1 files changed, 40 insertions, 10 deletions

diff --git a/internal/dup/dup.go b/internal/dup/dup.go
index e3c7424..32bcfb0 100644
--- a/internal/dup/dup.go
+++ b/internal/dup/dup.go
@@ -174,18 +174,47 @@ func (x *Index) Lookup(path string) (original string, dup bool, err error) {
 		return path, false, nil
 	}
 	orig := x.candidates[x.original(identical)].path
-	return orig, orig != path, nil
+	if orig == path {
+		return orig, false, nil
+	}
+	// Spec §5.5: two names for one file are never duplicates of each other.
+	// The other name may be a hardlink, or path itself indexed a second time
+	// under a DIR that overlaps the scanned tree. identicalTo gives every
+	// member of the content class the same set, so every lookup elects the
+	// same original, and no name for that original's file is reported as a
+	// duplicate: its content always keeps at least one name. Portable:
+	// os.SameFile, never a Stat_t.Dev/Ino read (that field's type differs
+	// across freebsd/openbsd, which `make ci` vets).
+	origInfo, err := os.Lstat(orig)
+	if err != nil {
+		return "", false, err
+	}
+	pathInfo, err := os.Lstat(path)
+	if err != nil {
+		return "", false, err
+	}
+	if os.SameFile(origInfo, pathInfo) {
+		return orig, false, nil
+	}
+	return orig, true, nil
 }
 
 // identicalTo returns the indexes in group (which all share idx's size,
 // idx included) whose content matches candidates[idx]: same partial hash,
-// then, only for those that collide, the same full hash. idx is the file
-// Lookup was asked about; a failure hashing it propagates, since Lookup can
-// answer nothing without it. A failure hashing any other candidate in group
-// only removes that candidate from consideration: a vanished candidate
-// (errors.Is fs.ErrNotExist) is dropped silently, any other failure is
-// recorded on the Index (see recordCandidateError) so the caller can warn
-// about it once matching is done.
+// then, only for those that collide, the same full hash. Every candidate
+// with identical bytes is included, whatever its path or inode: a hardlink
+// of idx, and idx's own path indexed a second time under an overlapping
+// extra directory, are both members. That keeps the set the same whichever
+// member Lookup was asked about, so every member elects the same original;
+// Lookup, not this function, decides that a name for the elected original's
+// own file is not a duplicate of it.
+//
+// idx is the file Lookup was asked about; a failure hashing it propagates,
+// since Lookup can answer nothing without it. A failure hashing any other
+// candidate in group only removes that candidate from consideration: a
+// vanished candidate (errors.Is fs.ErrNotExist) is dropped silently, any
+// other failure is recorded on the Index (see recordCandidateError) so the
+// caller can warn about it once matching is done.
 func (x *Index) identicalTo(idx int, group []int) ([]int, error) {
 	idxPartial, err := x.partialHash(x.candidates[idx].path)
 	if err != nil {
@@ -219,9 +248,10 @@ func (x *Index) identicalTo(idx int, group []int) ([]int, error) {
 			x.recordCandidateError(x.candidates[j].path, err)
 			continue
 		}
-		if jFull == idxFull {
-			same = append(same, j)
+		if jFull != idxFull {
+			continue
 		}
+		same = append(same, j)
 	}
 	return same, nil
 }