perf/search: keep folded mismatch checks constant-time

Most candidate mismatches occur well before the document boundary, where UTF-8's maximum rune width can prove the remaining span fits without decoding it. Reserve the exact rune walk for boundary-adjacent candidates so corruption remains observable without multiplying query-length work across broad candidate sets.

Amp-Thread-ID: https://ampcode.com/threads/T-019fd1ab-1ad4-77f5-9b59-c65232f5b3f1

Author: keegancsmith

index/bits.go

@@ -94,6 +94,19 @@ func caseFoldingEqualsRunes(lower, mixed []byte) (int, bool, bool) {
 }
 
 func hasEnoughRunes(mixed, lower []byte) bool {
+	// A UTF-8 rune occupies at most UTFMax bytes, and lower cannot contain
+	// more runes than bytes. A sufficiently long mixed suffix therefore proves
+	// the candidate span fits without walking either slice. In practice this
+	// keeps ordinary mismatches constant-time and reserves the exact scan below
+	// for candidates close enough to the document end to be truncated.
+	if len(mixed)/utf8.UTFMax >= len(lower) {
+		return true
+	}
+
+	// Compare only rune availability, not values: the caller already knows the
+	// candidate mismatches. Reaching the end of mixed first means the candidate's
+	// expected rune span crosses the document boundary, which is a corrupt index
+	// invariant rather than an ordinary non-match.
 	for len(lower) > 0 {
 		if len(mixed) == 0 {
 			return false