docs(research): #59 MTU/FEC-rate investigation findings + measurement tools

[vxfemboy]

Preserves the measure-first investigation of #59 (throughput lever 4c) as durable institutional memory. No production code — a research doc plus two reusable bench examples alongside the existing pipeline_profile.

Why

The #59 spike found both "cheap wins" are illusions once measured, and that's worth recording so the next person doesn't re-litigate it. Full analysis: docs/research/2026-07-22-mtu-packetization-findings.md. Headlines:

• Wire-padding constraint: every FEC symbol is zero-padded to full symbol_size on the wire (fec.rs "last zero-padded"; reassembler rejects data.len() != symbol_size; framer sends it untrimmed). So a blind symbol_size increase is a bandwidth regression for small packets.
• Two MTU levers, bounded: raising symbol_size to ~1500 helps only per-symbol costs and only for packets that currently split; jumbo inner MTU helps everything but only where the whole path carries jumbo. The issue's "reduces every per-packet cost proportionally" holds only for the jumbo lever.
• The FEC-rate lever is load-bearing, not waste. The Default class (DSCP 0 = nearly all traffic) permanently emits 1 source + 1 repair per single-symbol packet because its non-ARQ min_ratio floors at 0.10. Tempting to decay it to 0 — but the experiment measured the cost: clean-link datagrams halved (2.000 → 1.000), but 10% netem low-rate ping recovery regressed 10/10 → 7/10, 9/10, 8/10 (sparse traffic can't keep the controller armed; losses go unrecovered). The floor protects exactly the low-rate latency-sensitive traffic the class exists for. Change rejected.

Contents

• docs/research/2026-07-22-mtu-packetization-findings.md — the full write-up.
• crates/yip-bench/examples/mtu_symbol_sweep.rs — symbols/packet + outer bytes across inner-size × symbol_size (real Transport::encode).
• crates/yip-bench/examples/fec_rate_experiment.rs — whether the Default class decays on a clean link.

No behavior change; #59 updated with the same conclusion.