test(chunks): pin exact-overlap contracts for trusted merges

- Overlaps wider than the 400-rune text window must be trimmed once.
- Adjacent chunks built from one repeating row must keep every row.
- A length-preserving text mismatch must fall back and keep both bodies.

internal/application/service/chat_pipeline/merge_position_path_test.go

@@ -55,6 +55,52 @@ func TestGroupAndMerge_TrustedOverlapTrimsAndExtendsRange(t *testing.T) {
 	assert.ElementsMatch(t, []string{"c2"}, results[0].SubChunkID)
 }
 
+func TestGroupAndMerge_TrustedOverlapBeyondTextSearchWindow(t *testing.T) {
+	// A real overlap wider than the text matcher's 400-rune window: the position
+	// path knows the exact amount, so the overlap must not survive twice.
+	overlap := rangeText('o', 500)
+	first := trustedResult("c1", 1, 0, 1000, rangeText('a', 500)+overlap)
+	second := trustedResult("c2", 2, 500, 1500, overlap+rangeText('b', 500))
+
+	results := mergeGrouped(docChunks(first, second))
+
+	require.Len(t, results, 1)
+	assert.Equal(t, rangeText('a', 500)+overlap+rangeText('b', 500), results[0].Content)
+	assert.Equal(t, 1500, results[0].EndAt)
+}
+
+func TestGroupAndMerge_TrustedAdjacentRepeatedTextKeepsEveryRow(t *testing.T) {
+	// Adjacent trusted chunks (positional overlap 0) made of one repeating row:
+	// searching for the longest suffix match would mistake the repetition for an
+	// overlap and drop rows, so the exact-overlap path must concatenate verbatim.
+	row := "| cell | cell |\n"
+	firstBody := rangeText('x', 100) + strings.Repeat(row, 2)
+	secondBody := strings.Repeat(row, 3) + rangeText('y', 60)
+	first := trustedResult("c1", 1, 0, runeLen(firstBody), firstBody)
+	second := trustedResult("c2", 2, runeLen(firstBody), runeLen(firstBody)+runeLen(secondBody), secondBody)
+
+	results := mergeGrouped(docChunks(first, second))
+
+	require.Len(t, results, 1)
+	assert.Equal(t, firstBody+secondBody, results[0].Content)
+	assert.Equal(t, 5, strings.Count(results[0].Content, row), "no repeated row may be dropped")
+}
+
+func TestGroupAndMerge_TrustedOverlapTextMismatchFallsBackToTextMatch(t *testing.T) {
+	// Length invariant holds but the bodies disagree inside the overlap window
+	// (HTML entities, synthetic headers): the exact path must refuse and the text
+	// fallback must keep both bodies rather than trim on coordinates.
+	first := trustedResult("c1", 1, 0, 100, rangeText('a', 80)+rangeText('b', 20))
+	second := trustedResult("c2", 2, 80, 200, rangeText('c', 20)+rangeText('d', 100))
+
+	results := mergeGrouped(docChunks(first, second))
+
+	require.Len(t, results, 1)
+	assert.Contains(t, results[0].Content, rangeText('b', 20))
+	assert.Contains(t, results[0].Content, rangeText('c', 20))
+	assert.Contains(t, results[0].Content, rangeText('d', 100))
+}
+
 func TestGroupAndMerge_TrustedAdjacentJoinsSeamlessly(t *testing.T) {
 	// Adjacent trusted chunks are contiguous in the original document: the
 	// merged body must be the exact concatenation, with no separator inserted.

internal/searchutil/chunkmerge_test.go

@@ -58,6 +58,50 @@ func TestAppendWithOverlap_NoOverlap(t *testing.T) {
 	}
 }
 
+func TestAppendWithExactOverlap_TrimsKnownOverlap(t *testing.T) {
+	overlap := "shared boundary text"
+	got, ok := AppendWithExactOverlap("before "+overlap, overlap+" after", len([]rune(overlap)))
+	if !ok {
+		t.Fatal("exact overlap should be accepted")
+	}
+	want := "before " + overlap + " after"
+	if got != want {
+		t.Fatalf("exact overlap mismatch:\n got=%q\nwant=%q", got, want)
+	}
+}
+
+func TestAppendWithExactOverlap_ZeroOverlapConcatenatesRepeatedText(t *testing.T) {
+	// 首尾相接的两段都由同一周期性文本组成（表格行 / 日志行）。重叠量为 0 时
+	// 必须原样拼接，不能去搜索后缀匹配，否则会吃掉 next 开头的重复行。
+	row := "| cell | cell |\n"
+	acc := "前言\n" + row + row
+	next := row + row + row + "结尾\n"
+	got, ok := AppendWithExactOverlap(acc, next, 0)
+	if !ok {
+		t.Fatal("zero overlap should be accepted")
+	}
+	if got != acc+next {
+		t.Fatalf("zero overlap must concatenate verbatim:\n got=%q\nwant=%q", got, acc+next)
+	}
+}
+
+func TestAppendWithExactOverlap_RejectsMismatchedOverlap(t *testing.T) {
+	// 长度不变式成立但文本已经不同（HTML 实体、补写表头等），必须拒绝，
+	// 由调用方回退到按文本匹配。
+	if _, ok := AppendWithExactOverlap("abcdefghijkl", "XYZdefghijkl", 6); ok {
+		t.Fatal("mismatched overlap should be rejected")
+	}
+}
+
+func TestAppendWithExactOverlap_RejectsOverlapLongerThanBodies(t *testing.T) {
+	if _, ok := AppendWithExactOverlap("short", "shorter", 99); ok {
+		t.Fatal("overlap exceeding body length should be rejected")
+	}
+	if _, ok := AppendWithExactOverlap("short", "shorter", -1); ok {
+		t.Fatal("negative overlap should be rejected")
+	}
+}
+
 func TestJoinChunkContentUsesCurrentTextInsteadOfSourceOffsets(t *testing.T) {
 	first := "first edited body with no original overlap"
 	second := "second independently edited body"