caption(31)))
+    html_out = await parser2.get_post("testchan", 31, "html")
+    _assert_clean(html_out, "single-post html (media caption)")
+
+
+@pytest.mark.asyncio
+async def test_xss_in_media_caption_stripped_in_feeds(monkeypatch):
+    async def fake_get_chat(client, channel):
+        return SimpleNamespace(title="Test", username="testchan", id=-1001234567890)
+
+    async def fake_get_history(client, channel, limit=20):
+        return [_media_msg_with_payload_caption(32)]
+
+    monkeypatch.setattr("tg_cache.cached_get_chat", fake_get_chat, raising=False)
+    monkeypatch.setattr("tg_cache.cached_get_chat_history", fake_get_history, raising=False)
+
+    rss = await generate_channel_rss("testchan", client=SimpleNamespace(), limit=5)
+    for chunk in re.findall(r"<!\[CDATA\[(.*?)\]\]>", rss, re.DOTALL):
+        _assert_clean(chunk, "rss content (media caption)")
+
+    async def fake_get_history2(client, channel, limit=20):
+        return [_media_msg_with_payload_caption(33)]
+    monkeypatch.setattr("tg_cache.cached_get_chat_history", fake_get_history2, raising=False)
+    html_feed = await generate_channel_html("testchan", client=SimpleNamespace(), limit=5)
+    _assert_clean(html_feed, "html feed (media caption)")
+
+
+# --------------------------------------------------------------------------- #
+# Review round-1: the new bulk-upsert SQL executed for real (not mocked).
+# --------------------------------------------------------------------------- #
+def test_bulk_upsert_media_file_ids_real_sql(tmp_path):
+    import sqlite3
+    from file_io import upsert_media_file_ids_bulk_sync, init_db_sync
+
+    db = str(tmp_path / "t.db")
+    init_db_sync(db)
+
+    # Empty list is a no-op (no crash).
+    upsert_media_file_ids_bulk_sync(db, [])
+
+    # Multi-row insert.
+    upsert_media_file_ids_bulk_sync(db, [
+        ("chA", 1, "fidA", 100.0),
+        ("chB", 2, "fidB", 200.0),
+    ])
+    conn = sqlite3.connect(db)
+    rows = dict(((c, p, f), a) for c, p, f, a in
+                conn.execute("SELECT channel, post_id, file_unique_id, added FROM media_file_ids"))
+    assert rows[("chA", 1, "fidA")] == 100.0
+    assert rows[("chB", 2, "fidB")] == 200.0
+
+    # Re-upsert the SAME key updates `added` (ON CONFLICT ... DO UPDATE SET added=excluded.added).
+    upsert_media_file_ids_bulk_sync(db, [("chA", 1, "fidA", 999.0)])
+    a = conn.execute(
+        "SELECT added FROM media_file_ids WHERE channel='chA' AND post_id=1 AND file_unique_id='fidA'"
+    ).fetchone()[0]
+    assert a == 999.0
+    conn.close()
+
+
+# --------------------------------------------------------------------------- #
+# Review round-1: media ids collected before a render exception are still
+# flushed (the flush is in a finally). Removing the finally must break this.
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_pending_media_flushed_on_render_exception(monkeypatch):
+    async def fake_get_chat(client, channel):
+        return SimpleNamespace(title="Test", username="testchan", id=-1001234567890)
+
+    async def fake_get_history(client, channel, limit=20):
+        return [make_message(40, text="hi")]
+
+    monkeypatch.setattr("tg_cache.cached_get_chat", fake_get_chat, raising=False)
+    monkeypatch.setattr("tg_cache.cached_get_chat_history", fake_get_history, raising=False)
+
+    # A render pipeline that collects a pending id then raises mid-render.
+    def boom_pipeline(messages, post_parser, *a, **k):
+        post_parser._pending_media_ids.append(("chZ", 9, "fidZ", 1.0))
+        raise RuntimeError("render blew up")
+
+    monkeypatch.setattr(rss_module, "_render_pipeline", boom_pipeline)
+
+    flushed = {}
+    async def fake_bulk(db, entries):
+        flushed["entries"] = list(entries)
+    monkeypatch.setattr("post_parser.upsert_media_file_ids_bulk_sync",
+                        lambda db, entries: flushed.__setitem__("entries", list(entries)),
+                        raising=False)
+
+    with pytest.raises(Exception):
+        await generate_channel_rss("testchan", client=SimpleNamespace(), limit=5)
+
+    # The collected id was persisted despite the render exception (flush in finally).
+    assert flushed.get("entries") == [("chZ", 9, "fidZ", 1.0)]
+
+
+# --------------------------------------------------------------------------- #
+# Review round-1 [security]: if the ONLY sanitize pass throws, the feed must
+# FAIL CLOSED (html.escape the raw content), never emit the raw XSS payload.
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_rss_fails_closed_when_sanitizer_raises(monkeypatch):
+    async def fake_get_chat(client, channel):
+        return SimpleNamespace(title="Test", username="testchan", id=-1001234567890)
+
+    async def fake_get_history(client, channel, limit=20):
+        return [make_message(41, text=XSS_PAYLOAD)]
+
+    monkeypatch.setattr("tg_cache.cached_get_chat", fake_get_chat, raising=False)
+    monkeypatch.setattr("tg_cache.cached_get_chat_history", fake_get_history, raising=False)
+
+    # Force bleach to blow up (e.g. the RecursionError class already seen in prod).
+    # rss_generator imports it as `from bleach import clean as HTMLSanitizer`.
+    def boom(*a, **k):
+        raise RecursionError("bleach exploded")
+    monkeypatch.setattr(rss_module, "HTMLSanitizer", boom, raising=True)
+
+    rss = await generate_channel_rss("testchan", client=SimpleNamespace(), limit=5)
+    chunks = re.findall(r"<!\[CDATA\[(.*?)\]\]>", rss, re.DOTALL)
+    assert chunks, "expected CDATA content"
+    for chunk in chunks:
+        # Fail-closed: the raw payload was html.escaped, so NO live tag survived — every
+        # `<` became `&lt;`. (The letters "javascript:" still appear, but as inert text
+        # inside an escaped &quot;…&quot;, not a live href.)
+        assert "<script" not in chunk, "RSS fail-open: raw <script> reached the feed"
+        assert "<img" not in chunk, "RSS fail-open: raw <img onerror> reached the feed"
+        assert "<a " not in chunk, "RSS fail-open: raw <a href> reached the feed"
+        # ...and the escaping actually ran (payload present as escaped text, not dropped).
+        assert "&lt;script&gt;" in chunk, "expected the payload html-escaped, not dropped"

tests/test_stage5_sqlite.py

@@ -0,0 +1,258 @@
+# flake8: noqa
+# pylint: disable=protected-access, missing-function-docstring, missing-class-docstring
+# pylint: disable=redefined-outer-name, logging-fstring-interpolation, line-too-long
+# pylance: disable=reportMissingImports, reportMissingModuleSource
+"""
+Stage 5 (SQLite access-time batching) tests.
+
+Covers:
+- 5.1 accumulator: a /media cache hit records into api_server._access_updates and does NOT
+      call update_media_file_access_sync (no synchronous SQLite write on the hot path).
