From 21238381d41818c3c88026af90c9d4564a0c4df8 Mon Sep 17 00:00:00 2001
From: Aurelio <19254254+Aureliolo@users.noreply.github.com>
Date: Sun, 24 May 2026 16:55:46 +0200
Subject: [PATCH 1/3] ci: retry cosign sign on transient GHCR/Rekor failures

---
 .github/actions/publish-apko-base/action.yml  | 13 +--
 .../actions/publish-image-loaded/action.yml   | 15 ++--
 .github/scripts/cosign_sign_with_retry.sh     | 87 +++++++++++++++++++
 3 files changed, 103 insertions(+), 12 deletions(-)
 create mode 100644 .github/scripts/cosign_sign_with_retry.sh

diff --git a/.github/actions/publish-apko-base/action.yml b/.github/actions/publish-apko-base/action.yml
index 7cfbe90db4..7d5629f9f5 100644
--- a/.github/actions/publish-apko-base/action.yml
+++ b/.github/actions/publish-apko-base/action.yml
@@ -157,6 +157,12 @@ runs:
       if: steps.cosign_2.outcome == 'failure'
       uses: sigstore/cosign-installer@6f9f17788090df1f26f669e9d70d6ae9567deba6 # v4.1.2
 
+    # Wrap the cosign call in cosign_sign_with_retry.sh so a GHCR /
+    # Rekor / Fulcio 5xx during signature upload retries on the same
+    # budget every docker push already gets via docker_push_with_retry.sh.
+    # The helper preserves the `createLogEntryConflict` idempotency
+    # branch (already-signed digest -> success) and only retries on
+    # the shared transient regex.
     - name: Sign image
       shell: bash
       env:
@@ -165,11 +171,8 @@ runs:
       run: |
         set -euo pipefail
         if [ -z "$DIGEST" ]; then echo "::error::No digest"; exit 1; fi
-        if ! cosign sign --yes "ghcr.io/aureliolo/synthorg-${IMAGE_NAME}-base@${DIGEST}" 2>&1 | tee /tmp/cosign-out; then
-          if grep -q 'createLogEntryConflict' /tmp/cosign-out; then
-            echo "::notice::Image already signed -- skipping"
-          else cat /tmp/cosign-out; exit 1; fi
-        fi
+        bash "${GITHUB_WORKSPACE}/.github/scripts/cosign_sign_with_retry.sh" \
+          "ghcr.io/aureliolo/synthorg-${IMAGE_NAME}-base@${DIGEST}"
 
     # Confirm the input tag still resolves to the digest cosign just
     # signed and that the cosign signature artifact is reachable on
diff --git a/.github/actions/publish-image-loaded/action.yml b/.github/actions/publish-image-loaded/action.yml
index 86e4d7c1cc..f17fd2ec47 100644
--- a/.github/actions/publish-image-loaded/action.yml
+++ b/.github/actions/publish-image-loaded/action.yml
@@ -240,6 +240,12 @@ runs:
       if: steps.cosign_2.outcome == 'failure'
       uses: sigstore/cosign-installer@6f9f17788090df1f26f669e9d70d6ae9567deba6 # v4.1.2
 
+    # Wrap the cosign call in cosign_sign_with_retry.sh so a GHCR /
+    # Rekor / Fulcio 5xx during signature upload retries on the same
+    # budget every docker push already gets via docker_push_with_retry.sh.
+    # The helper preserves the `createLogEntryConflict` idempotency
+    # branch (already-signed digest -> success) and only retries on
+    # the shared transient regex.
     - name: Sign image
       shell: bash
       env:
@@ -251,13 +257,8 @@ runs:
           echo "::error::Push step did not produce a digest -- cannot sign image"
           exit 1
         fi
-        if ! cosign sign --yes "ghcr.io/aureliolo/synthorg-${IMAGE_NAME}@${DIGEST}" 2>&1 | tee /tmp/cosign-out; then
-          if grep -q 'createLogEntryConflict' /tmp/cosign-out; then
-            echo "::notice::Image already signed in transparency log -- skipping"
-          else
-            exit 1
-          fi
-        fi
+        bash "${GITHUB_WORKSPACE}/.github/scripts/cosign_sign_with_retry.sh" \
+          "ghcr.io/aureliolo/synthorg-${IMAGE_NAME}@${DIGEST}"
 
