TLDR
I have a recipe, based on @arinc9's instructions that will produce a kernel package with the same magic number as the official release, and does not require patches:
git clone https://git.openwrt.org/openwrt/openwrt.git
cd openwrt
git checkout v21.02.1
rm -rf package/kernel/{mac80211,ath10k-ct,mt76,rtl8812au-ct}
git checkout d1100c76b33ff68c6db0f5fa31a26532bdbb15c4 -- package/kernel/{mac80211,ath10k-ct,mt76,rtl8812au-ct}
rm -f package/kernel/mac80211/patches/ath/551-ath9k_ubnt_uap_plus_hsr.patch
./scripts/feeds update -a -f && ./scripts/feeds install -a -f
./scripts/feeds uninstall batman-adv
wget 'https://downloads.openwrt.org/releases/21.02.1/targets/mvebu/cortexa9/config.buildinfo'
cp config.buildinfo .config
make defconfig
make -j8 download && make -j8 BUILD_LOG=1
Detailed Explanation regarding the kernel vermagic hash
batman-adv fails to compile because of header mixup involving a mac80211-installed header. It could be trivially patched, but the purpose of the build is to allow openwrt package installation, so I don't imagine it would be much of a problem.
You can customize the configuration with make menuconfig, but beware that some packages will change the kernel config, changing the vermagic number. The value is computed at the beginning of make target/compile. So if you want to check if the configuration changes its value, run make target/compile, and wait a bit for the sources to be unpacked, and the .config files to be generated. It is saved to build_dir/target-arm_cortex-a9+vfpv3-d16_musl_eabi/linux-mvebu_cortexa9/linux-5.4.154/.vermagic.
Inclusion or exclusion of a package may change the hash. Notably, excluding the ath10k-ct, and other wireless driver packages will change the hash. The same happens if you don't install a feed. Luckily, omitting batman-adv did not alter the hash, so I took the shortcut and just removed it.
I'm just speculating, as my wrt3200s are used in production, and I can't just test things at will, but it may be possible to use the official image, and then just force-downgrade the mac80211 packages (kmod-cfg80211, kmod-mwifiex-sdio, kmod-mac80211) from a local build that shares the same magicver.
Edit: I'm adding kmod-mwlwifi to the list of packages that need to come from the local build. One of the functions has apparently changed name, judging by the difference when running strings /lib/modules/5.4.154/mwlwifi.ko:
--- strings 2021-11-17 16:07:23.206678891 -0300
+++ strings.new 2021-11-17 16:07:48.576678480 -0300
... minor binary stuff skipped.
@@ -934,10 +934,10 @@
__dev_kfree_skb_any
devm_ioremap_resource
filp_close
+ieee80211_channel_to_frequency
ieee80211_unregister_hw
cancel_work_sync
ieee80211_beacon_get_tim
-ieee80211_channel_to_freq_khz
skb_copy
__aeabi_uidivmod
ieee80211_hdrlen
It doesn't help that the symptoms are intermittent. I wish there was a more deterministic way to check for a problem.