+- 5.1 flush: seeding _access_updates + running the flush once writes the accumulated
+      timestamps to a real temp DB (hit -> flush -> value updated in DB), and clears the dict.
+- 5.1 bulk fn update_media_file_access_bulk_sync: empty no-op, multi-row, and updating an
+      EXISTING row's `added` (WHERE matches on the str channel key).
+- 5.1 snapshot-then-clear: an update arriving DURING the flush lands in the fresh dict and
+      is not lost.
+- gotcha: str(channel) key discipline — an int-ish channel on the hot path keys the
+      accumulator (and thus the UPDATE) by the string form.
+"""
+import sqlite3
+import asyncio
+from types import SimpleNamespace
+
+import pytest
+
+import api_server
+from file_io import init_db_sync, update_media_file_access_bulk_sync
+
+
+def _fake_plain_message():
+    """A non-poll, non-video message so download_media_file takes the normal cache flow."""
+    return SimpleNamespace(media=None, video=None, empty=False)
+
+
+@pytest.fixture(autouse=True)
+def _clean_accumulator():
+    """Each test starts with an empty accumulator and restores it afterwards."""
+    api_server._access_updates = {}
+    yield
+    api_server._access_updates = {}
+
+
+# --------------------------------------------------------------------------- #
+# 5.1 hot path: cache hit records into the accumulator, no synchronous SQLite.
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_cache_hit_records_accumulator_no_sqlite(tmp_path, monkeypatch):
+    monkeypatch.chdir(tmp_path)
+    channel, post_id, fid = "testchan", 5, "fidHIT"
+    cache_dir = tmp_path / "data" / "cache" / channel / str(post_id)
+    cache_dir.mkdir(parents=True)
+    (cache_dir / fid).write_bytes(b"cached-bytes")
+
+    async def fake_get(_cid, _pid):
+        return _fake_plain_message()
+    monkeypatch.setattr(api_server.client, "safe_get_messages", fake_get)
+
+    # Spy: the single-row synchronous updater must NOT be called on the hot path.
+    called = []
+    monkeypatch.setattr(api_server, "update_media_file_access_sync",
+                        lambda *a, **k: called.append(a))
+
+    path, delete_after = await api_server.download_media_file(channel, post_id, fid)
+
+    assert path == str(cache_dir / fid)
+    assert delete_after is False
+    assert called == []  # no synchronous SQLite access-write happened
+    assert (channel, post_id, fid) in api_server._access_updates
+
+
+# --------------------------------------------------------------------------- #
+# 5.1 hot path (DoD guard): get_media's pre-semaphore cache-hit — the hottest
+# changed site, the one this PR exists for — records into the accumulator and
+# does NO synchronous SQLite access-write. Mirrors the download_media_file spy
+# so a regression re-introducing a per-hit write into THIS branch goes red.
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_get_media_pre_semaphore_cache_hit_no_sqlite(tmp_path, monkeypatch):
+    monkeypatch.chdir(tmp_path)
+    channel, post_id, fid = "gmchan", 11, "fidGM"
+    cache_dir = tmp_path / "data" / "cache" / channel / str(post_id)
+    cache_dir.mkdir(parents=True)
+    (cache_dir / fid).write_bytes(b"cached-bytes")
+
+    # Bypass the HMAC digest gate and the FileResponse machinery — the test targets
+    # only the access-time write on the pre-semaphore cache-hit branch.
+    monkeypatch.setattr(api_server, "verify_media_digest", lambda *a, **k: True)
+    sentinel = object()
+    async def fake_prepare(cache_path, request=None, media_key=None):
+        return sentinel
+    monkeypatch.setattr(api_server, "prepare_file_response", fake_prepare)
+
+    # Spy: the single-row synchronous updater must NOT be called on the hot path.
+    called = []
+    monkeypatch.setattr(api_server, "update_media_file_access_sync",
+                        lambda *a, **k: called.append(a))
+
+    resp = await api_server.get_media(channel, post_id, fid, request=object(), digest="x")
+
+    assert resp is sentinel
+    assert called == []  # no synchronous SQLite access-write on the hottest path
+    assert (channel, post_id, fid) in api_server._access_updates
+
+
+# --------------------------------------------------------------------------- #
+# gotcha: str(channel) key discipline on the hot path (int-ish channel).
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_cache_hit_keys_channel_as_str(tmp_path, monkeypatch):
+    monkeypatch.chdir(tmp_path)
+    channel_int, post_id, fid = 12345, 7, "fidINT"
+    cache_dir = tmp_path / "data" / "cache" / str(channel_int) / str(post_id)
+    cache_dir.mkdir(parents=True)
+    (cache_dir / fid).write_bytes(b"x")
+
+    async def fake_get(_cid, _pid):
+        return _fake_plain_message()
+    monkeypatch.setattr(api_server.client, "safe_get_messages", fake_get)
+
+    await api_server.download_media_file(channel_int, post_id, fid)
+
+    assert ("12345", post_id, fid) in api_server._access_updates      # string form
+    assert (channel_int, post_id, fid) not in api_server._access_updates  # never the raw int
+
+
+# --------------------------------------------------------------------------- #
+# 5.1 flush: hit -> flush -> the accumulated timestamp is written to the DB.
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_flush_writes_accumulated_timestamps(tmp_path, monkeypatch):
+    db = str(tmp_path / "flush.db")
+    init_db_sync(db)
+    # Seed two existing rows with an old timestamp.
+    conn = sqlite3.connect(db)
+    conn.executemany(
+        "INSERT INTO media_file_ids (channel, post_id, file_unique_id, added) VALUES (?,?,?,?)",
+        [("chA", 1, "fA", 1.0), ("chB", 2, "fB", 2.0)],
+    )
+    conn.commit()
+    conn.close()
+
+    monkeypatch.setattr(api_server, "DB_PATH", db)
+    api_server._access_updates = {
+        ("chA", 1, "fA"): 111.0,
+        ("chB", 2, "fB"): 222.0,
+    }
+
+    await api_server._flush_access_updates()
+
+    # Dict cleared after flush.
+    assert api_server._access_updates == {}
+
+    conn = sqlite3.connect(db)
+    rows = dict(((c, p, f), a) for c, p, f, a in
+                conn.execute("SELECT channel, post_id, file_unique_id, added FROM media_file_ids"))
+    conn.close()
+    assert rows[("chA", 1, "fA")] == 111.0
+    assert rows[("chB", 2, "fB")] == 222.0
+
+
+@pytest.mark.asyncio
+async def test_flush_empty_is_noop(monkeypatch):
+    api_server._access_updates = {}
+    # Must not raise and must not touch the threadpool/DB.
+    monkeypatch.setattr(api_server, "update_media_file_access_bulk_sync",
+                        lambda *a, **k: pytest.fail("bulk sync should not run for an empty batch"))
+    await api_server._flush_access_updates()
+
+
+# --------------------------------------------------------------------------- #
+# 5.1 snapshot-then-clear: an update arriving DURING the flush is not lost.