     # Belt-and-braces: confirm every tag pushed in this step resolves
     # (in the registry, NOW) to the same digest cosign just signed. A
diff --git a/.github/scripts/cosign_sign_with_retry.sh b/.github/scripts/cosign_sign_with_retry.sh
new file mode 100644
index 0000000000..58997ef280
--- /dev/null
+++ b/.github/scripts/cosign_sign_with_retry.sh
@@ -0,0 +1,87 @@
+#!/usr/bin/env bash
+# Retry cosign sign on transient registry / Rekor / Fulcio failures.
+#
+# `cosign sign` against a published digest is idempotent: signing the
+# same digest twice either succeeds again or hits a Rekor
+# `createLogEntryConflict` (already-logged) response, which callers
+# treat as success. Transient GHCR/Rekor/Fulcio errors (5xx, 429, TLS
+# handshake stalls, connection resets) almost always settle inside the
+# next attempt window, so a bounded retry turns a noisy infra blip
+# into a green run instead of failing the whole Docker workflow.
+#
+# Usage:
+#   cosign_sign_with_retry.sh <ref>
+#     <ref> is the full image reference, e.g.
+#     ghcr.io/aureliolo/synthorg-sandbox-base@sha256:abc...
+#
+# Behaviour:
+#   - Captures combined stdout+stderr of `cosign sign --yes <ref>`.
+#   - On exit 0: prints captured output, exits 0.
+#   - On exit non-0:
+#     * Output contains `createLogEntryConflict` -> already signed,
+#       emit `::notice::` and exit 0 (preserves the idempotency branch
+#       the inline shell blocks used to carry).
+#     * Output matches the shared transient regex sourced from
+#       docker_push_with_retry.sh (single source of truth for
+#       "is this a registry-side flake?") -> warn + sleep + retry
+#       with exponential backoff.
+#     * Otherwise -> non-transient cosign / Rekor / Fulcio error,
+#       surface output and exit with cosign's exit code.
+#   - Final attempt with no terminal classification: surface all
+#     output and exit with cosign's exit code.
+set -euo pipefail
+
+REF="${1:?usage: cosign_sign_with_retry.sh <ref>}"
+
+# Same regex the docker push helper uses; `--print-transient-re` keeps
+# both scripts in lockstep so a new transient signature added in one
+# place automatically protects every signing call too.
+SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
+TRANSIENT_RE="$("$SCRIPT_DIR/docker_push_with_retry.sh" --print-transient-re)"
+
+# 4 attempts, backoff 15s -> 30s -> 60s = ~1m45s of wait in the worst
+# case before the final attempt. Matches docker_push_with_retry.sh so
+# cosign rides through GHCR's typical 30-90s unicorn windows under the
+# same budget the push step already does.
+ATTEMPTS=4
+BACKOFF=15
+
+for ((i = 1; i <= ATTEMPTS; i++)); do
+  out=""
+  rc=0
+  out="$(cosign sign --yes "$REF" 2>&1)" || rc=$?
+  if [ "$rc" -eq 0 ]; then
+    printf '%s\n' "$out"
+    exit 0
+  fi
+
+  # Idempotency branch: a re-sign that lost the createLogEntry race is
+  # success, not a transient error. Check this BEFORE the regex so
+  # attempt 1 -> 5xx -> attempt 2 -> conflict resolves cleanly.
+  if printf '%s' "$out" | grep -q 'createLogEntryConflict'; then
+    printf '%s\n' "$out"
+    echo "::notice::Image ${REF} already signed -- skipping"
+    exit 0
+  fi
+
+  if [ "$i" -eq "$ATTEMPTS" ]; then
+    printf '%s\n' "$out"
+    echo "::error::cosign sign ${REF} failed after ${ATTEMPTS} attempts (last exit ${rc})" >&2
+    exit "$rc"
+  fi
+
+  if printf '%s' "$out" | grep -qiE "$TRANSIENT_RE"; then
+    printf '%s\n' "$out" >&2
+    echo "::warning::cosign sign ${REF} hit transient error (attempt ${i}/${ATTEMPTS}, rc=${rc}); sleeping ${BACKOFF}s before retry" >&2
+    sleep "$BACKOFF"
+    BACKOFF=$((BACKOFF * 2))
+    continue
+  fi
+
+  # Non-transient: surface output and bubble up immediately. Auth
+  # denials, malformed digests, Rekor schema rejections, etc. will
+  # never improve on a retry.
+  printf '%s\n' "$out"
+  echo "::error::cosign sign ${REF} failed with non-transient error (exit ${rc}); not retrying" >&2
+  exit "$rc"
+done

