diff --git a/crates/yip-bench/examples/fec_rate_experiment.rs b/crates/yip-bench/examples/fec_rate_experiment.rs
new file mode 100644
index 0000000..9a3b4db
--- /dev/null
+++ b/crates/yip-bench/examples/fec_rate_experiment.rs
@@ -0,0 +1,40 @@
+//! #59 FEC-rate experiment: does the Default class's proactive repair actually
+//! DECAY on a clean link, or does its non-ARQ floor keep it permanently at
+//! ~1 repair per single-symbol object? Drives the real
+//! `Transport::observe_loss` clean-link feedback (loss = 0.0) many times, then
+//! measures symbols emitted per 1184-byte Default packet (1 source symbol).
+//!
+//! Run: cargo run --release -p yip-bench --example fec_rate_experiment
+use yip_transport::{FlowClass, Transport};
+
+fn measure(clean_reports: u32) -> f64 {
+    let mut tx = Transport::new(vec![], 1200);
+    // Simulate a sustained clean link: the receiver's periodic (30 ms)
+    // LossReport drives observe_loss(Default, 0.0) each interval.
+    for _ in 0..clean_reports {
+        tx.observe_loss(FlowClass::Default, 0.0);
+    }
+    // A 1184-byte inner packet -> 1200-byte ciphertext -> exactly 1 source symbol.
+    let inner = vec![0xABu8; 1184];
+    let ciphertext = vec![0xCDu8; inner.len() + 16];
+    let iters = 2000u32;
+    let mut nsym = 0usize;
+    for _ in 0..iters {
+        // Keep the link clean while sending so the ratio stays decayed.
+        tx.observe_loss(FlowClass::Default, 0.0);
+        let (_c, syms) = tx.encode(&ciphertext, &inner, false, 0);
+        nsym += syms.len();
+    }
+    nsym as f64 / f64::from(iters)
+}
+
+fn main() {
+    println!("Default class, 1184-byte packet (1 source symbol), after clean-link decay:");
+    for reports in [0u32, 10, 50, 200, 1000] {
+        let spp = measure(reports);
+        println!("  {reports:>5} clean LossReports -> {spp:.3} symbols/packet (datagrams/packet)");
+    }
+    println!();
+    println!("Interpretation: 2.000 = permanent +1 repair (floor never reaches 0);");
+    println!("                1.000 = repair decayed to 0 (single datagram per packet).");
+}
diff --git a/crates/yip-bench/examples/mtu_symbol_sweep.rs b/crates/yip-bench/examples/mtu_symbol_sweep.rs
new file mode 100644
index 0000000..59ac8ee
--- /dev/null
+++ b/crates/yip-bench/examples/mtu_symbol_sweep.rs
@@ -0,0 +1,50 @@
+//! #59 measurement spike: how the FEC symbol count per inner packet varies
+//! with the inner packet size and the FEC `symbol_size`. Real code path
+//! (`yip_transport::Transport::encode`), so the counts are authoritative, not
+//! arithmetic. Run:
+//!   cargo run --release -p yip-bench --example mtu_symbol_sweep
+//!
+//! Reports, for each (inner_size, symbol_size): the source-symbol count
+//! (= ceil((inner+16)/symbol_size)), the total symbols emitted (source +
+//! adaptive FEC repair), and the resulting outer wire bytes per inner packet
+//! (symbols x (symbol_size + per-symbol overhead)). Overhead per symbol is the
+//! yip-wire frame (HEADER_LEN 15 + TAG_LEN 8 = 23) + obf envelope
+//! (MIN_ENVELOPE 11) + IP/UDP (28 v4), a lower bound (ignores Symbol metadata
+//! + random obf pad).