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_snapshot_then_clear_does_not_lose_late_update(tmp_path, monkeypatch):
+    db = str(tmp_path / "race.db")
+    init_db_sync(db)
+    monkeypatch.setattr(api_server, "DB_PATH", db)
+
+    late_key = ("chLate", 9, "fLate")
+
+    def fake_bulk(_db, entries):
+        # Simulates a cache hit landing WHILE the flush's to_thread runs: because the flush
+        # already replaced the module dict with a fresh one, this write goes into the NEW dict.
+        api_server._access_updates[late_key] = 999.0
+
+    monkeypatch.setattr(api_server, "update_media_file_access_bulk_sync", fake_bulk)
+
+    api_server._access_updates = {("chA", 1, "fA"): 111.0}
+    await api_server._flush_access_updates()
+
+    # The snapshot (chA) was flushed and cleared; the late update survives in the new dict.
+    assert api_server._access_updates == {late_key: 999.0}
+
+
+# --------------------------------------------------------------------------- #
+# 5.1 re-queue on failure: a failed bulk write is not lost; setdefault keeps a
+# fresher write that arrived during the flush.
+# --------------------------------------------------------------------------- #
+@pytest.mark.asyncio
+async def test_failed_flush_requeues_batch_without_clobbering_fresh(tmp_path, monkeypatch):
+    db = str(tmp_path / "fail.db")
+    init_db_sync(db)
+    monkeypatch.setattr(api_server, "DB_PATH", db)
+
+    stale_key = ("chStale", 1, "fS")   # only in the failing snapshot
+    fresh_key = ("chFresh", 2, "fF")   # re-written FRESHER during the flush
+
+    def fake_bulk(_db, _entries):
+        # A newer cache hit for fresh_key lands in the fresh dict WHILE the bulk write runs,
+        # then the write fails. The re-queue must restore stale_key but must NOT overwrite
+        # the newer fresh_key value (setdefault, not assignment).
+        api_server._access_updates[fresh_key] = 999.0
+        raise sqlite3.OperationalError("database is locked")
+
+    monkeypatch.setattr(api_server, "update_media_file_access_bulk_sync", fake_bulk)
+
+    api_server._access_updates = {stale_key: 1.0, fresh_key: 2.0}
+    with pytest.raises(sqlite3.OperationalError):
+        await api_server._flush_access_updates()
+
+    # stale_key restored with its snapshot value; fresh_key keeps the NEWER value (not 2.0).
+    assert api_server._access_updates == {stale_key: 1.0, fresh_key: 999.0}
+
+
+# --------------------------------------------------------------------------- #
+# 5.1 bulk fn: empty no-op, multi-row, and existing-row update via str channel key.
+# --------------------------------------------------------------------------- #
+def test_bulk_access_update_real_sql(tmp_path):
+    db = str(tmp_path / "bulk.db")
+    init_db_sync(db)
+
+    # Empty batch is a no-op (no crash).
+    update_media_file_access_bulk_sync(db, [])
+
+    # Seed existing rows.
+    conn = sqlite3.connect(db)
+    conn.executemany(
+        "INSERT INTO media_file_ids (channel, post_id, file_unique_id, added) VALUES (?,?,?,?)",
+        [("chA", 1, "fA", 1.0), ("chB", 2, "fB", 2.0)],
+    )
+    conn.commit()
+    conn.close()
+
+    # Multi-row update of EXISTING rows: WHERE matches on the str channel key.
+    update_media_file_access_bulk_sync(db, [
+        ("chA", 1, "fA", 500.0),
+        ("chB", 2, "fB", 600.0),
+    ])
+
+    conn = sqlite3.connect(db)
+    rows = dict(((c, p, f), a) for c, p, f, a in
+                conn.execute("SELECT channel, post_id, file_unique_id, added FROM media_file_ids"))
+    # A row keyed by an int channel does NOT match the string "chA" WHERE (documents the gotcha).
+    n = conn.execute("SELECT COUNT(*) FROM media_file_ids WHERE channel = 1").fetchone()[0]
+    conn.close()
+    assert rows[("chA", 1, "fA")] == 500.0
+    assert rows[("chB", 2, "fB")] == 600.0
+    assert n == 0

tests/test_stage6_healthcheck.py

@@ -0,0 +1,182 @@
+# flake8: noqa
+# pylint: disable=protected-access, missing-function-docstring, missing-class-docstring
+# pylint: disable=redefined-outer-name, logging-fstring-interpolation, line-too-long
+# pylance: disable=reportMissingImports, reportMissingModuleSource
+"""
+Stage 6 (lightweight /ping healthcheck) regression tests.
+
+Covers:
+- /ping returns 200 "ok" when connected and the last probe is recent (age < threshold).
+- /ping returns 503 "degraded" when connected but the last probe is stale (age > threshold).
+- /ping returns 503 "degraded" when disconnected, regardless of probe age.
+- /ping returns 200 "ok" on a fresh boot (age is None) while connected — a freshly-started
+  container must NOT be killed before the watchdog's first probe.
+- ANTI-REGRESSION (the critical invariant): /ping issues ZERO Telegram RPC. A spy on the
+  fake client's get_me / safe_get_messages proves neither is ever called.
+- TelegramClient.watchdog_last_ok_age(): None when never probed; a positive float afterwards.
+"""
+import time
+
+import pytest
+
+from fastapi.testclient import TestClient
+
+import api_server
+from telegram_client import TelegramClient
+
+
+# --------------------------------------------------------------------------- #
+# Fakes
+# --------------------------------------------------------------------------- #
+class _FakeKurigramClient:
+    """Stands in for TelegramClient.client — exposes is_connected and RPC spies."""
+    def __init__(self, is_connected=True):
+        self.is_connected = is_connected
+        self.get_me_calls = 0
+
+    async def get_me(self):
+        # If /ping ever touches this, the whole point of the endpoint is defeated.
+        self.get_me_calls += 1
+        raise AssertionError("/ping must never call get_me()")
+
+
+class _FakeTelegramClient:
+    """Stands in for api_server.client with a controllable probe age + RPC spies."""
+    def __init__(self, age, is_connected=True):
+        self._age = age
+        self.client = _FakeKurigramClient(is_connected=is_connected)
+        self.safe_get_messages_calls = 0
+
+    def watchdog_last_ok_age(self):
+        return self._age
+
+    async def safe_get_messages(self, *args, **kwargs):
+        self.safe_get_messages_calls += 1
+        raise AssertionError("/ping must never call safe_get_messages()")
+
+
+@pytest.fixture
+def patch_client(monkeypatch):
+    """Return a factory that installs a fake api_server.client and yields a TestClient."""
