diff --git a/.github/shared-files.sha256 b/.github/shared-files.sha256
index 9af6918..de628f5 100644
--- a/.github/shared-files.sha256
+++ b/.github/shared-files.sha256
@@ -68,5 +68,5 @@ fcb6cd0e2a71205b918f545f2680d378c790b774b9852129c5f9c7f3f8eb9860  .github/workfl
 09230b04c5f14bdda50a0c862152ff197a656cbc3fd0d110899c5f8150f8aa9a  .github/workflows/publish.yml#check-snapshot
 51e987d59efd9887b43e345edc0e0f3db21686879e3ca6ea9f0fdbc493d37121  .github/workflows/publish.yml#check-tag
 a8637f88765da759a6eaba243d0d0d695053174df581ccbbcc5e5f3a475fc9e9  .github/workflows/publish.yml#verify-signing-key-gradle
-f20f0a991ce68fa1dcc8ef4daffc5e5dea2093d9b578aa41d17f18a56555d79f  .github/workflows/publish.yml#github-snapshot
+83a1288c55d35e4d3972e013009577fb594a9a5fb8b9b7c96a7f6a19db20d8d1  .github/workflows/publish.yml#github-snapshot
 ce90d90fcb1f2de35075b91c74411e9eb7d3b43e7965fd592744e914fd04c5db  .github/workflows/publish.yml#github-release
diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index 57c1925..7967479 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -528,23 +528,58 @@ jobs:
         # upload on purpose: a failed upload must not leave the release emptier than it was. And only
         # here: a versioned v* release is history, github-release never deletes anything. Every
         # deletion is printed by name, so the job log is the audit trail.
+        #
+        # Asset by asset, never the release. The first form of this job (2026-05-07) deleted the whole
+        # `snapshot` release with its tag and recreated it; on 2026-05-10 the runs of several main
+        # pushes minutes apart deleted it under each other, and every job of each burst but the last
+        # failed in that step with its artifacts present. The job then went with the Central
+        # migration and came back as `view || create` + `upload --clobber` without any cleanup, which
+        # is how the backlog grew (workspace/policies/fat-jar-release-assets.md has the runs). Hence
+        # three guards: an empty upload is refused (every asset would count as stale; the download
+        # and upload steps above fail first, the guard keeps it so), an asset that is already gone
+        # counts as removed (two overlapping dispatches may both pick it), and a failed deletion is
+        # reported and the rest still tried -- the assets of this upload are on the release either
+        # way, so this is "attach first, then red" like the signature check.
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         run: |
           set -euo pipefail
           keep="$(ls -1 snapshot-assets | sort)"
-          current="$(gh release view snapshot --repo ${{ github.repository }} --json assets --jq '.assets[].name' | sort)"
+          if [ -z "$keep" ]; then
+            echo "::error::refusing to prune: this run uploaded nothing, so every asset of the release would count as stale"
+            exit 1
+          fi
+          list_assets() {
+            gh release view snapshot --repo ${{ github.repository }} --json assets --jq '.assets[].name' | sort
+          }
+          current="$(list_assets)"
           stale="$(comm -13 <(printf '%s\n' "$keep") <(printf '%s\n' "$current"))"
           if [ -z "$stale" ]; then
             echo "nothing to remove: the release holds exactly the $(printf '%s\n' "$keep" | wc -l) asset(s) of this upload"
             exit 0
           fi
           echo "removing $(printf '%s\n' "$stale" | wc -l) asset(s) that are not part of this upload:"
+          failed=0
           while IFS= read -r name; do
             [ -n "$name" ] || continue
             echo "  deleting $name"
-            gh release delete-asset snapshot "$name" --repo ${{ github.repository }} --yes
+            if gh release delete-asset snapshot "$name" --repo ${{ github.repository }} --yes; then
+              continue
+            fi
+            if ! remaining="$(list_assets)"; then
+              echo "::error::could not delete $name, and could not re-read the release to tell whether it is gone"
+              failed=$((failed + 1))
+            elif printf '%s\n' "$remaining" | grep -qxF -- "$name"; then
+              echo "::error::could not delete $name"
+              failed=$((failed + 1))
+            else
+              echo "  $name was already gone"
+            fi
           done <<< "$stale"
+          if [ "$failed" -ne 0 ]; then
+            echo "::error::$failed asset(s) could not be removed; the $(printf '%s\n' "$keep" | wc -l) asset(s) of this upload are on the release"
+            exit 1
+          fi
           echo "kept the $(printf '%s\n' "$keep" | wc -l) asset(s) of this upload"
       - name: Fail if anything was attached unsigned
         # After the upload on purpose: the assets must exist even when the signature does not.
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 4baf127..f4b128e 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -79,8 +79,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   signature check existed) and three SpotBugs plugin jars (`fb-contrib-7.6.4.jar`, `fb-contrib-7.7.4.jar`,
   `findsecbugs-plugin-1.14.0.jar`) an old collect step uploaded from `target/spotbugs/` in 2026-05, 5 of its
   11 assets. `github-snapshot` now deletes every asset
-  the run did not upload, after the upload and each printed by name; versioned releases are never
-  pruned, and the next snapshot run clears the backlog by itself. The collect steps copy
+  the run did not upload, after the upload and each printed by name; it refuses an empty upload,
+  treats an asset that is already gone as removed and fails only after trying every deletion (asset
+  by asset on purpose: the job's first form deleted and recreated the release, and failed in every
+  burst of main pushes in May 2026); versioned releases are never pruned, and the next snapshot run
+  clears the backlog by itself. The collect steps copy
   `<artifactId>-<version>.pom` + `.pom.asc` as well (Central had them, the release did not), and
   the signature check counts every attachable type (`*.jar`, `*.aar`, `*.pom`) instead of jars
   alone. Shared job text with the two sibling repositories (`shared-files.sha256`), the convention
diff --git a/CLAUDE.md b/CLAUDE.md
index 4302c48..6bcd510 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -251,8 +251,10 @@ exists here. The attach jobs are shared text with the two sibling Maven-only rep
 signature check counts every attachable type (`*.jar`, `*.aar`, `*.pom`), and `github-snapshot` ends
 by deleting every asset of the `snapshot` pre-release that the run did not upload, each printed by
 name -- `gh release upload --clobber` never removed anything, so the release accumulated every past
-version. Versioned releases are never pruned. See the policy's "Attach first, then go red" and "The
-`snapshot` pre-release holds exactly the current upload".
+version. The prune refuses an empty upload, treats an asset that is already gone as removed and
+fails only after trying every deletion; versioned releases are never pruned. See the policy's
+"Attach first, then go red" and "The `snapshot` pre-release holds exactly the current upload", which
+also records the delete-and-recreate form of this job (May 2026) and why it is gone.
 
 ## Shared files and the release gate (`shared-files` job)