+use yip_transport::Transport;
+
+const AEAD_TAG: usize = 16; // ChaCha20-Poly1305 tag (Session::seal)
+const WIRE_OVERHEAD: usize = 15 + 8; // yip-wire HEADER_LEN + TAG_LEN
+const OBF_ENVELOPE: usize = 11; // yip-obf MIN_ENVELOPE (nonce 8 + type 1 + len 2)
+const IP_UDP_V4: usize = 28; // 20 IPv4 + 8 UDP
+
+fn main() {
+    // A fresh Transport classifies filler as the Default class; its adaptive
+    // repair ratio starts at the class default (no observed loss here), so the
+    // repair count is the clean-link baseline.
+    let inner_sizes = [576usize, 1184, 1400, 1448, 1500, 4000, 8972];
+    let symbol_sizes = [1200u16, 1452, 1500, 9000];
+
+    println!(
+        "{:>10} {:>11} {:>7} {:>7} {:>6} {:>11} {:>9}",
+        "inner", "symbol_size", "source", "total", "repair", "outer_bytes", "overhead%"
+    );
+    for &inner_size in &inner_sizes {
+        let inner = vec![0xABu8; inner_size];
+        let ciphertext = vec![0xCDu8; inner_size + AEAD_TAG];
+        for &symbol_size in &symbol_sizes {
+            let mut tx = Transport::new(vec![], symbol_size);
+            let (_class, syms) = tx.encode(&ciphertext, &inner, false, 0);
+            let source = ciphertext.len().div_ceil(usize::from(symbol_size));
+            let total = syms.len();
+            let repair = total.saturating_sub(source);
+            let per_symbol = usize::from(symbol_size) + WIRE_OVERHEAD + OBF_ENVELOPE + IP_UDP_V4;
+            let outer_bytes = total * per_symbol;
+            let overhead_pct = 100.0 * (outer_bytes as f64 - inner_size as f64) / inner_size as f64;
+            println!(
+                "{inner_size:>10} {symbol_size:>11} {source:>7} {total:>7} {repair:>6} {outer_bytes:>11} {overhead_pct:>8.0}%"
+            );
+        }
+        println!();
+    }
+}
diff --git a/docs/research/2026-07-22-mtu-packetization-findings.md b/docs/research/2026-07-22-mtu-packetization-findings.md
new file mode 100644
index 0000000..516ac9e
--- /dev/null
+++ b/docs/research/2026-07-22-mtu-packetization-findings.md
@@ -0,0 +1,214 @@
+# #59 MTU-aware packetization — measurement findings (throughput lever 4c)
+
+**Date:** 2026-07-22
+**Status:** investigation complete; feeds the mechanism design (deferred pending these findings).
+**Method:** real code path (`yip_transport::Transport::encode`) via
+`crates/yip-bench/examples/mtu_symbol_sweep.rs`; source reading of the
+send-path layering; the 4a/4b profiling in `crates/yip-bench/RESULTS.md`.
+
+## The send-path layering (confirmed from source)
+
+```
+inner IP packet (≤ TUN MTU)
+  → Session::seal_into          +16  (ChaCha20-Poly1305 tag)          [per inner packet]
+  → Transport::encode           split into K = ceil((inner+16)/symbol_size) SOURCE symbols
+                                 + R repair symbols (adaptive)          [per symbol]
+      each symbol is EXACTLY symbol_size on the wire (last zero-padded)
+  → Codec::frame                +23  (yip-wire HEADER_LEN 15 + TAG_LEN 8) [per symbol]
+  → obf envelope                +11  (yip-obf MIN_ENVELOPE) + random pad   [per symbol]
+  → UDP/IP                      +28 v4 / +48 v6                            [per datagram]
+```
+
+Two cost tiers matter:
+- **per-inner-packet:** AEAD seal/open (now the dominant per-packet cost per
+  RESULTS.md), TUN write (~20%). One each per inner packet, independent of
+  `symbol_size`.
+- **per-symbol / per-datagram:** UDP `sendmsg` (~23%), yip-wire SipHash
+  header-auth (~9%, #58), FEC encode (~1.3µs), framing. One each per symbol.
+
+`yipd` does **not** currently set the TUN MTU (no `SIOCSIFMTU` in `yip-device`,
+nothing in the netns scripts) — the interface takes the kernel default (1500).
