feat(release-notes): agent-generated release notes action

[feng-shiplight]

Summary

A composite action that writes user-facing release notes for a commit range using a headless agent, returning them as a file for gh release create --notes-file. Consumer repos keep a short step plus their own notes-focus, mirroring how claude-review is wired.

Motivation: gh release create --generate-notes produces a list of merged PR titles, and a script like monots' gen-changelog.mjs can only reformat commit subjects. Neither can say what a release means for someone using the software. That is a writing task.

Design

The shape follows ci-triage/scripts/run-triage-agent.sh, the proven in-org pattern for running an agent non-interactively in CI:

• Model fallback chain, best first — each Claude model, then each Codex model, both overridable. A model the CI token cannot access fails fast and the next is tried, so a missing entitlement degrades instead of dead-ending.
• Success gated on a non-empty artifact, not an exit code, which also covers service-unavailable and unanticipated errors. A too-short result counts as no result, so a model refusal cannot become the release notes.
• Never fails the job. Notes are not a gate — a caller that gets nothing falls back to --generate-notes rather than blocking a release on a model outage. generated is exposed as an output so the caller can branch.

Two deliberate departures from run-triage-agent.sh, both because this task is write a paragraph rather than investigate and fix:

• Output is captured from stdout rather than written by the agent, so the agent needs no tools at all. Commit messages are author-controlled input, and this runs in a release job holding the caller's token; an agent with no write access cannot be steered by instructions embedded in a commit message. The prompt carries the whole range inline and instructs it not to use tools.
• No actions/checkout. This runs late in a release job, after the artifact is built and packaged — a fresh checkout would wipe those untracked outputs. The caller's checkout is reused and only deepened, since release jobs check out at depth 1 and git log from..to needs history.

Verification

Run end to end against a real range — screen-recorder v0.1.0..main, 10 commits, a 16.5 kB prompt — with claude --print:

• Grouped changes under themes a user would recognise (plan limits, recording reliability, sign-in and upload)
• Described behaviour rather than files
• Folded CI and refactor work into a single closing line instead of padding
• Invented nothing — every claim checks out against the commits, including one whose effect is limited to unpacked development installs, which it correctly qualified

Also: shellcheck clean on the runner, and both files parse.

Notes for reviewers

• v1 is not moved by this PR. The first consumer (screen-recorder) will pin to this PR's merge commit SHA, so the five repos on claude-review@v1 are untouched. Folding this into v1 later is a separate decision.
• The Codex leg installs only when openai_api_key is supplied, so repos without it simply run the Claude chain.
• Model defaults are pinned in the runner for the same reason ci-triage pins them: a runner otherwise inherits a model the CI token may not be entitled to.