diff --git a/app/ui/worddiff/worddiff.go b/app/ui/worddiff/worddiff.go
index 4eb66115..bcab2b8b 100644
--- a/app/ui/worddiff/worddiff.go
+++ b/app/ui/worddiff/worddiff.go
@@ -36,10 +36,18 @@ type Pair struct {
 	AddIdx    int
 }
 
-// maxLineLenForDiff caps intra-line diff to lines up to this many bytes.
-// longer lines skip word-level highlighting to avoid O(m*n) LCS memory blowup
-// on pathological input (minified files, very long configs).
-const maxLineLenForDiff = 500
+// maxLineLenForDiff caps intra-line diff to lines up to this many bytes. it is a cheap
+// pre-filter that keeps pathological input (minified bundles, single-line JSON) out of the
+// tokenizer; maxDiffCells is the actual cost guard.
+const maxLineLenForDiff = 20000
+
+// maxDiffCells caps the LCS table at this many cells (minus tokens * plus tokens).
+// the table dominates cost at 8 bytes per cell, so the budget is ~32MB and ~17ms per pair. it bounds
+// one pair and nothing wider: recomputeIntraRanges calls ComputeIntraRanges once per paired
+// remove/add line, so a file whose diff holds many long pairs pays this for each of them.
+// byte length is a poor proxy for the cost: 500 repeated letters tokenize to one token,
+// 500 bytes of minified JSON to nearly 300.
+const maxDiffCells = 4_000_000
 
 // similarityThreshold is the minimum percentage of common tokens for highlighting.
 // pairs with less than this percentage of common content get no intra-line overlay.
@@ -57,8 +65,8 @@ var tokenPattern = regexp.MustCompile(`[\pL\pN_]+|\s+|[^\pL\pN_\s]+`)
 
 // ComputeIntraRanges computes changed byte-offset ranges for a pair of minus/plus lines.
 // returns ranges for the minus line and plus line respectively.
-// returns nil ranges if either line is empty, exceeds maxLineLenForDiff,
-// or fails the similarity gate (< 30% common non-whitespace tokens).
+// returns nil ranges if either line is empty, exceeds maxLineLenForDiff, would need more than
+// maxDiffCells LCS cells, or fails the similarity gate (< 30% common non-whitespace tokens).
 func (d *Differ) ComputeIntraRanges(minusLine, plusLine string) ([]Range, []Range) {
 	if minusLine == "" || plusLine == "" {
 		return nil, nil
@@ -72,6 +80,9 @@ func (d *Differ) ComputeIntraRanges(minusLine, plusLine string) ([]Range, []Rang
 	if len(minusToks) == 0 || len(plusToks) == 0 {
 		return nil, nil
 	}
+	if len(minusToks)*len(plusToks) > maxDiffCells {
+		return nil, nil
+	}
 
 	keepMinus, keepPlus := d.lcsKeptTokens(minusToks, plusToks)
 	minusRanges := d.buildChangedRanges(minusToks, keepMinus)
diff --git a/app/ui/worddiff/worddiff_test.go b/app/ui/worddiff/worddiff_test.go
index 7bfb5584..688ee652 100644
--- a/app/ui/worddiff/worddiff_test.go
+++ b/app/ui/worddiff/worddiff_test.go
@@ -22,6 +22,9 @@ func changedRangesHelper(d *Differ, minusLine, plusLine string) ([]Range, []Rang
 	if len(minusToks) == 0 || len(plusToks) == 0 {
 		return nil, nil
 	}
+	if len(minusToks)*len(plusToks) > maxDiffCells {
+		return nil, nil
+	}
 	keepMinus, keepPlus := d.lcsKeptTokens(minusToks, plusToks)
 	return d.buildChangedRanges(minusToks, keepMinus), d.buildChangedRanges(plusToks, keepPlus)
 }
@@ -309,8 +312,8 @@ func TestChangedRanges_MultibytePrecision(t *testing.T) {
 
 func TestChangedRanges_SkipsVeryLongLines(t *testing.T) {
 	d := New()
-	// lines above maxLineLenForDiff must skip intra-line diff to prevent
-	// LCS memory blowup on pathological input (minified content).
+	// the byte cap is a pre-filter that keeps long lines out of the tokenizer;
+	// maxDiffCells is what guards the LCS cost itself.
 	longMinus := strings.Repeat("a", maxLineLenForDiff+1)
 	longPlus := strings.Repeat("b", maxLineLenForDiff+1)
 
@@ -332,6 +335,76 @@ func TestChangedRanges_SkipsVeryLongLines(t *testing.T) {
 	assert.Nil(t, pr2, "asymmetric long minus should skip")
 }
 