+The QUIC `MtuDiscoveryConfig` (1200→1452) is only the QUIC-*mimicry* transport,
+a separate path; the real data plane has no PMTUD.
+
+## The measurement (`mtu_symbol_sweep`, real `Transport::encode`)
+
+Symbols emitted per inner packet, and lower-bound outer wire bytes
+(symbols × (symbol_size + 23 + 11 + 28)), fresh Transport (default-class
+repair ratio, clean-link baseline):
+
+| inner | symbol_size | source | total | repair | outer_bytes | overhead |
+|------:|------------:|-------:|------:|-------:|------------:|---------:|
+|  1184 |        1200 |      1 |     2 |      1 |        2524 |     113% |
+|  1400 |        1200 |      2 |     3 |      1 |        3786 |     170% |
+|  1400 |        1500 |      1 |     2 |      1 |        3124 |     123% |
+|  1448 |        1200 |      2 |     3 |      1 |        3786 |     161% |
+|  1448 |        1452 |      2 |     3 |      1 |        4542 |     214% |
+|  1448 |        1500 |      1 |     2 |      1 |        3124 |     116% |
+|  8972 |        1200 |      8 |     9 |      1 |       11358 |      27% |
+|  8972 |        1500 |      6 |     7 |      1 |       10934 |      22% |
+|  8972 |        9000 |      1 |     2 |      1 |       18124 |     102% |
+|   576 |        9000 |      1 |     2 |      1 |       18124 |    3047% |
+
+## Findings
+
+### 1. Every symbol is zero-padded to full `symbol_size` on the wire — so a blind `symbol_size` increase is a bandwidth REGRESSION for small packets.
+
+`fec.rs` (“last zero-padded”), the reassembler rejects `data.len() != symbol_size`
+(fec.rs:206), and `wire_glue::symbol_to_frame` frames the full `sym.data`
+untrimmed. A 576-byte packet at `symbol_size` 9000 emits two 9000-byte symbols
+(3047% overhead). `symbol_size` therefore cannot simply be cranked up: it is a
+per-datagram floor that every packet, however small, pays in full.
+
+### 2. Two independent MTU levers with very different reach.
+
+- **Lever 1 — raise `symbol_size` to ~1500.** Collapses a standard 1400–1500 B
+  inner packet from 2 source symbols to 1 (1448 B: 3→2 datagrams, −33%).
+  Reduces only the **per-symbol** costs (sendmsg, SipHash, FEC, framing) — NOT
+  AEAD or TUN-write. Internet-safe (path MTU ~1500). Caveat: `symbol_size` must
+  be ≥ inner+16, so 1452 does NOT fit a 1448 B packet — need 1500. And it pays
+  Finding 1's padding penalty on sub-`symbol_size` traffic.
+- **Lever 2 — raise the inner TUN MTU to jumbo (~9000) with a matching
+  `symbol_size`.** Fewer, bigger inner packets → reduces **all** costs including
+  AEAD + TUN-write (the two largest): moving 8972 B as one jumbo inner packet
+  (2 datagrams, 1 seal, 1 TUN write) vs ~6× 1448 B packets (~12 datagrams,
+  6 seals, 6 TUN writes) is ~6× fewer of everything. But it needs the **whole
+  path** to carry jumbo, and only helps when traffic is actually jumbo-sized
+  (Finding 1 punishes small packets hard at `symbol_size` 9000).
+
+The issue’s framing — raising MTU “reduces every per-packet cost proportionally”
+— is fully true only for **Lever 2**. Lever 1’s ceiling is the per-symbol share
+(~sendmsg 23% + SipHash 9% ≈ 32%), and only for packets that currently split.
+
+### 3. The FEC repair floor (+1 per object, min-1) doubles single-symbol objects — but it is loss-adaptive.
