perf(scan): memchr2 cross-chunk jump in NEON and AVX2 fast probe

README.md

@@ -103,10 +103,10 @@ ARM64 (Apple M4, NEON/PMULL scanner, same workload):
 
 | Size | cjson | `qd.parse` | `qd.decode + t.f x3` | speedup vs. cjson |
 |---:|---:|---:|---:|---:|
-|   2 KB | 254,738 | 654,108 | 392,711 | 2.6× / 1.5× |
-| 100 KB |  15,281 | 108,932 |  99,701 | 7.1× / 6.5× |
-|   1 MB |   1,523 |  11,905 |  11,876 | 7.8× / 7.8× |
-|  10 MB |     153 |   1,218 |   1,222 | 8.0× / 8.0× |
+|   2 KB | 237,124 | 705,000 | 390,000 |  3.0× /  1.6× |
+| 100 KB |  14,667 | 232,000 | 208,000 | 15.8× / 14.2× |
+|   1 MB |   1,494 |  33,700 |  33,000 | 22.6× / 22.1× |
+|  10 MB |     150 |   3,376 |   3,454 | 22.5× / 23.0× |
 
 See [`docs/benchmarks.md`](docs/benchmarks.md) for the full size ladder,
 memory numbers, an "encode round-trip" row (passthrough emit via

src/scan/avx2.rs

@@ -38,6 +38,22 @@ unsafe fn scan_avx2_impl(buf: &[u8], out: &mut Vec<u32>) -> Result<(), usize> {
             if interesting == 0 {
                 bs_carry = 0;
                 i += 64;
+                // Cross-chunk jump: no quote/backslash means in_string polarity
+                // cannot flip and no escape can start, so jump straight to the
+                // 64B-aligned chunk containing the next interesting byte.
+                // The 4 KB remaining-buffer threshold suppresses the memchr2
+                // call entirely on small payloads (≤4 KB total), where the per-
+                // call libc overhead exceeds the in-string probe loop it would
+                // replace. On larger payloads only the last 4 KB foregoes the
+                // jump — negligible against MB-scale gains.
+                if i + 4096 <= buf.len() {
+                    let scan_end = buf.len() - (buf.len() % 64);
+                    let jump = match memchr::memchr2(b'"', b'\\', &buf[i..scan_end]) {
+                        Some(rel) => rel & !63,
+                        None      => scan_end - i,
+                    };
+                    i += jump;
+                }
                 continue;
             }
         }

src/scan/neon.rs

@@ -176,6 +176,24 @@ unsafe fn scan_neon_impl(buf: &[u8], out: &mut Vec<u32>) -> Result<(), usize> {
             if (quote_probe | backslash_probe) == 0 {
                 bs_carry = 0;
                 i += 64;
+                // Cross-chunk jump: with no quote/backslash in the chunk we just
+                // skipped, in_string polarity cannot flip and no escape can start,
+                // so we can use memchr2 to skip ahead to the 64B-aligned chunk
+                // containing the next interesting byte. Bounded by the last full
+                // 64B chunk; the <64B tail is handled by the scalar resume path.
+                // The 4 KB remaining-buffer threshold suppresses the memchr2
+                // call entirely on small payloads (≤4 KB total), where the per-
+                // call libc overhead exceeds the in-string probe loop it would
+                // replace. On larger payloads only the last 4 KB foregoes the
+                // jump — negligible against MB-scale gains.
+                if i + 4096 <= buf.len() {
+                    let scan_end = buf.len() - (buf.len() % 64);
+                    let jump = match memchr::memchr2(b'"', b'\\', &buf[i..scan_end]) {
+                        Some(rel) => rel & !63,
+                        None      => scan_end - i,
+                    };
+                    i += jump;
+                }
                 continue;
             }
         }