This is the most significant measurement error in the toolkit. additions / max(deletions, 1) for a commit of entirely new files yields a ratio approaching infinity, which explodes both the median and p90. The RETROSPECTIVE noted p90=446x is "almost certainly dominated by these net-new-file commits" but leaves it as documentation debt.
Proposed fix: Compute a second ratio alongside the current one:
net_additions_ratio = (additions - deletions) / total_lines_changed
This is bounded 0–1, immune to the new-file distortion, and more directly measures "what percentage of this commit is truly new code vs replacement." A ratio consistently above 0.8 means 80% of every commit is net-new code — the actual batch-acceptance signal.
Alternatively, exclude commits where prod_files_count > 0 AND total_deletions < 5 from the ratio distribution (they're new-file bootstraps, not drift signals), and document this filtering in the output.
This is the most significant measurement error in the toolkit. additions / max(deletions, 1) for a commit of entirely new files yields a ratio approaching infinity, which explodes both the median and p90. The RETROSPECTIVE noted p90=446x is "almost certainly dominated by these net-new-file commits" but leaves it as documentation debt.
Proposed fix: Compute a second ratio alongside the current one:
net_additions_ratio = (additions - deletions) / total_lines_changed
This is bounded 0–1, immune to the new-file distortion, and more directly measures "what percentage of this commit is truly new code vs replacement." A ratio consistently above 0.8 means 80% of every commit is net-new code — the actual batch-acceptance signal.
Alternatively, exclude commits where prod_files_count > 0 AND total_deletions < 5 from the ratio distribution (they're new-file bootstraps, not drift signals), and document this filtering in the output.