+    def _install(age, is_connected=True):
+        fake = _FakeTelegramClient(age=age, is_connected=is_connected)
+        monkeypatch.setattr(api_server, "client", fake)
+        return fake, TestClient(api_server.app)
+    return _install
+
+
+THRESHOLD = api_server.Config["tg_ping_unhealthy_after"]  # 250 in mock_config
+
+
+# --------------------------------------------------------------------------- #
+# /ping endpoint behavior
+# --------------------------------------------------------------------------- #
+def test_ping_healthy_connected_recent(patch_client):
+    fake, tc = patch_client(age=THRESHOLD - 10, is_connected=True)
+    r = tc.get("/ping")
+    assert r.status_code == 200
+    body = r.json()
+    assert body["status"] == "ok"
+    assert body["connected"] is True
+    assert body["last_probe_age_s"] == round(THRESHOLD - 10, 1)
+    assert body["threshold_s"] == THRESHOLD
+    assert fake.client.get_me_calls == 0
+
+
+def test_ping_degraded_stale_probe(patch_client):
+    fake, tc = patch_client(age=THRESHOLD + 100, is_connected=True)
+    r = tc.get("/ping")
+    assert r.status_code == 503
+    body = r.json()
+    assert body["status"] == "degraded"
+    assert body["connected"] is True
+    assert fake.client.get_me_calls == 0
+
+
+def test_ping_watchdog_disabled_stale_age_still_healthy(patch_client, monkeypatch):
+    # With the watchdog OFF, nothing refreshes age (a disconnect-flap restart can stamp it
+    # once, then it only grows). A stale age must NOT drive /ping to 503 on a live connection
+    # — that would spuriously fail the healthcheck and trigger an autoheal restart. So with the
+    # watchdog disabled, /ping is a pure connectivity check: connected + stale age => healthy.
+    monkeypatch.setitem(api_server.Config, "tg_watchdog_enabled", False)
+    fake, tc = patch_client(age=THRESHOLD + 100, is_connected=True)
+    r = tc.get("/ping")
+    assert r.status_code == 200
+    body = r.json()
+    assert body["status"] == "ok"
+    assert body["connected"] is True
+    assert fake.client.get_me_calls == 0
+
+
+def test_ping_degraded_disconnected(patch_client):
+    # Even with a fresh probe age, a disconnected client is unhealthy.
+    fake, tc = patch_client(age=1.0, is_connected=False)
+    r = tc.get("/ping")
+    assert r.status_code == 503
+    body = r.json()
+    assert body["status"] == "degraded"
+    assert body["connected"] is False
+
+
+def test_ping_degraded_disconnected_even_when_age_none(patch_client):
+    fake, tc = patch_client(age=None, is_connected=False)
+    r = tc.get("/ping")
+    assert r.status_code == 503
+    assert r.json()["status"] == "degraded"
+
+
+def test_ping_fresh_boot_age_none_is_healthy(patch_client):
+    # Right after boot the watchdog hasn't probed yet (age None); connected => healthy.
+    fake, tc = patch_client(age=None, is_connected=True)
+    r = tc.get("/ping")
+    assert r.status_code == 200
+    body = r.json()
+    assert body["status"] == "ok"
+    assert body["last_probe_age_s"] is None
+    assert body["threshold_s"] == THRESHOLD
+
+
+def test_ping_pre_start_connected_none_is_degraded_bool(patch_client):
+    # Before client.start(), Kurigram's is_connected is None. /ping must not 500: it coerces
+    # to a bool, so "connected" is false (never null) and the endpoint reports 503 degraded.
+    fake, tc = patch_client(age=None, is_connected=None)
+    r = tc.get("/ping")
+    assert r.status_code == 503
+    body = r.json()
+    assert body["status"] == "degraded"
+    assert body["connected"] is False  # bool, not null
+    assert fake.client.get_me_calls == 0
+
+
+def test_ping_issues_no_tg_rpc(patch_client):
+    """The critical invariant: /ping never issues any Telegram RPC in any branch."""
+    for age, connected in [(1.0, True), (THRESHOLD + 500, True), (None, True), (1.0, False)]:
+        fake, tc = patch_client(age=age, is_connected=connected)
+        tc.get("/ping")
+        assert fake.client.get_me_calls == 0, f"get_me called (age={age}, connected={connected})"
+        assert fake.safe_get_messages_calls == 0, f"safe_get_messages called (age={age}, connected={connected})"
+
+
+def test_ping_route_needs_no_token(patch_client):
+    # /ping is unauthenticated by design (no token path variant); it just works.
+    fake, tc = patch_client(age=1.0, is_connected=True)
+    assert tc.get("/ping").status_code == 200
+
+
+# --------------------------------------------------------------------------- #
+# TelegramClient.watchdog_last_ok_age accessor
+# --------------------------------------------------------------------------- #
+def test_watchdog_last_ok_age_none_when_never_probed():
+    tgc = TelegramClient()
+    assert tgc._wd_last_ok_monotonic is None
+    assert tgc.watchdog_last_ok_age() is None
+
+
+def test_watchdog_last_ok_age_positive_after_probe():
+    tgc = TelegramClient()
+    tgc._wd_last_ok_monotonic = time.monotonic() - 5
+    age = tgc.watchdog_last_ok_age()
+    assert age is not None
+    assert age >= 5.0
+    # Sanity: a plausible upper bound so we didn't accidentally read the wrong field.
+    assert age < 60.0

tests/test_stage7_integration.py

@@ -0,0 +1,343 @@
+# flake8: noqa
+# pylint: disable=protected-access, missing-function-docstring, missing-class-docstring
+# pylint: disable=redefined-outer-name, logging-fstring-interpolation, line-too-long
+# pylance: disable=reportMissingImports, reportMissingModuleSource
+"""
+Stage 7 — cross-stage END-TO-END integration tests.
+
+Unlike the per-stage suites (which pin one seam in isolation), these drive the real
+public entry points (`get_media`, `ping`, the access-time flush) so a regression that
+only shows up when the stages interact goes red. Every scenario here maps to one of the
+plan's "Стадия 7 — сквозные ручные сценарии"; the ones that genuinely need a running
+server + real downloads + lsof (fd-leak counting) are NOT faked here — they stay in
+docs/stability-verification.md for the operator's prod observation.
+
+Automated here:
+- Range semantics at the /media ROUTE level (get_media -> prepare_file_response ->
+  FileResponse driven through real ASGI): bytes=0-99 -> 206, bytes=-100 -> 206,
+  bytes=999999999- -> 416. (Stage 3 pins these on prepare_file_response directly; this
+  adds the end-to-end assertion that get_media's cache-hit branch reaches FileResponse
+  with a live Range still honored — i.e. stages 2+3 wired together.)
+- /ping (stage 6) stays fast + correct while a deliberately-slow op is in flight, issuing
+  ZERO Telegram RPC — proving the healthcheck is decoupled from the hot/blocked paths.
+- In-flight dedup + disconnect cleanup (stages 1/2) through the REAL get_media entry
+  point: concurrent requests share one download and the _inflight registry drains; a
+  cancelled request (client disconnect) leaves neither a stuck key nor a hung sibling.
+- str(channel) access-time key consistency (stage 5) end-to-end: a /media cache hit for
+  an int-ish channel records the str-keyed timestamp, and the flush UPDATE matches the
+  str-keyed DB row (hit -> flush -> DB), the exact affinity gotcha the plan warns about.
+"""
+import time
+import sqlite3
+import asyncio
+from types import SimpleNamespace
+
+import pytest
+
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+import api_server
+from pyrogram import errors
+from file_io import init_db_sync
+
+
+# 2048 deterministic bytes so Range slices are byte-checkable.
+BODY = bytes(range(256)) * 8
+SIZE = len(BODY)
+
+
+def _media_app():
+    """A bare app that mounts the REAL get_media (and ping) with NO lifespan, so
+    client.start() never runs — a pure cache hit never touches Telegram, and FileResponse
+    computes Range/206/416 at send time, which only happens when driven through ASGI."""
