fuzz: adopt fuzzing (wave 1) + alkcall 0.8.1 / alktunnels 0.2.1 bumps

Adopt the alk*/fuzzing posture (alkcall 0.8.1 + alktunnels 0.2.1
precedent) before the v0.1 publish — the cheapest moment this crate
will ever have for it:

- alkcall 0.8.1 + alktunnels 0.2.1: the duplicate-adopt/open fix is
  directly relevant to this crate's consumer open/adopt paths
- fuzz/ workspace (the sibling layout): nightly-pinned subtree, stable
  fuzz/shared library (corpus replay = the fuzz gate), detached runner,
  JSON dict, deterministic seed generator; 129 committed seeds
- three wave-1 targets over already-pub APIs: datagram_codec (the
  ASSOCIATE frame codec — the alkcall-fuzz-doc-assigned downstream
  target), rfc1928_codec (the forked RFC 1928 core's handshake state
  machines, both protocol directions), params_parse (the wire-stable
  open-op surface)
- one diagnostic: DatagramReader::retained_bytes() (the decoder-state
  bound's assertion target, the alktunnels retained_bytes() posture)
- docs/research/fuzzing.md: the scoped plan + implementation log;
  AGENTS.md gains the corpus-replay verification line
- campaigns: 3 × 10-min detached runs exited 0 (datagram ~1.9M execs,
  rfc1928 ~2.0M, params ~3.1M; oom/timeout/crash 0/0/0 on all 33 fork
  jobs per target; coverage saturated on both codec targets). No crate
  bugs; three harness-model defects found + fixed (2 pre-campaign by
  corpus replay, 2 campaign-found, crash inputs promoted as seeds 900/
  901): the len==0 well-framed decode partition, the fieldless-struct
  serde-seq form, the BadAtyp re-frame arm, the short-domain count-walk
  gate
- wave 2 (session_pumps stateful target) deferred to post-0.1 — the
  byte-level targets carry the attacker-reachable load for the publish

Verification: cargo test (15 suites ok, default + all-features), clippy
(all-targets + wasm + fuzz/shared), fmt, cargo doc, publish --dry-run,
corpus replay 129 seeds green, cargo fuzz build clean.

Refs: alkcall docs/research/fuzzing.md (§7.9 posture), alktunnels
docs/research/fuzzing.md; docs/research/fuzzing.md (this crate's copy)

src/wire.rs

@@ -260,6 +260,15 @@ impl DatagramReader {
     pub fn is_mid_frame(&self) -> bool {
         self.current.is_some()
     }
+
+    /// The decoder's retained byte count: the buffered remainder plus
+    /// any in-flight frame's accumulated payload. The fuzz harness's
+    /// bound assertion (`fuzz/shared`'s `datagram_codec` target)
+    /// asserts against this diagnostic (the alktunnels
+    /// `retained_bytes()` posture); it is not a protocol surface.
+    pub fn retained_bytes(&self) -> usize {
+        self.buf.len() + self.current.as_ref().map_or(0, |(_, acc)| acc.len())
+    }
 }
 
 /// Decode one frame's inner bytes (`addr + data`, the length prefix