The >50-deleted-lines do-not-auto-merge heuristic misfires on extraction refactors, test migrations, and generated artifacts #124

The gap

Phase 8's scope gate treats a large deletion count as a signal that a human should look at the change. The heuristic is sound for its intended target (a PR quietly removing behaviour) but it has no exemption for the common cases where a large deletion count is structurally meaningless:

• Extraction refactors — lines move out of one file into a new module. The deletion count is large; the net behaviour change is zero.
• Test migrations — a suite is ported to a new runner or fixture style. Deletions are large and entirely mechanical.
• Generated artifacts — regenerating a lockfile, a vendored bundle, or a generated doc can delete hundreds of lines that no human should be reading line by line.

The result is that instances doing legitimate cleanup work stall on a human gate that carries no information, which trains the operator to wave the gate through — weakening it for the case it actually exists to catch.

Reported independently by: asn1c-dev-loop#7, foodspoilage-dev-loop#8, repo-quality-dashboard-dev-loop#4, parpt-dev-loop#7, kreatures-dev-loop#1.

Suggested resolution

Keep the ceiling, but make it measure net removal rather than raw deletions, and name the exemptions:

The deletion ceiling applies to net removed lines. A diff whose deletions are matched by additions elsewhere in the same PR (an extraction refactor, a file rename, a test migration) does not trip the gate. Paths that are generated rather than authored — lockfiles, vendored trees, generated documentation — are excluded from the count entirely; list them in the instance's generated-paths set at generation time.

Requiring the instance to record its own generated-paths set during Step 2 keeps the exemption honest, rather than letting a cycle decide case by case that its own deletions do not count.

drafted by Claude on behalf of Daniel Stephenson