test(proto): add decoder fuzzing

Add isolated cargo-fuzz/libFuzzer infrastructure for irtt-proto, fuzzing
decode_request, decode_open_reply, decode_echo_reply, Params::decode(_with_presence),
and the varint codec against arbitrary and structurally-derived wire input.
Each target's core invariant is that arbitrary bytes must never panic; a
small clean-room seed corpus (generated from irtt-proto's own encoders) and
a bounded nightly smoke CI workflow are included. The fuzz project is its
own isolated cargo workspace and does not affect the root workspace, its
lockfile, or any published crate's dependency graph.

fuzz/fuzz_targets/decode_varint.rs

@@ -0,0 +1,38 @@
+#![no_main]
+
+use irtt_proto::varint::{decode_uvarint, decode_varint, encode_uvarint, encode_varint};
+use libfuzzer_sys::fuzz_target;
+
+/// See `decode_request.rs` for the rationale behind this bound. Varint
+/// grammar is trivial, but the bound is kept for consistency with the other
+/// targets.
+const MAX_INPUT_LEN: usize = 128 * 1024;
+
+fuzz_target!(|data: &[u8]| {
+    if data.len() > MAX_INPUT_LEN {
+        return;
+    }
+
+    if let Ok((value, used)) = decode_uvarint(data) {
+        assert!(used > 0);
+        assert!(used <= data.len());
+        assert!(used <= 10);
+
+        let mut encoded = Vec::new();
+        encode_uvarint(value, &mut encoded);
+        // Canonical re-encoding need not reproduce noncanonical fuzz input,
+        // but it must decode back to the same value and consume exactly its
+        // own bytes.
+        assert_eq!(decode_uvarint(&encoded), Ok((value, encoded.len())));
+    }
+
+    if let Ok((value, used)) = decode_varint(data) {
+        assert!(used > 0);
+        assert!(used <= data.len());
+        assert!(used <= 10);
+
+        let mut encoded = Vec::new();
+        encode_varint(value, &mut encoded);
+        assert_eq!(decode_varint(&encoded), Ok((value, encoded.len())));
+    }
+});