Since rEABFbc9898f1a13a342ca44e928cee353ab5bf32e12e, the following is in AbuseFilterRunner::checkEmergencyDisable:

$filterProfile = $this->filterProfiler->getFilterProfile( $filter );
$matchCount = ( $filterProfile['matches'] ?? 0 ) + 1;

However, FilterProfiler::getFilterProfile returns stats in a different format:

public function getFilterProfile( $filter ) : array {
  $profile = $this->objectStash->get( $this->filterProfileKey( $filter ) );
  if ( $profile === false ) {
    return [ 0, 0, 0.0, 0.0 ];
  }
  $curCount = $profile['count'];
  $curTotalTime = $profile['total-time'];
  $curTotalConds = $profile['total-cond'];
  // Return in milliseconds, rounded to 2dp
  $avgTime = round( $curTotalTime / $curCount, 2 );
  $avgCond = round( $curTotalConds / $curCount, 1 );
  return [ $curCount, $profile['matches'], $avgTime, $avgCond ];
}

This probably broke EmergencyDisable feature. Since $matchCount is always 1 and AbuseFilterEmergencyDisableCount is (by default) 2, it would never throttle a filter. (Which maybe is better than every changed filter.)

Hotfix: change $filterProfile['matches'] to $filterProfile[1].
Further changes:

• Make FilterProfiler::getFilterProfile return stats unchanged and move its logic to AbuseFilterViewEdit (where all the dividing and rounding can be done). Now it's inconsistent with its sibling getGroupProfile method and this probably was the source of confusion behind the regression.
• EmergencyWatcher service & better test coverage (work in progress).