+// prose builds a line of exactly nbytes of space-separated words, deterministic across runs.
+func prose(nbytes int) string {
+	words := []string{"the", "review", "annotation", "paragraph", "diff", "highlight", "maintainer", "line"}
+	var sb strings.Builder
+	for i := 0; sb.Len() < nbytes; i++ {
+		sb.WriteString(words[i%len(words)])
+		sb.WriteByte(' ')
+	}
+	return sb.String()[:nbytes]
+}
+
+// pins issue #323: a prose pair far past the old 500-byte cap must still be word-diffed.
+// the guard is the LCS table size, not the byte length of either line.
+func TestComputeIntraRanges_LongProsePairIsDiffed(t *testing.T) {
+	d := New()
+	minus := prose(3602)
+	plus := minus + " and a trailing sentence appended to the paragraph"
+
+	minusRanges, plusRanges := d.ComputeIntraRanges(minus, plus)
+	assert.Empty(t, minusRanges, "nothing removed from the minus line")
+	require.NotEmpty(t, plusRanges, "the appended tail must be highlighted")
+	// whitespace tokens are excluded, so the tail arrives as one range per word
+	assert.Equal(t, len(minus)+1, plusRanges[0].Start, "highlight starts at the first appended word")
+	assert.Equal(t, len(plus), plusRanges[len(plusRanges)-1].End, "highlight runs to the end of the line")
+}
+
+func TestComputeIntraRanges_CostGuards(t *testing.T) {
+	d := New()
+
+	t.Run("over byte pre-filter returns nil", func(t *testing.T) {
+		// one long run is a single token, so the pair clears the cell budget and the similarity
+		// gate — the byte pre-filter is the only thing that can reject it
+		minus := strings.Repeat("a", maxLineLenForDiff+1)
+		plus := minus + " b"
+		require.LessOrEqual(t, len(d.tokenizeLineWithOffsets(minus))*len(d.tokenizeLineWithOffsets(plus)), maxDiffCells,
+			"fixture must clear the cell budget so only the byte pre-filter can reject it")
+
+		minusRanges, plusRanges := d.ComputeIntraRanges(minus, plus)
+		assert.Nil(t, minusRanges)
+		assert.Nil(t, plusRanges)
+
+		// the same shape sized to fit the pre-filter is diffed, so the nil above is the byte gate
+		under := strings.Repeat("a", maxLineLenForDiff-2)
+		_, underPlus := d.ComputeIntraRanges(under, under+" b")
+		require.Len(t, underPlus, 1, "pair under the pre-filter must still be diffed")
+	})
+
+	t.Run("token product over budget returns nil", func(t *testing.T) {
+		minus, plus := prose(maxLineLenForDiff), prose(maxLineLenForDiff-5)+" zzzz"
+		require.LessOrEqual(t, len(minus), maxLineLenForDiff, "must clear the byte pre-filter")
+		require.LessOrEqual(t, len(plus), maxLineLenForDiff, "must clear the byte pre-filter")
+		require.Greater(t, len(d.tokenizeLineWithOffsets(minus))*len(d.tokenizeLineWithOffsets(plus)), maxDiffCells,
+			"fixture must exceed the cell budget")
+
+		minusRanges, plusRanges := d.ComputeIntraRanges(minus, plus)
+		assert.Nil(t, minusRanges)
+		assert.Nil(t, plusRanges)
+	})
+
+	t.Run("few tokens over the old byte cap are diffed", func(t *testing.T) {
+		// a single long run tokenizes to one token, so cost is trivial regardless of byte length
+		minus := strings.Repeat("a", 4000)
+		plus := strings.Repeat("a", 4000) + " b"
+		minusRanges, plusRanges := d.ComputeIntraRanges(minus, plus)
+		assert.Empty(t, minusRanges)
+		require.Len(t, plusRanges, 1)
+		assert.Equal(t, "b", plus[plusRanges[0].Start:plusRanges[0].End])
+	})
+}
+
 func TestPassesSimilarityGateFromKeep(t *testing.T) {
 	d := New()
 	tests := []struct {
diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md
index c583bd50..b34af507 100644
--- a/docs/ARCHITECTURE.md
+++ b/docs/ARCHITECTURE.md
@@ -293,8 +293,10 @@ Single stateless type `Differ` grouping all word-diff algorithms:
 - `InsertHighlightMarkers()` — ANSI-aware highlight insertion, shared by both word-diff and search
   highlighting
 
-30% similarity gate discards ranges for dissimilar pairs. Ranges are byte offsets on tab-replaced
-content, aligning with `prepareLineContent` output.
+30% similarity gate discards ranges for dissimilar pairs. Cost is guarded by a budget on the LCS
+table size (token count of one line times the other), with a byte pre-filter that keeps minified
+input out of the tokenizer — line length alone is not the cost driver. Ranges are byte offsets on
+tab-replaced content, aligning with `prepareLineContent` output.
 
 ### app/highlight/ — syntax highlighting