+
+Every case above shows `repair = 1`: `repair_count = max(1, ceil(K × ratio))`,
+so a 1-source-symbol object is 1 source + 1 repair = **100% packet overhead**,
+while an 8-source jumbo object amortizes the same +1 over 8 (12.5%). This is a
+large packet-count multiplier for MTU-sized single-packet flows — BUT
+`repair_count` returns **0** once the adaptive ratio decays to zero on a clean
+link (ARQ-eligible classes), so steady-state clean traffic would not pay it. The
+sweep shows the fresh/default-ratio baseline, not steady state. This is an
+**FEC-rate** lever (which classes FEC-protect single-symbol objects on clean
+links), adjacent to but distinct from #59’s MTU scope — flag for separate
+investigation; it may be a bigger single-core win than the MTU change.
+
+### 4. The structural fix that removes Finding 1’s tradeoff: variable-length last symbol (wire-trim).
+
+Zero-pad only for the GF math on encode/decode; carry the true byte length and
+trim the partial final source symbol on the wire (`object_size` is already
+transmitted, so the receiver can reconstruct). This decouples FEC granularity
+from wire bytes: `symbol_size` can rise to collapse source-symbol counts
+**without** padding small packets. It is more invasive (touches the RS codec
+wire representation of the last symbol) but removes the Lever-1 tradeoff
+entirely and is the enabler for a safe, always-on MTU increase.
+
+## Recommendation (feeds the mechanism design)
+
+Ranked, given the padding constraint:
+
+1. **Wire-trim the last symbol (Finding 4) + static configurable inner MTU with
+   a derived `symbol_size`.** The wire-trim makes Lever 1 safe on the internet
+   (no small-packet penalty); the configurable MTU ships Lever 2 for
+   controlled/jumbo paths (the biggest win, which the issue itself calls
+   “largest on controlled/LAN paths”). `yipd` gains a `SIOCSIFMTU` on the TUN
+   device and the derivation `path_MTU → symbol_size → inner MTU`, with the safe
+   default = today’s behavior.
+2. **If wire-trim is out of scope for v1:** static configurable MTU only,
+   `symbol_size` capped at 1500, operator opt-in, documented as bulk-optimized
+   (accepts Finding 1’s small-packet padding). Ships Lever 1+2 without a codec
+   change but is a bandwidth regression on mixed traffic if raised blindly.
+3. **Active PLPMTUD (RFC 8899)** remains a later automation layer over either.
+
+Separately: open an FEC-rate investigation for Finding 3 (suppress the +1 repair
+on clean single-symbol objects) — plausibly a larger single-core win than the
+MTU change, and independent of it.
+
+## Finding 3 resolved (2026-07-22): the +1 repair is PERMANENT for the common traffic class.
+
+Followed the pivot into the FEC-rate lever. The adaptive controller
+(`control.rs`) sets `min_ratio = 0.0` for ARQ classes but
+`min_ratio = initial_repair_ratio` for non-ARQ classes — so non-ARQ classes
+**floor above zero and can never decay to repair = 0**. The class taxonomy
+(`lib.rs::params`):
+
+| Class | DSCP | `initial_repair_ratio` | `arq` | Clean-link repair (K=1) |
+|-------|------|-----------------------:|:-----:|------------------------:|
+| Realtime | EF/CS5-7 | 0.15 | false | **1 (permanent)** |
+| Bulk | CS1/AF1x | 0.05 | **true** | 0 (decays) |
+| Default | 0 (**~all traffic**) | 0.10 | false | **1 (permanent)** |
+
+The feedback loop is live (`dataplane.rs:440-443` drives `observe_loss` per
+class from each peer `LossReport`; `FEEDBACK_INTERVAL_MS = 30`, so clean-link
+reports fire and decay works). So Bulk genuinely reaches 0 on clean links, but
+**Default and Realtime are stuck at their floor forever** —
+`repair_count(1) = max(1, ceil(1 × 0.10)) = 1`. Since Default is the class for
+essentially all traffic that doesn't set DSCP, the common case permanently emits
+1 source + 1 repair = **2 datagrams per single-symbol packet, on a pristine
+link**. The bench's “2.00 symbols/packet” is this steady state, not an artifact.
