Function scan_merge 

pub fn scan_merge(
    merge: &str,
    paths: &[(String, Blobs)],
    report: &mut ScanReport,
)

Applies the safety check to every path of one merge, counting into report.

Pure: the caller supplies the blobs. merges_scanned is incremented here, so a caller that skips a merge before reaching this function must count that skip itself.

Relocation

crate::safety::check compares one path against itself, so a merge that moves content from one file to another looks like a reversion in the source and an injection in the destination. That is not a silent revert: nothing was lost. The first corpus this was pointed at produced exactly that, a documentation section moved between two files during the merge, and it was 40% of the raw findings.

So after every path of a merge is checked, a line that appears as reverted in one path and injected in another path of the same merge is cancelled from both and counted in ScanReport::lines_relocated. The cancellation is per merge and never across merges: content leaving one commit and appearing in another, later, is not a move, and treating it as one would hide the exact class this scan exists to count.

Survival, which is a second number rather than a second filter

That cancellation requires the line to be unattributable at both ends. A refactor that moves a function into a file the author was already editing does not qualify: the destination addition is attributable to base -> proposed, so it never enters the injected set, so it cannot cancel anything, and the source removal is reported as a reversion of work that is sitting in the result untouched. On git/git that is most of what the raw findings are – the object database refactor moving blocks out of object-file.c reads as fourteen reverted lines.

ScanReport::lines_surviving_elsewhere counts reverted lines that are present somewhere in this merge’s result, and ScanReport::findings_all_surviving counts findings made entirely of them.

They are counted and still reported. Cancelling them would be the stronger detector and the weaker measurement: line presence anywhere in a result is a cheap test that a short or idiomatic line passes by accident, so silently dropping on it would remove true findings with no way to see how many. The kill criterion for this measurement was fixed before any number was known, and a filter added after seeing the data is exactly the move that discipline forbids. Two numbers let a reader bound the answer from both sides; one number chosen after the fact lets them do neither.