Add `.github/llm-review-rules.md` to define project-specific patterns
and guidelines for the LLM review bot, alongside the equivalents in
openwrt/openwrt and openwrt/luci.
The rules cover what the review routine cannot infer from the diff
alone: quilt-managed patch refresh, the generic version check and the
test.sh / test-version.sh overrides, procd init scripts, conffiles
layout, source and mirror macro preferences, and Makefile indentation
per block type.
They also set the review posture. Formatting that changes what gets
built is treated as a defect, while cosmetic style is not worth
blocking a merge over; findings the formality bot already posts are
not repeated; and CI results are a reference rather than a gate.
Patches only need to be upstream-ready when they are candidates for
upstream, since many here are OpenWrt-specific hacks with nowhere to
go. Stable branches may carry a minor LTS update that has no
counterpart on master, so not every stable-branch PR is a cherry-pick.
Signed-off-by: Josef Schlehofer <redacted> Co-Authored-By: Claude Opus 5 <redacted>