+
+This is a large, internet-universal, single-core packet-count tax that is
+independent of MTU and the wire-padding constraint. It is the recommended focus
+(user pivot, 2026-07-22). The design is a **policy tradeoff** — proactive
+protection (instant recovery, permanent cost) vs efficiency (zero clean-link
+cost, ARQ/slower recovery on loss onset) — for the non-ARQ classes that floor
+above zero. Options: (A) let Default decay to 0 on a proven-clean link and
+re-arm on the first observed loss (keep Realtime’s floor); (B) make Default
+ARQ-eligible so it decays like Bulk (biggest reach, changes recovery to
+reactive); (C) suppress repair only for K=1 objects on clean links. The MTU work
+(Findings 1–2, 4) is parked behind this.
+
+## Finding 3 — the cheap experiment (2026-07-22): the floor is load-bearing, NOT waste.
+
+Ran the experiment (set non-ARQ `min_ratio` to 0 so Default decays to 0 on a
+clean link; `crates/yip-bench/examples/fec_rate_experiment.rs` for the mechanism,
+netns for recovery). Results:
+
+**Clean-link win (measured, deterministic).** Baseline: Default holds **2.000
+symbols/packet** permanently, even after 1000 clean LossReports (the 0.10 floor
+never decays). Experiment: decays to **1.000** after ~200 clean reports (~6 s at
+the 30 ms feedback interval). So on a *sustained-clean* link the change is a real
+**2× reduction in send-path datagrams** for Default traffic.
+
+**Lossy low-rate REGRESSION (measured, netns 10% netem, low-rate ping).**
+Baseline delivers **10/10** (RESULTS.md: “RS FEC recovers the injected loss”).
+Experiment: **7/10, 9/10, 8/10** across three runs — ~20% delivered loss, i.e.
+the netem loss passing through *unrecovered*. Sparse traffic cannot sustain the
+feedback loop, so the ratio decays to ~0 between packets and each loss is
+unprotected (Default is non-ARQ — no retransmit to catch it).
+
+**High-rate is orthogonal.** `run-arq-integrity.sh` (20000×1400 B, 5% loss)
+still passes at 99.3% — but that path recovers via **ARQ** (132 retransmits,
+Bulk class), which the change never touched. It does not validate the Default
+proactive-FEC path.
+
+**The floor is a deliberate, tested invariant.** The change breaks
+`control::tests::non_arq_class_keeps_floor`, which explicitly asserts non-ARQ
+classes keep their floor. The permanent +1 repair exists precisely to protect
+the **low-rate, latency-sensitive** traffic (SSH, DNS, VoIP, gaming) that
+non-ARQ classes serve — exactly the traffic that cannot keep the controller
+armed and cannot wait for ARQ. It is buying real loss protection, not wasting
+bandwidth.
+
+### Conclusion
+
+The apparent FEC-rate inefficiency is **load-bearing**. Decaying Default to 0
+(Option A) trades a clean-link throughput win for a lossy-link recovery
+regression on exactly the low-rate traffic the class exists to protect — a bad
+trade for a VPN. Option C (suppress K=1 repair on “clean” links) has the same
+flaw: “clean” is unknowable for sparse traffic, which is when it matters. Option
+B (Default→ARQ) likely regresses low-rate recovery differently (gap-detection
+latency). **Recommendation: do not pursue the FEC-rate change; the current
+tuning is right.**
+
+Net for #59: the MTU levers are real but constrained (wire-padding; jumbo only
+on controlled paths — Findings 1, 2, 4), and the FEC-rate lever is a mirage. The
+defensible remaining work is **static configurable MTU + last-symbol wire-trim**
+(Finding 4) for operators on controlled/jumbo paths, or closing #59 with these
+findings until a controlled-path deployment makes the MTU work worth it.
+
+
+**9b knock-on:** wire-trim + a raised `symbol_size` ceiling gives the PQ
+handshake more headroom on controlled paths, but the handshake must still fit
+the safe floor (ML-KEM-512 under `OBF_MTU_BUDGET`), so 9b’s sizing decision is
+unchanged by this.