diff --git a/.changeset/release-dispatch.md b/.changeset/release-dispatch.md
new file mode 100644
index 0000000..ab15093
--- /dev/null
+++ b/.changeset/release-dispatch.md
@@ -0,0 +1,15 @@
+---
+"resilix": patch
+---
+
+`release.yml` can now be triggered manually (`workflow_dispatch`).
+
+The publish fires on a push to `main`, and a push gives exactly one chance to run it. During the
+2026-08-26 GitHub Actions outage no runs were created at all, which means a release commit landing
+in that window would have consumed its changesets and left the version on `main` with nothing
+published — recoverable only by pushing another commit purely to fire the event. Exposing the
+dispatch is safe: `changeset publish` is a no-op for a version already on the registry.
+
+`CONTRIBUTING.md` also now states when the branch-protection bypass is legitimate — a platform
+outage or a production incident, with the local checks run first — and that it never extends to
+force-push, branch deletion or tag immutability.
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 03be0ab..9e26aa3 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -3,6 +3,15 @@ name: Release
 on:
   push:
     branches: [main]
+  # Re-triggerable by hand. The publish is driven by a push to main, and a push produces exactly
+  # one chance to run it: if Actions is unavailable when the release commit lands — as during the
+  # 2026-08-26 major outage — no run is ever created, changesets has already consumed its
+  # changesets, and the version sits on main with nothing on npm. Without this the only recovery
+  # is pushing another commit to main purely to fire the event.
+  #
+  # Safe to expose: `changeset publish` is a no-op for a version that is already on the registry,
+  # so an accidental dispatch cannot republish or overwrite anything.
+  workflow_dispatch:
 
 permissions:
   contents: write
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index 6e5d864..a288f80 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -131,6 +131,35 @@ maintainer.
 pointing at the commit whose provenance attestation says it built `resilix@0.5.0`; a moved tag
 makes that attestation a lie.
 
+### When the admin bypass is legitimate
+
+The bypass on `main: PR + green checks` exists so a solo maintainer is never locked out. It is not
+a convenience, and "the checks are slow" is not a reason. There are two:
+
+1. **A platform outage means the checks cannot run at all.** On 2026-08-26 GitHub Actions was in
+   `major_outage` for over an hour and created zero runs repo-wide. A PR in that window is
+   `BLOCKED` forever with nothing to wait for. Check
+   [githubstatus.com](https://www.githubstatus.com) before concluding it is your branch — the
+   first hour of that outage was spent suspecting a workflow-file edit and then a push credential,
+   both wrong.
+2. **A production incident** where the fix must land faster than a full matrix run.
+
+In both cases, **run the checks locally first and say so in the merge**:
+
+```bash
+pnpm verify        # lint, paths, typecheck, coverage, build, package, smoke, docs
+pnpm test:compat   # opossum's own suite — not in verify, it needs a network fetch
+pnpm test:perf
+```
+
+`pnpm verify` covers everything the nine required checks do except the multi-runtime matrix, which
+only CI can provide. If you bypass, the next green CI run on `main` is what actually confirms it —
+watch it.
+
+The bypass does **not** extend to `main: no force-push, no deletion` or to tag immutability. Those
+have no bypass actor at all, deliberately: they are the rules whose violation destroys work rather
+than merely making a mess.
+
 ### Required checks, and the two that are deliberately absent
 
 Required: `build`, `node 18`, `node 20`, `node 22`, `node 24`, `bun`, `deno`,