+    app = FastAPI()
+    app.add_api_route("/media/{channel}/{post_id}/{file_unique_id}/{digest}", api_server.get_media, methods=["GET"])
+    app.add_api_route("/media/{channel}/{post_id}/{file_unique_id}", api_server.get_media, methods=["GET"])
+    return app
+
+
+def _seed_cache(tmp_path, channel, post_id, fid, body=BODY):
+    cache_dir = tmp_path / "data" / "cache" / str(channel) / str(post_id)
+    cache_dir.mkdir(parents=True, exist_ok=True)
+    (cache_dir / fid).write_bytes(body)
+    return cache_dir / fid
+
+
+# --------------------------------------------------------------------------- #
+# Range semantics at the /media ROUTE level (stages 2 + 3 wired together).
+# Plan scenario: `curl -H "Range: bytes=0-99" / "bytes=-100" / "bytes=999999999-"`.
+# Regression caught: any change that makes get_media's cache-hit branch stop reaching
+# FileResponse (e.g. re-buffering the body, hand-rolling headers, dropping the media_key
+# path) or that breaks the digest gate wiring — the Range would stop being honored.
+# --------------------------------------------------------------------------- #
+def test_media_route_range_prefix_0_99(monkeypatch, tmp_path):
+    monkeypatch.chdir(tmp_path)
+    _seed_cache(tmp_path, "chan", 3, "fidR")
+    monkeypatch.setattr(api_server, "verify_media_digest", lambda url, digest: True)
+    # Keep the MIME path DB-free; the FileResponse/Range machinery is what we exercise.
+    monkeypatch.setattr(api_server, "get_mime_type_sync", lambda *a, **k: None)
+    monkeypatch.setattr(api_server, "set_mime_type_sync", lambda *a, **k: None)
+
+    c = TestClient(_media_app())
+    r = c.get("/media/chan/3/fidR/anydigest", headers={"Range": "bytes=0-99"})
+    assert r.status_code == 206
+    assert r.headers["content-range"] == f"bytes 0-99/{SIZE}"
+    assert r.headers["content-length"] == "100"
+    assert r.content == BODY[:100]
+    assert r.headers["accept-ranges"] == "bytes"
+
+
+def test_media_route_range_suffix_last_100(monkeypatch, tmp_path):
+    monkeypatch.chdir(tmp_path)
+    _seed_cache(tmp_path, "chan", 3, "fidS")
+    monkeypatch.setattr(api_server, "verify_media_digest", lambda url, digest: True)
+    monkeypatch.setattr(api_server, "get_mime_type_sync", lambda *a, **k: None)
+    monkeypatch.setattr(api_server, "set_mime_type_sync", lambda *a, **k: None)
+
+    c = TestClient(_media_app())
+    r = c.get("/media/chan/3/fidS/anydigest", headers={"Range": "bytes=-100"})
+    assert r.status_code == 206
+    assert r.headers["content-range"] == f"bytes {SIZE - 100}-{SIZE - 1}/{SIZE}"
+    assert r.headers["content-length"] == "100"
+    assert r.content == BODY[-100:]
+
+
+def test_media_route_range_unsatisfiable_416(monkeypatch, tmp_path):
+    monkeypatch.chdir(tmp_path)
+    _seed_cache(tmp_path, "chan", 3, "fidU")
+    monkeypatch.setattr(api_server, "verify_media_digest", lambda url, digest: True)
+    monkeypatch.setattr(api_server, "get_mime_type_sync", lambda *a, **k: None)
+    monkeypatch.setattr(api_server, "set_mime_type_sync", lambda *a, **k: None)
+
+    c = TestClient(_media_app())
+    r = c.get("/media/chan/3/fidU/anydigest", headers={"Range": "bytes=999999999-"})
+    assert r.status_code == 416
+    # Starlette's 416 Content-Range is `*/size` (documented stage-3 RFC-7233 delta).
+    assert r.headers["content-range"] == f"*/{SIZE}"
+
+
+# --------------------------------------------------------------------------- #
+# /ping (stage 6) stays fast + correct while a slow op is in flight, zero TG RPC.
+# Plan scenario: "Генерация фида на 100+ сообщений + параллельный /ping -> ping < 100 мс".
+# We model the concurrent slow op as a coroutine parked on an Event that is NEVER set
+# during the ping, and assert ping resolves while it is still pending AND touches no RPC.
+# Regression caught: re-coupling /ping to a TG RPC (get_me / safe_get_messages) or to any
+# awaitable that a blocked hot path could stall — the ping would no longer return promptly
+# or the spy counts would go non-zero.
+# --------------------------------------------------------------------------- #
+class _FakeKurigram:
+    def __init__(self, is_connected=True):
+        self.is_connected = is_connected
+        self.get_me_calls = 0
+
+    async def get_me(self):
+        self.get_me_calls += 1
+        raise AssertionError("/ping must never call get_me()")
+
+
+class _FakeTelegramClient:
+    def __init__(self, age, is_connected=True):
+        self._age = age
+        self.client = _FakeKurigram(is_connected=is_connected)
+        self.safe_get_messages_calls = 0
+
+    def watchdog_last_ok_age(self):
+        return self._age
+
+    async def safe_get_messages(self, *a, **k):
+        self.safe_get_messages_calls += 1
+        raise AssertionError("/ping must never call safe_get_messages()")
+
+
+async def test_ping_prompt_and_rpc_free_while_slow_op_pending(monkeypatch):
+    threshold = api_server.Config["tg_ping_unhealthy_after"]
+    fake = _FakeTelegramClient(age=threshold - 10, is_connected=True)
+    monkeypatch.setattr(api_server, "client", fake)
+
+    # A concurrent slow operation (a stand-in for a hung feed/RPC hot path) parked on an
+    # Event we deliberately never set for the duration of the ping.
+    gate = asyncio.Event()
+
+    async def slow_op():
+        await gate.wait()
+
+    slow = asyncio.create_task(slow_op())
+    await asyncio.sleep(0)  # let slow_op start and park on the gate
+
+    # The real proof of decoupling: ping() returns under a tight deadline while the slow op
+    # is parked, AND issues zero TG RPC. wait_for reds if ping ever blocks; the spies (which
+    # raise if touched) red if ping recouples to any RPC. `not slow.done()` is only a sanity
+    # check that ping did not somehow drive the parked op — the gate keeps it pending anyway.
+    resp = await asyncio.wait_for(api_server.ping(), timeout=1.0)
+
+    assert resp.status_code == 200            # correct health while connected + fresh
+    assert not slow.done()                    # sanity: slow op still parked, ping did not await it
+    assert fake.client.get_me_calls == 0      # decoupled: zero TG RPC
+    assert fake.safe_get_messages_calls == 0
+
+    gate.set()
+    await slow
+
+
+async def test_ping_reports_degraded_promptly_while_slow_op_pending(monkeypatch):
+    threshold = api_server.Config["tg_ping_unhealthy_after"]
+    fake = _FakeTelegramClient(age=threshold + 100, is_connected=True)  # stale probe
+    monkeypatch.setattr(api_server, "client", fake)
+
+    gate = asyncio.Event()
+
+    async def slow_op():
+        await gate.wait()
+
+    slow = asyncio.create_task(slow_op())
+    await asyncio.sleep(0)
+
+    resp = await asyncio.wait_for(api_server.ping(), timeout=1.0)
+
+    assert resp.status_code == 503            # stale watchdog probe -> degraded, still instant
+    assert not slow.done()                    # sanity: slow op still parked, ping did not await it
+    assert fake.client.get_me_calls == 0
+    assert fake.safe_get_messages_calls == 0
+
+    gate.set()
+    await slow
+
+
+# --------------------------------------------------------------------------- #
+# In-flight dedup + disconnect cleanup (stages 1/2) through the REAL get_media entry.