From edb4a83e8b86d7d06776135a0ef344e991e77ce4 Mon Sep 17 00:00:00 2001
From: Aurelio <19254254+Aureliolo@users.noreply.github.com>
Date: Sun, 24 May 2026 17:13:11 +0200
Subject: [PATCH 2/3] ci: strip pre-existing comment rot in
 publish-image-loaded action

---
 .../actions/publish-image-loaded/action.yml   | 34 +++++++++----------
 1 file changed, 16 insertions(+), 18 deletions(-)

diff --git a/.github/actions/publish-image-loaded/action.yml b/.github/actions/publish-image-loaded/action.yml
index f17fd2ec47..df63403f88 100644
--- a/.github/actions/publish-image-loaded/action.yml
+++ b/.github/actions/publish-image-loaded/action.yml
@@ -14,16 +14,15 @@ description: >-
   the expected names). This action then runs the security-critical
   publish path that must be byte-identical across every caller.
 
-  Background on the digest-pinning: a previous version used
-  `${REPO}:sha-X-amd64` tag refs in `docker manifest create`.
-  `docker manifest create` re-resolves tag references against the
-  registry on every call, so when a concurrent main-push run overwrote
-  the per-arch tag between two iterations of a tag-push run's loop,
-  the manifest lists for different destination tags ended up with
-  different per-arch digests. cosign signed only the last one, leaving
-  the user-facing version tag unsigned. Capturing the per-arch digest
-  immediately after push and referencing manifest list members by
-  `@sha256:digest` makes every iteration produce byte-identical bytes.
+  Why digest-pinning: `docker manifest create` re-resolves tag
+  references against the registry on every call. When a concurrent
+  main-push run overwrites a per-arch tag between two iterations of a
+  tag-push run's loop, the manifest lists for different destination
+  tags end up with different per-arch digests, and cosign signs only
+  the last one, leaving the user-facing version tag unsigned.
+  Capturing the per-arch digest immediately after push and referencing
+  manifest list members by `@sha256:digest` makes every iteration
+  produce byte-identical bytes, immune to concurrent tag mutations.
 
   Caller responsibilities:
     - `packages: write`, `id-token: write`, `attestations: write`
@@ -61,11 +60,11 @@ outputs:
 runs:
   using: composite
   steps:
-    # GHCR's token endpoint occasionally exceeds the docker client's default
-    # 30s deadline (the failure mode that took down main 2026-05-13). Drive
-    # `docker login` directly through nick-fields/retry so the login is
-    # re-attempted on `context deadline exceeded`. Token via env + stdin so
-    # it never lands on a command line.
+    # GHCR's token endpoint occasionally exceeds the docker client's
+    # default 30s deadline. Drive `docker login` directly through
+    # nick-fields/retry so the login is re-attempted on `context
+    # deadline exceeded`. Token via env + stdin so it never lands on
+    # a command line.
     - name: Log in to GHCR
       uses: nick-fields/retry@ad984534de44a9489a53aefd81eb77f87c70dc60 # v4.0.0
       env:
@@ -109,9 +108,8 @@ runs:
     #
     # Note: the manifest-list digest below is still read via
     # `docker buildx imagetools inspect` because `docker manifest
-    # push` stdout is unreliable across Docker versions (PR #1650
-    # documented this: some versions emit the digest, some only
-    # print `Created manifest list`).
+    # push` stdout is unreliable across Docker versions: some emit
+    # the digest, some only print `Created manifest list`.
     - name: Push per-arch images and assemble manifest list
       id: push
       shell: bash

