fast, readable log: stop routine noise, keyset paging, token search

The log had 6.45M rows, of which 434 were actual actions: every mail x
every non-matching rule wrote "Keine Übereinstimmung" (5.3M rows, ~1 GB),
plus "keine neuen Mails" per rule every 30s and a "Poll gestartet" row.
Free-text searches took 50-70s.

- Routine entries are only written with LOG_VERBOSE=true; the poll
  summary is logged only when there were new mails, actions or errors.
  "Why didn't my rule match" is now answered by the rule tester.
- The scheduler commits every 25 evaluated mails and after each rule,
  since the dropped log calls used to be the implicit commit points.
- Migration rebuilds filter_logs without the routine rows (live:
  6.45M -> ~38k) via copy instead of DELETE.
- /api/logs: keyset pagination (before_id) instead of OFFSET, count
  capped at 1000, since/until range, multi-level filter, per-field
  filters; ordner: also matches the move target stored in details.
- New log page: one search field with shortcuts (von:, betreff:,
  regel:, ordner:, text:), live search with debounce and abort, active
  filters as removable chips, click on rule/sender/folder in a row to
  filter, time range (default 7 days), level quick filter, expandable
  row details, "Ältere laden", filters kept in the URL.
- Disable Pico's colour transition on segmented buttons (they briefly
  flashed in the primary colour on load).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

app/config.py

@@ -9,6 +9,9 @@ class Settings(BaseSettings):
     encryption_key: str = ""
     yaml_sync_on_startup: bool = True
     log_level: str = "INFO"
+    # Ausführliches Verarbeitungslog (jede Nicht-Übereinstimmung, jede Regel pro
+    # Poll). Erzeugt Millionen Einträge — nur zur Fehlersuche einschalten.
+    log_verbose: bool = False
 
     model_config = {"env_file": ".env", "env_file_encoding": "utf-8"}