The device tree asks ath11k for the Fairphone_Gen_6 board data entry.
Upstream linux-firmware's WCN6750 board-2.bin has no Fairphone entry, so
the driver silently falls back to the generic board-id 255 data: the FP6
Plus does not associate on it at all (user report 2026-10-14), and the FP6
keeps its channel list but its scans no longer find the 5 GHz networks
around it (two scans found 0 and 1 5 GHz BSSes against 14 with the
device's data, the home AP at -37 dBm among the missing), so it
associates on 2.4 GHz only (dev phone A/B). The stock modem partition
carries the device's own calibration as image/qca6750/bdwlan.elf.
New manifest directive ath11k-bdf: copy that ELF and wrap it into a
container under /lib/firmware/updates (searched first by the loader) that
is upstream's board-2.bin with every entry kept plus one BOARD entry under
the requested name. Verified structurally (ELF magic, size, the container
magic), no hash pin: the ELF belongs to the stock modem image and can
change with it. The sha256 of the base is kept beside the
override: a changed base (a linux-firmware upgrade) rebuilds it at the
next run, and a rebuild that reproduces the existing file installs
nothing; an entry upstream ships under the same name is taken first by the
driver, so the appended copy then goes unused. The fragment has no
rebind line: the service is ordered before udev's coldplug, which loads
the driver, so a fresh install builds the container before the first
probe (dev phone, the override removed and rebooted: built at 7.2 s,
coldplug from 7.3 s, the driver's board-data load at 8.7 s, the first
association on 5 GHz), and an installed phone gets it at its next boot; a
re-probe from a package hook would restart wifi in the middle of the
upgrade.
zstd joins the dependencies to read the compressed base. Dev phone: the
built container is byte-identical to the one hand-made from the stock
image by the reporter (sha256 186202ee...); the hook's call with the
override absent built it over a wifi session that stayed up, a second
call installed nothing and a --refresh call left the container unchanged.
