Respond to code review feedback

This makes two slight changes after discussing code review feedback:

1. We move the assignment of the @Detections instance variable down
   below the first guard clause in the Detector::MlCitation initializer.
   This better accounts for the fact that the Term model implements its
   own guard clause, resulting in the rest of the application seeing a
   detector result of null if the detector is not enabled via env vars.

2. Adds a note to the enough_nonzero_values? method comment explaining
   how we came to decide that three non-zero feature values is
   sufficient. This includes links to the notebooks where we conducted
   that analysis.

Author: matt-bernhardt

app/models/detector/ml_citation.rb

@@ -9,9 +9,10 @@ class MlCitation
     #   @param phrase String. Often a `Term.phrase`.
     #   @return Nothing intentional. Data is written to Boolean `@detections` during processing.
     def initialize(phrase)
-      @detections = false
       return unless self.class.expected_env?
 
+      @detections = false
+
       features = extract_features(phrase)
       return unless enough_nonzero_values?(features)
 
@@ -158,6 +159,11 @@ def fetch(features)
 
     # Enough_nonzero_values? checks that a provided hash contains at least three values which are not zero.
     #
+    # @note We chose 3 as our value here after analyzing the behavior of the citation detector across nearly a year of
+    #   search traffic. For searches which had only one or two features that are not zero, we found no actual citations.
+    #   To see the analyses, look at the "Filtering results" and "Surprising predictions" notebooks at
+    #   https://github.com/MITLibraries/tacos-notebooks/tree/main/notebooks/explorations
+    #
     # @param hash Hash
     # @return Integer
     def enough_nonzero_values?(hash)