+# Plan scenario: "Параллельные запросы одного большого видео -> нет частичной отдачи" and
+# "Отключение клиента на середине стрима -> нет утечки тасков/фд".
+# The pure-unit stage-2 tests pin _download_deduped directly; these drive get_media so the
+# HTTP semaphore + dedup registry + serve path are proven wired together.
+# Regression caught: moving the download back into the request coroutine (so a disconnect
+# cancels it), or dropping the finally-pop, would leave a stuck _inflight key / hung sibling.
+# --------------------------------------------------------------------------- #
+async def test_get_media_concurrent_shares_one_download_and_drains_registry(monkeypatch, tmp_path):
+    monkeypatch.chdir(tmp_path)
+    api_server._inflight.clear()
+    monkeypatch.setattr(api_server, "verify_media_digest", lambda url, digest: True)
+    # Serve step is not under test here; keep it to a sentinel so we assert on dedup + registry.
+    sentinel = object()
+
+    async def fake_prepare(file_path, request=None, delete_after=False, media_key=None):
+        return sentinel
+
+    monkeypatch.setattr(api_server, "prepare_file_response", fake_prepare)
+
+    calls = []
+
+    async def slow_dl(channel, post_id, fid):
+        calls.append(fid)
+        await asyncio.sleep(0.05)  # real overlap window for the two requests
+        return (f"/final/{fid}", False)
+
+    monkeypatch.setattr(api_server, "download_media_file", slow_dl)
+
+    req = SimpleNamespace(headers={})
+    t1 = asyncio.create_task(api_server.get_media("chan", 9, "vfid", request=req, digest="x"))
+    t2 = asyncio.create_task(api_server.get_media("chan", 9, "vfid", request=req, digest="x"))
+    r1, r2 = await asyncio.gather(t1, t2)
+
+    assert r1 is sentinel and r2 is sentinel
+    assert calls == ["vfid"]            # exactly ONE real download, shared by both requests
+    assert not api_server._inflight     # registry drained — no forever-busy key
+
+
+async def test_get_media_request_cancel_does_not_stick_registry_or_hang_sibling(monkeypatch, tmp_path):
+    monkeypatch.chdir(tmp_path)
+    api_server._inflight.clear()
+    monkeypatch.setattr(api_server, "verify_media_digest", lambda url, digest: True)
+    sentinel = object()
+
+    async def fake_prepare(file_path, request=None, delete_after=False, media_key=None):
+        return sentinel
+
+    monkeypatch.setattr(api_server, "prepare_file_response", fake_prepare)
+
+    gate = asyncio.Event()
+    calls = []
+
+    async def held_dl(channel, post_id, fid):
+        calls.append(fid)
+        await gate.wait()  # hold the shared download open until we release it
+        return (f"/final/{fid}", False)
+
+    monkeypatch.setattr(api_server, "download_media_file", held_dl)
+
+    req = SimpleNamespace(headers={})
+    t1 = asyncio.create_task(api_server.get_media("chan", 9, "cfid", request=req, digest="x"))
+    await asyncio.sleep(0.02)  # t1 registers the future + starts the detached download
+    t2 = asyncio.create_task(api_server.get_media("chan", 9, "cfid", request=req, digest="x"))
+    await asyncio.sleep(0.02)
+
+    # First client disconnects: its request coroutine is cancelled mid-wait.
+    t1.cancel()
+    with pytest.raises(asyncio.CancelledError):
+        await t1
+
+    # The detached download is unaffected; releasing it resolves the surviving request.
+    gate.set()
+    r2 = await asyncio.wait_for(t2, timeout=2.0)
+    assert r2 is sentinel
+    assert calls == ["cfid"]           # download ran exactly once (not restarted)
+    assert not api_server._inflight     # registry drained despite the disconnect
+
+
+# --------------------------------------------------------------------------- #
+# str(channel) access-time key consistency (stage 5) END-TO-END: hit -> flush -> DB.
+# Plan gotcha #9: "Ключи SQLite: channel всегда str(...)"; if the key the hot path RECORDS
+# and the key the flush UPDATEs by ever diverge, the timestamp silently stops updating and
+# the file eventually falls out of the cache. Stage 5 pins hit and flush separately; this
+# stitches them: the /media cache-hit records a (str-channel) key, and the flush's bulk
+# UPDATE must land on that exact DB row. (get_media's route channel is already a str, so the
+# str() there is a defensive no-op; the live int/str hazard is in download_media_file, pinned
+# by stage-5's test_cache_hit_keys_channel_as_str — this test covers the get_media+flush leg.)
+# Regression caught: any accumulator-key vs UPDATE-WHERE inconsistency (e.g. a transposed
+# key-column order in the bulk SQL, verified to turn this test red) leaves `added` stale.
+# --------------------------------------------------------------------------- #
+async def test_media_cache_hit_flush_updates_str_keyed_db_row(monkeypatch, tmp_path):
+    monkeypatch.chdir(tmp_path)
+    api_server._access_updates = {}
+
+    channel, post_id, fid = "12345", 7, "fidINT"  # int-ish channel, passed as the route str
+    _seed_cache(tmp_path, channel, post_id, fid)
+
+    db = str(tmp_path / "access.db")
+    init_db_sync(db)
+    # Seed the row keyed by the STRING channel with an old access time.
+    conn = sqlite3.connect(db)
+    conn.execute(
+        "INSERT INTO media_file_ids (channel, post_id, file_unique_id, added) VALUES (?,?,?,?)",
+        (channel, post_id, fid, 1.0),
+    )
+    conn.commit()
+    conn.close()
+    monkeypatch.setattr(api_server, "DB_PATH", db)
+
+    monkeypatch.setattr(api_server, "verify_media_digest", lambda url, digest: True)
+    sentinel = object()
+
+    async def fake_prepare(file_path, request=None, media_key=None):
+        return sentinel
+
+    monkeypatch.setattr(api_server, "prepare_file_response", fake_prepare)
+
+    before = time.time()
+    resp = await api_server.get_media(channel, post_id, fid, request=SimpleNamespace(headers={}), digest="x")
+    assert resp is sentinel
+
+    # Hot path recorded the str-keyed access time (never the raw int form).
+    assert (channel, post_id, fid) in api_server._access_updates
+
+    # Flush the accumulator; the bulk UPDATE must match the str-keyed row and refresh `added`.
