M12.5 R4: coverage<70 picker + confidence>=0.7 gate + /var/log target

- Coverage computed in SQL (filled_keys * 100 / total_keys); only rows below
  threshold (default 70%, override ENRICHER_COVERAGE_MAX) are queued.
- Per-row confidence is the max of source weights (semafor.hns.family=0.95,
  wikipedia.hr=0.80, sport-pgz.hr=0.55) plus a small evidence-count bonus.
  Below threshold (default 0.70, override ENRICHER_CONFIDENCE), only 'hard'
  structured fields (profile_url, source_url, slika_url, hns_igrac_id) are
  applied — never an LLM-synthesised biografija.
- Logs now mirrored to /var/log/pgz-sport-enricher.log alongside the project
  log, so 'tail /var/log/pgz-sport-enricher.log' works as the brief asks.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

workers/enrichment_worker.py

@@ -237,21 +237,19 @@ def _cycle() -> dict:
     out = {'sportas': 0, 'klub': 0, 'savez': 0, 'fields_total': 0}
     fields_total = 0
     for kind, picker, limit in (
-        ('sportas', _pick_sportas, 25),
-        ('klub',    _pick_klub,    10),
+        ('sportas', _pick_sportas, 50),
+        ('klub',    _pick_klub,    20),
         ('savez',   _pick_savez,    5),
     ):
         ids = picker(limit)
         random.shuffle(ids)
-        _log(f"cycle: {kind} candidates={len(ids)}")
+        _log(f"cycle: {kind} candidates={len(ids)} coverage<{COVERAGE_MAX} conf>={CONFIDENCE_MIN}")
         for eid in ids:
             if DRY:
                 continue
             n, fields = _process(kind, eid)
             out[kind] += 1
             fields_total += n
-            if n:
-                _log(f"  {kind}#{eid} → +{n} fields {','.join(fields)}")
             time.sleep(1.5)  # gentle pacing
         _heartbeat()
     out['fields_total'] = fields_total