bench: per-chip TX throughput/latency harness; replace README "TX + RX" with real numbers

[josephnef]

What

Replaces the generic "TX + RX" band cells in the README "Hardware landscape" table with measured devourer TX throughput, and adds the harness that produces it.

New: tests/bench_tput.py

Per-chip TX throughput + per-frame latency, devourer vs the host kernel driver, across bands (2.4 / UNII-1 / UNII-2·3) and PSDU sizes (1500 / 3994 B). Reuses regress.py for DUT discovery, kernel bind/unbind, USB power-cycle, process hygiene and log parsers. Resumable, --quick smoke, CSV + markdown output.

Method — the clean metric: TX rate = usbmon bulk-OUT completions at the source chip (true frames-accepted rate). Counting frames at a sniffer instead measures the sniffer's RX ceiling (~336 fps here) — a trap that makes every transmitter look identical. devourer has no host-side TX backpressure (it pipelines URBs), so latency is taken from a separate non-saturating pass.

Driver/injector support

• txdemo: DEVOURER_TX_PAYLOAD_BYTES=N pads the 802.11 PSDU to N bytes (on-wire N+40; PKT_SIZE is 16-bit, so 3994 fits).
• inject_beacon.py: --size N + --max-rate (blocking AF_PACKET blaster ≈ kernel TX-completion rate).

Headline results (HT MCS7, 20 MHz, monitor injection)

Band	Part	TX dev 1500/3994 (Mbps)	TX ker 1500/3994	lat dev (µs)
2.4 GHz	RTL8812AU	46 / 58	0.9 / —	116
2.4 GHz	RTL8814AU	0.3 / 22 ⚠	1.0 / —	29
2.4 GHz	RTL8821AU	41 / 60	0.8 / —	128
UNII-1	RTL8812AU	49 / 61	5.8 / —	116
UNII-2/3	RTL8812AU	49 / 62	5.9 / —	115

• devourer direct-USB TX is 8–60× faster than kernel AF_PACKET monitor injection (it pipelines bulk-OUT URBs; the kernel path blocks on the TX ring per frame).
• The kernel monitor path can't inject 3994 B frames at all (AF_PACKET > iface MTU); devourer sends them at 58–62 Mbps.
• Throughput scales with frame size; devourer per-frame latency 17–128 µs.
• 8814 TX is the family's least reliable (high run-to-run variance — flagged ⚠).
• 8821AU 2.4 GHz on par with 8812; ~½ throughput at 5 GHz UNII (documented UNII-2/3 asymmetry).

RX — honestly, not tabulated

RX throughput cannot be measured cleanly on a 2-USB-bus bench: same-bus TX/RX pairs (8812 + 8821 share a host controller) contend, the only reliable cross-bus flooder (8812 → 8814) saturates the receiver at full TX rate, and the 8814 RX path is itself intermittent. A clean cross-bus moderate-rate flood (8812 → 8814) does receive ~3100 frames / 12 s, confirming RX works; a capacity number needs a 3-bus rig or a calibrated SDR transmitter. Full caveats in tests/README.md.

Test

cmake --build build; ctest green. Benchmark: sudo tests/bench_tput.py --quick, then sudo tests/bench_tput.py --directions tx. Run on RTL8812AU/8814AU/8821AU.

🤖 Generated with Claude Code