docs: give a hand-driven PR a stopping point, and name the stacked-PR blind spot

CONTRIBUTING.md

@@ -330,6 +330,17 @@ The short version for contributors:
   its window reopens.
   Do not relabel per push: each application spends a full opus session, and relabeling after
   every fix is what turned #643 into 14 review rounds.
+- **Know when to stop fixing.** The autonomous loop stops after `MAX_ITERATIONS` fix rounds
+  (`claude-pr-loop.yml` sets it) and hands the PR to a human with the open threads listed. Give
+  a hand-driven PR the same stopping point. Once a file has been round-tripped that many times
+  and what is left is wording rather than behaviour, reply once naming the findings you are
+  leaving and why, then take the PR to human review. The reviewers keep reviewing; what changes
+  is that you stop treating every finding as blocking. A finding that names a defect in
+  behaviour still earns another round.
+- **A PR based on another PR's branch is reviewed by fewer of them.** CodeRabbit auto-reviews
+  only the base branches listed in `.coderabbit.yaml`, and the deep review cannot run on a
+  branch that carries an unmerged change to its own workflow. Land a stack one PR at a time, or
+  retarget the upper PR once its base merges — the docs guide has the detail.
 - **PR titles must be scoped conventional commits** — the title becomes the squash commit and
   drives semantic-release (see [Commit Message Guidelines](#commit-message-guidelines)).
 

docs/docs/guides/getting-started/ai-development.md

@@ -39,6 +39,15 @@ Every PR gets AI review before human review:
   spend the maintainer's Claude usage). External contributors don't need them — just push.
 - **Copilot review** may also appear when the maintainers have it enabled.
 
+**A PR based on another PR's branch is reviewed by fewer of them, and nothing announces it.**
+CodeRabbit auto-reviews only the base branches listed under `auto_review.base_branches` in
+`.coderabbit.yaml`, which is the default branch alone; on any other base it posts a "Review
+skipped" notice and does not retry, and a manual mention does not override it. The Claude deep
+review is separately unavailable to a branch carrying an unmerged change to its own workflow
+file. A stacked PR can therefore reach merge having been read by Copilot alone. Either land a
+stack one PR at a time, or retarget the upper PR at the default branch once its base merges and
+let the reviewers see it there.
+
 A green AI review is not approval: a human maintainer still reviews and merges every PR, and
 the review-time requirements (Storybook story for UI changes, docs page for API changes,
 `skills/` updates for component changes) still apply.