[sujal-maheshwari2004]

What was wrong

In changes(), the inner hunk loop stops when either side runs out. When that happened mid-hunk (editing the last line of a template with no trailing newline, or replacing the last N lines with M lines), the hunk was closed at that point, and the rest of the other side became a separate trailing hunk. One replace came out as a delete plus an insert, each with an empty side, so both had point delta 1.

Before, with the real Analyze on a 3-line prompt whose last line changed from "Always answer in English." to "Always answer in French.":

change={OldStart:2 OldEnd:3 NewStart:2 NewEnd:2} pointDelta=1.000 class="localized tweak"
change={OldStart:3 OldEnd:3 NewStart:2 NewEnd:3} pointDelta=1.000 class="structural"

After:

change={OldStart:2 OldEnd:3 NewStart:2 NewEnd:3} pointDelta=0.250 class="structural"

The fix

After the inner loop, if either side is exhausted, extend the hunk to the end of both sides (i, j = n, m). Nothing left can match, so the remainder belongs to the same contiguous hunk. The existing trailing check still covers hunks that start after the last matched line.

Note: a single replace at the very end of a file can still be classified structural. The down window hits the file edge at radius 2 with delta equal to the point delta, so the classifier counts it as persisting to the boundary. That is existing classifier behavior for edge-of-file edits and is not changed here. This PR only fixes the hunk split.

Tests

New TestEditAtEndIsOneReplace covers editing the last line, the only line, last line -> two lines, and last two lines -> one line.

Before the fix (all 4 fail):

--- FAIL: TestEditAtEndIsOneReplace/edit_last_line
    changes = [{OldStart:2 OldEnd:3 NewStart:2 NewEnd:2} {OldStart:3 OldEnd:3 NewStart:2 NewEnd:3}], want [{OldStart:2 OldEnd:3 NewStart:2 NewEnd:3}]

After:

$ GOWORK=off go build ./... && go vet ./... && go test -count=1 ./...
ok  github.com/priompt/proto/semdiff
ok  github.com/priompt/proto/validate

Fixes #5

🤖 Generated with Claude Code