Startup time: all-chip devourer-vs-kernel benchmark (host-kernel cells) + docs/startup-time.md

[josephnef]

What

Closes the README "Startup time" gap: the "faster than the kernel driver" claim was only ever substantiated for Jaguar1 (vs aircrack in the VM). This PR measures every supported generation on the host against the vendor drivers built from reference/, publishes the numbers + methodology in a new docs/startup-time.md, and shrinks the README section to one line + link.

Results (true cold plug, VBUS-cycled per rep, median of 2, ch 6)

chip	devourer	vendor kernel driver
RTL8814AU	5.7 s	6.9 s (88XXau)
RTL8821AU (T2U+)	2.2 s	5.3 s (88XXau)
RTL8822BU (T3U)	7.2 s	11.1 s (rtl88x2bu)
RTL8821CU	3.4 s	5.1 s (8821cu)
RTL8812CU	1.2 s	3.1 s (rtl88x2cu)
RTL8812EU	1.0 s	3.0 s (rtl88x2eu)

TX-ready tracks RX-ready within ~1%. The doc also carries the warm (authorized-toggle) table — the vendor drivers pay full calibration only from true cold, so warm-only benchmarks understate the first-plug gap (and the 8821AU kernel even edges ahead by ~60 ms warm; stated honestly in the doc).

The 8812AU rows are absent: the bench unit doesn't come up RX-functional on first init from true cold under either driver — #205.

Harness (tests/bench_init.py)

• --kernel-host — kernel cells via insmod of the reference/ vendor .ko on the host (netdev → monitor → first frame, same stage names as the VM path; KMOD_FOR_VIDPID maps PIDs for all three generations). VM mode remains for driver-behaviour comparisons but is documented as unfit for startup timing (virtualized USB adds latency to every stage).
• --traffic-from PID — dedicates a plugged non-DUT adapter to a devourer beacon flood on the bench channel. In an RF-quiet room there are zero ambient frames and every first-frame marker (devourer and kernel tcpdump alike) starves; with a ~500 fps flood, first-frame latency measures RX-path readiness instead of beacon-interval luck.
• udev-rename-tolerant iface configuration in the kernel cell (the 88XXau netdev renames right after registration).
• tests/regress.py: SUPPORTED_DUTS learns the Jaguar2 PIDs (2357:012d, 0bda:b82c, 0bda:c811) so DUT discovery covers them.

Notes

• All four vendor comparators (rtl8812au/88XXau, rtl88x2bu, rtl88x2cu, rtl88x2eu, 8821cu) build against the host 6.18 kernel with local compat patches in the (gitignored) reference/ trees — including aircrack-lineage 88XXau, so no VM is needed anywhere in the table.
• Methodology traps documented in the doc: warm-vs-cold chip state, and the traffic-source requirement.
• 36+ harness cells all 3/3–5/5 green across two full runs; .ko debug-strip verified irrelevant to kernel.probe (the T3U's 5.5 s probe is genuine vendor work, not insmod cost).

🤖 Generated with Claude Code