fix: completeness now accounts for missed files

Previously, completeness was calculated only across matched files
(files both the tool and human touched), ignoring missed files entirely.
This gave inflated scores -- e.g., 97.2% when only 4 out of 31 Go files
were matched.

Now, every missed Go file in the ground truth contributes a 0% score to
the average. If the tool matches 4 files at 97% but misses 10 files,
completeness = (4*97% + 10*0%) / 14 = 27.7%, not 97%.

This gives an honest picture of how much of the total implementation
the tool actually covered.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: rausingh-rh

benchmark/compare.py

@@ -403,12 +403,13 @@ def compare_iteration(
                         ),
                     ))
 
+    missed_go_files = go_truth_files - go_gen_files
+    for _missed in missed_go_files:
+        completeness_scores.append(0.0)
+
     completeness = statistics.mean(completeness_scores) * 100 if completeness_scores else 0.0
     convention = statistics.mean(convention_scores) * 100 if convention_scores else 0.0
 
-    if not matched_files and go_truth_files:
-        completeness = 0.0
-
     score = IterationScore(
         iteration=gen_result.iteration,
         completeness=round(completeness, 1),

benchmark/report.py

@@ -63,7 +63,7 @@ def generate_ep_report(
     lines.append("")
     lines.append("**Metric definitions:**")
     lines.append("")
-    lines.append("- **Completeness**: What % of the human's structs, fields, and functions did the tool also generate? Higher = tool covered more of the ground truth.")
+    lines.append("- **Completeness**: What % of the human's Go implementation did the tool cover? Accounts for ALL Go files in the ground truth -- matched files are scored by struct/field/function overlap, missed files count as 0%.")
     lines.append("- **Convention**: What % of kubebuilder markers (`+kubebuilder:validation:Required`, `+optional`, etc.) match between generated and ground truth?")
     lines.append("- **Build**: Did `make build` pass on the generated code?")
     lines.append("- **Matched**: Files the tool generated that also exist in the human implementation (true positives).")