[nc-intezra]

Adds dyode probe, the measurement tool for the streaming hardware trials (trial plan). iperf3 can't be used through a diode or a receive-only radio because it needs a TCP control channel.

What it does

• probe send sends numbered, timestamped, CRC-32-checked UDP packets at a fixed rate (--mbps or --pps). Each stream gets one thread and one UDP port (--streams N: port + i), paced on a schedule. --burst N tests how the link handles bursts.
• probe recv runs one thread per port and prints a live line every second, then a report per stream:
  • loss, judged per sequence number through a reorder window (duplicates, reordered and late packets are counted separately; end markers count losses at the tail)
  • burst-length and burst-duration histograms
  • a Gilbert-Elliott fit, plus the tc netem loss gemodel p r line that replays it on the bench
  • delay and RFC 3550 jitter
  • a RaptorQ what-if (--fec-sim 200ms:1,...): what a span T and repair ratio r would have left unrecovered (systematic code: only source packets lost in failed groups count)
  • output: --log writes a per-packet CSV, --json a full report with per-port counters (damaged/foreign datagrams, buffer size actually granted), and --keep keeps listening for field soaks
• probe analyze LOG replays a log, for example with other --fec-sim settings.
• Docs: docs/PROBE.md (usage, how to read the report, notes per hardware class). CHANGELOG (Unreleased) and README are updated.

New file src/probe.rs; main.rs gains the probe command and help text. Nothing else changes; no new dependencies (it uses crc32fast, libc, getrandom, serde_json, which are already in the tree).

Testing (sandbox, offline build against the 2.0.2 vendored sources)

• cargo test: 64 pass (8 new: codec/CRC, FEC-spec parsing, bursts + Gilbert, reorder/duplicate/late, tail loss with and without the end marker, late joiner, FEC simulation, and a 2-stream loopback run including a log replay). Clippy and rustfmt are clean.
• 4 streams at 2 Gbit/s total over loopback: 1,428,572 packets with no loss on a 2-core VM.
• Gilbert relay (a userspace UDP relay dropping with p = 1%, r = 25%; tc isn't available in the sandbox): the probe measured p = 0.94%, r = 24.95%, mean burst 4.01 packets (theory: 4). Its loss matched the relay's own count exactly (3.620%).
• Low rate (5 pkt/s, 64 bytes, --burst 3), Ctrl-C on an idle receiver, and the usage errors were checked by hand.
• Not yet run on diode or radio hardware.

Notes

• Probe packets carry no HMAC. It's a trial tool; don't point it at production DYODE ports. Default port 7000 stays clear of the lidi ports, and DYODE's nftables rules only restrict lidi ports.
• The sender busy-waits the last ~50 µs before each packet for accurate pacing, so expect one busy core per stream at high rates.
• This PR and PR 2.1.0: local API (phase 1) and the input box's status through the diode #6 both edit CHANGELOG.md and src/main.rs, so expect a small conflict to resolve when merging the second one.

---

Generated by Claude Code