Extract the fat inline bash steps from upstream-release-docs.yml into scripts/

[danbarr]

.github/workflows/upstream-release-docs.yml is 1667 lines across 41 steps, including 960 lines of inline bash. The size is concentrated in four steps:

Step	Inline bash lines
Augment PR body (marker-delimited section)	313
Assign reviewers and prepare contributor mentions	121
Resolve release owner and contributors	99
Capture skill signal files	74
remaining 37 steps	353 (avg 9)

Two reasons to extract the big ones:

• Nothing lints the inline bash. There is no shellcheck or actionlint in CI, so checking a change today means hand-extracting steps[].run from the YAML, stubbing ${{ }}, and piping it to shellcheck.
• scripts/upstream-release/ already holds five .mjs files and this workflow already invokes one, so a script there inherits Prettier and ESLint from existing CI with no new tooling.

Augment PR body is the best first candidate: its only I/O is roughly 25 env vars in and one gh pr edit --body-file out, and the PR body is the artifact reviewers actually read.

Scope

In:

• Augment PR body to scripts/upstream-release/render-pr-body.mjs
• Assign reviewers and Resolve release owner and contributors as separate follow-ups if the first goes cleanly

Out:

• The two Claude step prompts (88 lines). Their content is almost entirely interpolated run context, and the durable instructions already live in the upstream-release-docs skill.
• Capture skill signal files, mostly file reads and GITHUB_OUTPUT heredocs.
• The 37 short steps, which belong inline.

Sequencing

After #1111 has run live on two or three release PRs. That change is unvalidated, and stacking a refactor on it means a live failure has two candidate causes instead of one.

#1111 and this refactor touch the same steps, so if #1111 needs iteration, extract Augment PR body first (#1111 only added two table rows and two sections to it) and leave the reviewer steps until after validation.

Verification

Behavior-preserving: render the extracted script against the env of a real past run and diff the output against that run's actual PR body.