+    await api_server._flush_access_updates()
+    assert api_server._access_updates == {}
+
+    conn = sqlite3.connect(db)
+    added = conn.execute(
+        "SELECT added FROM media_file_ids WHERE channel = ? AND post_id = ? AND file_unique_id = ?",
+        (channel, post_id, fid),
+    ).fetchone()[0]
+    conn.close()
+    assert added >= before  # refreshed from the stale 1.0 to ~now via the str-keyed WHERE

tg_cache.py

@@ -0,0 +1,230 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+# flake8: noqa
+# pylint: disable=broad-exception-raised, raise-missing-from, too-many-arguments, redefined-outer-name
+# pylint: disable=multiple-statements, logging-fstring-interpolation, trailing-whitespace, line-too-long
+# pylint: disable=broad-exception-caught, missing-function-docstring, missing-class-docstring
+# pylint: disable=f-string-without-interpolation
+# pylance: disable=reportMissingImports, reportMissingModuleSource
+
+
+import os
+import json
+import pickle
+import logging
+import random
+import asyncio
+import time
+from datetime import datetime, timedelta
+from types import SimpleNamespace
+from typing import Any, Optional, Union, List
+from pyrogram import Client
+from pyrogram.types import Chat, Message
+from tg_throttle import tg_rpc_bounded
+from config import get_settings
+
+logger = logging.getLogger(__name__)
+
+Config = get_settings()
+
+# Path to cache directory
+CACHE_DIR = os.path.join('data', 'tgcache')
+
+# TTL (hours) for the channel-info cache. Channel title/username/id change rarely,
+# so a long TTL removes the rate-limited GetFullChannel from the feed-poll hot path.
+try:
+    CHAT_CACHE_TTL_HOURS = int(os.getenv("TG_CHAT_CACHE_TTL_HOURS", "12"))
+    if CHAT_CACHE_TTL_HOURS < 1:
+        CHAT_CACHE_TTL_HOURS = 12
+except ValueError:
+    CHAT_CACHE_TTL_HOURS = 12
+
+def _get_history_cache_file_path(channel_id: Union[str, int]) -> str:
+    """Returns path to the message history cache file for the channel"""
+    if not os.path.exists(CACHE_DIR):
+        os.makedirs(CACHE_DIR, exist_ok=True)
+        logger.info(f"cache_dir_created: path {CACHE_DIR}")
+    # Convert to string for uniformity
+    channel_id_str = str(channel_id)
+    # Replace potentially problematic characters
+    safe_filename = channel_id_str.replace('/', '_').replace('\\', '_')
+    return os.path.join(CACHE_DIR, f"{safe_filename}.cache")
+
+def _save_history_to_cache(channel_id: Union[str, int], messages: List[Message], limit: int) -> None:
+    """Saves message history to cache"""
+    try:
+        cache_file = _get_history_cache_file_path(channel_id)
+        
+        # Create cache metadata — store messages directly (no inner pickle.dumps)
+        cache_data = {
+            'timestamp': time.time(),
+            'limit': limit,
+            'messages': messages
+        }
+        
+        with open(cache_file, 'wb') as f:
+            pickle.dump(cache_data, f)
+        
+        logger.info(f"history_cache_saved: channel {channel_id}, limit {limit}, messages {len(messages)}, file {cache_file}")
+    except Exception as e:
+        logger.error(f"history_cache_save_error: channel {channel_id}, limit {limit}, error {str(e)}")
+
+def _get_history_from_cache(channel_id: Union[str, int], limit: int, max_age_hours: int = 8) -> Optional[List[Message]]:
+    """
+    Retrieves message history from cache if not older than specified age and matches the limit
+    
+    Args:
+        channel_id: Channel ID or username
+        limit: Required message limit
+        max_age_hours: Maximum cache age in hours (default 8 hours)
+        
+    Returns:
+        List of messages or None if cache not found, expired or limit doesn't match
+    """
+    try:
+        cache_file = _get_history_cache_file_path(channel_id)
+        
+        if not os.path.exists(cache_file):
+            logger.info(f"history_cache_miss: channel {channel_id}, limit {limit}, cache file not found")
+            return None
+        
+        with open(cache_file, 'rb') as f:
+            cache_data = pickle.load(f)
+        
+        # Check cache age with randomization
+        cache_age = time.time() - cache_data['timestamp']
+        # Add randomness up to 20% of max_age_seconds
+        random_factor = 1 - random.uniform(0, 0.2)
+        adjusted_max_age = max_age_hours * 3600 * random_factor
+        
+        if cache_age > adjusted_max_age:
+            logger.info(f"history_cache_expired: channel {channel_id}, limit {limit}, age {cache_age:.1f}s > adjusted max {adjusted_max_age:.1f}s (random factor: {random_factor:.2f})")
+            return None
+        
+        # Check if limit matches
+        cached_limit = cache_data.get('limit', 0)
+        if cached_limit != limit:
+            logger.info(f"history_cache_limit_mismatch: channel {channel_id}, cached limit {cached_limit}, requested limit {limit}")
+            return None
+        
+        # Restore message list; handle old cache files that used double-pickle (bytes = old format)
+        raw = cache_data['messages']
+        if isinstance(raw, bytes):
+            messages = pickle.loads(raw)
+        else:
+            messages = raw
+        logger.info(f"history_cache_hit: channel {channel_id}, limit {limit}, messages {len(messages)}, age {cache_age:.1f}s")
+        return messages
+    
+    except Exception as e:
+        logger.error(f"history_cache_read_error: channel {channel_id}, limit {limit}, error {str(e)}")
+        # In case of cache read error, better return None and request fresh data
+        return None
+
+async def cached_get_chat_history(client: Client, channel_id: Union[str, int], limit: int = 20) -> List[Message]:
+    """
+    Gets chat message history with caching.
+    
+    Args:
+        client: Pyrogram client
+        channel_id: Channel ID or username
+        limit: Maximum number of messages to retrieve
+        
+    Returns:
+        List of messages, same as original client.get_chat_history()
+    """
+    cached_messages = await asyncio.to_thread(_get_history_from_cache, channel_id, limit)
+    
+    if cached_messages is not None:
+        return cached_messages
+    
+    try:
+        logger.info(f"history_cache_request: fetching fresh history for channel {channel_id}, limit {limit}")
+        # Hold the global RPC gate for the live fetch and bound the RPC body with the
+        # timeout — gate outside, timeout inside — via the shared tg_rpc_bounded (so the
+        # tricky nesting is not re-derived here). The timeout covers the whole paginated
+        # fetch; see the note in tg_rpc_bounded.
+        async with tg_rpc_bounded(Config["tg_rpc_timeout"]):
+            messages = [m async for m in client.get_chat_history(channel_id, limit=limit)]
+        await asyncio.to_thread(_save_history_to_cache, channel_id, messages, limit)
+        
+        return messages
+    except Exception as e:
+        logger.error(f"history_cache_request_error: channel {channel_id}, limit {limit}, error {str(e)}")
+        raise
+
+
+def _get_chat_cache_file_path(channel_id: Union[str, int]) -> str:
+    """Returns path to the channel-info cache file for the channel."""
+    if not os.path.exists(CACHE_DIR):
+        os.makedirs(CACHE_DIR, exist_ok=True)
+        logger.info(f"cache_dir_created: path {CACHE_DIR}")
+    channel_id_str = str(channel_id)
+    safe_filename = channel_id_str.replace('/', '_').replace('\\', '_')
+    # Distinct suffix so channel-info files never collide with history (.cache) files.
