docs: catch the README and both primers up with two weeks of work

[josephnef]

47 commits landed since 12 July — the Kestrel Wi-Fi 6 HAL and its HE features,
six firmware-offload PRs, adaptive channel migration, LDPC as a first-class
primitive — and none of it reached the pages a newcomer actually reads.

What had drifted, all verified before writing:

• six docs linked from nowhere in the README (he-trigger-ul,
  he-extended-range, 8852c-quirks, and the whole channel-migration trio —
  an entire subsystem with two demos, undiscoverable)
• four demos missing from the binaries table (chanmig, chanscout,
  dwelltx, kestrelprobe)
• the CMake option list predating DEVOURER_KESTREL_8852B/C and DEVOURER_PCIE
• "Why devourer" not mentioning 802.11ax at all — the hardware table carried
  the 8852B/C rows, but the pitch had no HE, no extended range, no trigger
  vocabulary
• the RF primer teaching the OFDM comb in §1 and never saying a Wi-Fi 6 radio
  slices it between users; §3 introducing FEC and never showing what it buys

Four new figures

House style throughout — one generator per figure in tools/, monitor_style
palette and chrome, save_gif(colors=48), Pillow only.

figure	where	what it shows
he_ofdma_trigger.gif	RF primer §11	a grant table composed on the host, the Trigger frame built around it, and a monitor reading every parameter back off the air unchanged
he_extended_range.gif	RF primer §12	the ER SU / DCM ladder — three ~3 dB rungs, each paid for in rate
ldpc_waterfall.gif	RF primer §3	the coding-gain waterfall, and which receivers can actually decode it
fw_offload.gif	driver primer §14	H2C as a speed technique rather than a message pipe

Each figure is scoped to what was measured, and each states its limit in the
same breath as its win:

• The trigger figure shows the validated round trip — command in, frame out,
  parameters verified — which is what makes the adapter an instrument for 11ax
  work. One footer line carries the boundary: generating the schedule is
  available, owning it needs firmware these client parts don't carry, because
  only a firmware-scheduled trigger arms the receiving MAC's SIFS window.
• The LDPC curve shape is labelled illustrative on its face; the measured
  quantity is the ~3 dB gap at the 10 %-delivery crossing. The truth table
  beside it is bench-derived from AdapterCaps — the 8821A misses VHT-LDPC,
  the 8814A decodes but reports no per-frame flag.
• The offload figure ends on the honest half: offload frees the host, not the
  radio. The ~1.5 ms settle floor does not care who wrote the registers, so
  time-to-usable goes ~5 ms → ~1.7 ms, not to zero.

README

A Start-here reading order at the top of "Going deeper" (there are ~30 links
below it), Wi-Fi 6 and channel-migration value bullets, the missing demos and
build options, and homes for all six orphans.

Two stale facts corrected

• Binary sizes were measured, not guessed. The page claimed ~1.0 MB /
  ~2.6 MB; it is 1.60 MB for an 8812AU-only rxdemo against 6.30 MB
  all-on, with 4.17 MB at 11ac-only — the verbatim-vendored halbb/halrf
  plane alone accounts for ~2.1 MB. CLAUDE.md carried the same stale pair and
  is fixed too.
• Termux → PixelPilot as the example behind the no-kernel-driver claim: an
  FPV ground station on an unmodified phone is the real demonstration.

Verification

Checked as a loop, not by eye: zero orphan docs; the figure counts written in
both primer intros and the README equal the actual ![...] counts (15 and 14);
no missing image files; 46/46 ctest green.
docs/experiments/issue-330-rx-ring-starvation.md is deliberately left
unlinked — it is an ongoing experiment, not a feature.

🤖 Generated with Claude Code