feat(iace): 25 new robot cell patterns (HP1650-HP1697) + matcher fix

New patterns from GT benchmark gap analysis:
- HP1650-1655: Robot arm motion limit, restart safety, tool/workpiece
  crushing, workpiece penetrates fence, reaching over fence
- HP1660-1661: Centering gripper crushing (outside/inside cell)
- HP1665-1666: Machine tool loading door, machining workspace
- HP1670-1671: Coolant splash eyes, compressed air injury
- HP1675: Coolant hose burst/detachment
- HP1680: Workpiece/tunnel crushing at conveyor
- HP1685-1689: Indirect contact, cabinet contact, liquid ingress fire,
  potential differences, RCD socket protection
- HP1690-1691: Ergonomic loading/control position
- HP1695: Burns from hot workpieces
- HP1697: Machine collapse through floor

Matcher: keyword overlap penalty — matches without shared hazard-type
keywords AND low zone score get 0.5x penalty to prevent false matches.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

ai-compliance-sdk/internal/iace/benchmark_matcher.go

@@ -217,10 +217,17 @@ func fuzzyMatchScore(gt *GroundTruthEntry, h *Hazard) (float64, string) {
 	// Penalty: if engine hazard mentions a machine-specific term not in the GT context,
 	// it's likely a wrong-machine match (e.g. "Spielplatz" for a robot cell GT entry)
 	if hasWrongMachineTerm(h.Name, h.Scenario, gt.HazardCause, gt.ComponentZone) {
-		score *= 0.3 // Heavy penalty
+		score *= 0.3
 		reasons = append(reasons, "Strafabzug:FremdMaschine")
 	}
 
+	// Minimum keyword overlap required: if GT and Engine share no hazard-type
+	// keywords at all, the match is unreliable regardless of category/zone score
+	if kwScore == 0 && zoneScore < 0.5 {
+		score *= 0.5
+		reasons = append(reasons, "Strafabzug:KeineKeywords")
+	}
+
 	return score, strings.Join(reasons, "+")
 }