+    return os.path.join(CACHE_DIR, f"{safe_filename}.chatinfo")
+
+
+def _save_chat_to_cache(channel_id: Union[str, int], data: dict) -> None:
+    """Saves channel-info (id/title/username) to cache."""
+    try:
+        cache_file = _get_chat_cache_file_path(channel_id)
+        cache_data = {'timestamp': time.time(), 'data': data}
+        with open(cache_file, 'wb') as f:
+            pickle.dump(cache_data, f)
+        logger.info(f"chatinfo_cache_saved: channel {channel_id}, file {cache_file}")
+    except Exception as e:
+        logger.error(f"chatinfo_cache_save_error: channel {channel_id}, error {str(e)}")
+
+
+def _get_chat_from_cache(channel_id: Union[str, int], max_age_hours: int = CHAT_CACHE_TTL_HOURS) -> Optional[dict]:
+    """Retrieves channel-info from cache if not older than max_age_hours (with up-to-20% randomization)."""
+    try:
+        cache_file = _get_chat_cache_file_path(channel_id)
+        if not os.path.exists(cache_file):
+            logger.info(f"chatinfo_cache_miss: channel {channel_id}, cache file not found")
+            return None
+        with open(cache_file, 'rb') as f:
+            cache_data = pickle.load(f)
+        cache_age = time.time() - cache_data['timestamp']
+        # Add randomness up to 20% of max age, same approach as the history cache.
+        random_factor = 1 - random.uniform(0, 0.2)
+        adjusted_max_age = max_age_hours * 3600 * random_factor
+        if cache_age > adjusted_max_age:
+            logger.info(f"chatinfo_cache_expired: channel {channel_id}, age {cache_age:.1f}s > adjusted max {adjusted_max_age:.1f}s (random factor: {random_factor:.2f})")
+            return None
+        data = cache_data.get('data')
+        if not isinstance(data, dict):
+            logger.warning(f"chatinfo_cache_invalid: channel {channel_id}, unexpected payload type {type(data).__name__}")
+            return None
+        logger.info(f"chatinfo_cache_hit: channel {channel_id}, age {cache_age:.1f}s")
+        return data
+    except Exception as e:
+        logger.error(f"chatinfo_cache_read_error: channel {channel_id}, error {str(e)}")
+        return None
+
+
+async def cached_get_chat(client: Client, channel_id: Union[str, int]) -> SimpleNamespace:
+    """Gets channel info (id/title/username) with disk TTL caching.
+
+    Avoids the rate-limited channels.GetFullChannel on every feed poll. Mirrors
+    cached_get_chat_history. On a cache miss the live get_chat is throttled and its
+    exceptions (FloodWait, UsernameInvalid, ...) propagate to the caller unchanged;
+    only successful lookups are cached. The returned object exposes .id/.title/.username.
+    """
+    cached = await asyncio.to_thread(_get_chat_from_cache, channel_id)
+    if cached is not None:
+        return SimpleNamespace(**cached)
+
+    logger.info(f"chatinfo_cache_request: fetching fresh chat info for channel {channel_id}")
+    async with tg_rpc_bounded(Config["tg_rpc_timeout"]):
+        chat = await client.get_chat(channel_id)
+
+    data = {
+        'id': getattr(chat, 'id', None),
+        'title': getattr(chat, 'title', None),
+        'username': getattr(chat, 'username', None),
+    }
+    await asyncio.to_thread(_save_chat_to_cache, channel_id, data)
+    return SimpleNamespace(**data)

tg_throttle.py

@@ -0,0 +1,92 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+# flake8: noqa
+# pylint: disable=broad-exception-caught, global-statement, missing-function-docstring, missing-class-docstring
+# pylint: disable=logging-fstring-interpolation, line-too-long
+
+import os
+import time
+import asyncio
+import logging
+from contextlib import asynccontextmanager
+
+logger = logging.getLogger(__name__)
+
+
+def _parse_int_env(name: str, default: int, minimum: int) -> int:
+    """Parse an int env var, falling back to default on missing/invalid/out-of-range value."""
+    raw = os.getenv(name)
+    if raw is None or raw.strip() == "":
+        return default
+    try:
+        value = int(raw)
+    except ValueError:
+        logger.warning(f"tg_throttle: {name} is not a valid integer ({raw!r}); using default {default}")
+        return default
+    if value < minimum:
+        logger.warning(f"tg_throttle: {name}={value} below minimum {minimum}; using default {default}")
+        return default
+    return value
+
+
+# Global throttle for live Telegram MTProto RPC calls. Serializes bursts (e.g. miniflux
+# batching ~47 feeds at once) so they do not trip Telegram's FLOOD_WAIT (420).
+_CONCURRENCY = _parse_int_env("TG_RPC_CONCURRENCY", 1, 1)               # max concurrent Telegram RPCs
+_MIN_INTERVAL = _parse_int_env("TG_RPC_MIN_INTERVAL_MS", 500, 0) / 1000.0  # min gap between RPC starts (seconds)
+
+_sem = asyncio.Semaphore(_CONCURRENCY)
+_lock = asyncio.Lock()
+_last_start = 0.0
+
+logger.info(f"tg_throttle: initialized (concurrency={_CONCURRENCY}, min_interval={_MIN_INTERVAL*1000:.0f}ms)")
+
+
+class _TgRpcGate:
+    """Async context manager that caps concurrency and enforces a minimum spacing between RPC starts."""
+
+    async def __aenter__(self):
+        await _sem.acquire()
+        global _last_start
+        try:
+            async with _lock:
+                wait = _last_start + _MIN_INTERVAL - time.monotonic()
+                if wait > 0:
+                    await asyncio.sleep(wait)
+                _last_start = time.monotonic()
+        except BaseException:
+            # Do not leak a semaphore permit if cancelled while waiting for the spacing.
+            _sem.release()
+            raise
+        return self
+
+    async def __aexit__(self, exc_type, exc, tb):
+        _sem.release()
+        return False
+
+
+def tg_rpc():
+    """Return an async context manager that throttles a single live Telegram RPC call."""
+    return _TgRpcGate()
+
+
+@asynccontextmanager
+async def tg_rpc_bounded(timeout: float):
+    """Throttle a live Telegram RPC AND bound it with a timeout, correctly nested.
+
+    The single tricky invariant this centralizes: the timeout must bound ONLY the
+    RPC body, never the gate ENTRY — timing out the `_sem.acquire()` / spacing wait
+    would turn legitimate queue backpressure (e.g. ~47 feeds queueing) into false
+    timeouts. So the gate is the OUTER context and the timeout is the INNER one; a
+    TimeoutError raised inside propagates out through the gate's `__aexit__`, which
+    releases the permit (no leak). Call as:
+
+        async with tg_rpc_bounded(Config["tg_rpc_timeout"]):
+            result = await client.get_chat(channel_id)
+
+    Every gated+bounded RPC uses this so no call site re-derives the nesting by hand
+    (getting it wrong silently reopens the hang-under-backpressure class).
+    """
+    async with _TgRpcGate():
+        async with asyncio.timeout(timeout):
+            yield