diff --git a/.claude/skills/ci-gates/SKILL.md b/.claude/skills/ci-gates/SKILL.md
index b8c40eb05..635d12b4d 100644
--- a/.claude/skills/ci-gates/SKILL.md
+++ b/.claude/skills/ci-gates/SKILL.md
@@ -10265,3 +10265,48 @@ from a saved copy with an assert on each anchor. Cheap, deterministic, and it
 ends with a diff whose deleted lines are exactly the anchors you meant to
 replace — which is itself the check: `git diff | grep '^-'` should show nothing
 you did not intend to remove.
+
+
+## 408. The stash stack is shared across every worktree
+
+A separate `git worktree` does not get a separate stash stack -- stashes live in
+the shared `.git`. This repository had **nineteen**, from several sessions and
+branches (`WIP on w790`, `WIP on pr-1462`).
+
+The way it bites: a command of the shape `git stash -q -u; …; git stash pop -q`
+run on a tree with **nothing to stash**. The push stores nothing, and the pop
+takes the top of the SHARED stack -- somebody else's work in progress. Then
+`git add -A` committed 44 lines of another session's `compiler.rs` and its
+`FROZEN_HASH` under my name, into my pull request.
+
+**The symptom is a gate going red on a change that cannot touch it.** A one-line
+edit to a TEST file -- which `cargo build --release` does not even compile --
+turned `seal-coverage` red. A release build of the branch read **152 drifted
+seals** against master's **0**. The signal was true and the diff was lying about
+where it came from, which is why the first move is `git diff --name-status
+<base>...HEAD` rather than reading the failure.
+
+Never pair a blind `stash` with a blind `pop`. Label what you push
+(`git stash push -m`), take it back by name, or better, do not stash at all --
+that is what a second worktree is for. And a successful `pop` **drops** the
+entry: on finding somebody else's work, save the patch and return it to them
+before reverting, because the stack no longer has it.
+
+## 409. Grep before you file, and cede a locus twenty minutes old
+
+Two collisions in one pass, both avoidable by one command.
+
+An issue I opened on contradictions in the competitor table duplicated one that
+already existed -- and its chip was in the header of the dashboard I had loaded
+in full that same hour. `gh issue list --search` costs a second. A duplicate is
+not free: it splits the discussion and tells the next reader two things are open.
+
+Then a shared gate went red because of a file another session had merged twenty
+minutes earlier. I filed it, wrote the one-line repair, opened a pull request --
+and they landed the identical repair first. Mine closed as superseded, after two
+CI rounds.
+
+A file that new belongs to whoever wrote it: **file the issue and cede the fix**.
+What survived from that detour was not the patch but a measurement their PR did
+not carry -- the 152-against-0 reading above. The unique complement is worth more
+than the contested one, every time.
diff --git a/docs/now/2026-08-31-the-stash-stack-is-shared-and-a-one-line-test-fix-went-red-o.md b/docs/now/2026-08-31-the-stash-stack-is-shared-and-a-one-line-test-fix-went-red-o.md
new file mode 100644
index 000000000..f416f2ecf
--- /dev/null
+++ b/docs/now/2026-08-31-the-stash-stack-is-shared-and-a-one-line-test-fix-went-red-o.md
@@ -0,0 +1,8 @@
+# NOW -- The stash stack is shared, and a one-line test fix went red on seal-coverage (2026-08-31)
+
+## The stash stack is shared, and a one-line test fix went red on seal-coverage (Refs #2968)
+
+- A separate git worktree does not get a separate stash stack. A blind stash-then-pop on a tree with nothing to stash popped another session's work in progress, and git add -A committed 44 lines of their compiler.rs and FROZEN_HASH into my branch.
+- The symptom was a gate going red on a change that cannot touch it: a one-line edit to a test file, which cargo build --release does not compile, turned seal-coverage red. A release build of the branch read 152 drifted seals against master's 0.
+- Their work was saved to the issue verbatim before the revert, because a successful pop drops the stash entry. It has since landed on its own as #2971.
+- Two collisions in one pass: an issue duplicating one whose chip was in the dashboard I had just read in full, and a one-line repair another session landed first. A file twenty minutes old belongs to whoever wrote it -- file it and cede it.