From 67877f8f68fad67fbfd70ebfd6bb7b7da555d08d Mon Sep 17 00:00:00 2001
From: Aurelio <19254254+Aureliolo@users.noreply.github.com>
Date: Sun, 24 May 2026 17:27:25 +0200
Subject: [PATCH 3/3] fix: babysit round 1, address 3 gemini findings on cosign
 retry helper

G1 (CRITICAL): invoke sibling docker_push_with_retry.sh via bash (script ships 100644, bare invocation would fail Permission denied on Linux runner).
G2: printf -- '%s\n' for hyphen-safety and grep compat.
G3: guard grep -E against empty TRANSIENT_RE (would otherwise match every line and retry non-transient errors).
---
 .github/scripts/cosign_sign_with_retry.sh | 14 +++++++++++---
 1 file changed, 11 insertions(+), 3 deletions(-)

diff --git a/.github/scripts/cosign_sign_with_retry.sh b/.github/scripts/cosign_sign_with_retry.sh
index 58997ef280..22959dadb3 100644
--- a/.github/scripts/cosign_sign_with_retry.sh
+++ b/.github/scripts/cosign_sign_with_retry.sh
@@ -37,7 +37,12 @@ REF="${1:?usage: cosign_sign_with_retry.sh <ref>}"
 # both scripts in lockstep so a new transient signature added in one
 # place automatically protects every signing call too.
 SCRIPT_DIR="$(dirname "$(readlink -f "$0")")"
-TRANSIENT_RE="$("$SCRIPT_DIR/docker_push_with_retry.sh" --print-transient-re)"
+# Invoke the sibling helper via `bash` rather than relying on the
+# execute bit. Both scripts ship as git mode 100644 (no execute bit),
+# matching the existing publish-* action call sites which always use
+# `bash "$RETRY" ...`; a bare `"$SCRIPT_DIR/..."` would fail
+# "Permission denied" on the Linux runner.
+TRANSIENT_RE="$(bash "$SCRIPT_DIR/docker_push_with_retry.sh" --print-transient-re)"
 
 # 4 attempts, backoff 15s -> 30s -> 60s = ~1m45s of wait in the worst
 # case before the final attempt. Matches docker_push_with_retry.sh so
@@ -58,7 +63,7 @@ for ((i = 1; i <= ATTEMPTS; i++)); do
   # Idempotency branch: a re-sign that lost the createLogEntry race is
   # success, not a transient error. Check this BEFORE the regex so
   # attempt 1 -> 5xx -> attempt 2 -> conflict resolves cleanly.
-  if printf '%s' "$out" | grep -q 'createLogEntryConflict'; then
+  if printf -- '%s\n' "$out" | grep -q 'createLogEntryConflict'; then
     printf '%s\n' "$out"
     echo "::notice::Image ${REF} already signed -- skipping"
     exit 0
@@ -70,7 +75,10 @@ for ((i = 1; i <= ATTEMPTS; i++)); do
     exit "$rc"
   fi
 
-  if printf '%s' "$out" | grep -qiE "$TRANSIENT_RE"; then
+  # Guard against an empty `$TRANSIENT_RE` (e.g. sibling helper drift
+  # that silently prints nothing) - `grep -E ""` matches every line,
+  # which would retry auth failures and other non-transient errors.
+  if [[ -n "$TRANSIENT_RE" ]] && printf -- '%s\n' "$out" | grep -qiE "$TRANSIENT_RE"; then
     printf '%s\n' "$out" >&2
     echo "::warning::cosign sign ${REF} hit transient error (attempt ${i}/${ATTEMPTS}, rc=${rc}); sleeping ${BACKOFF}s before retry" >&2
     sleep "$BACKOFF"