fix(notify): keep bump PRs open on update, retry failed repos with backoff

[ssestak]

Summary

Fixes two problems in the consumer-repo notification script observed during the 2.6.0 rollout.

1. Updating an existing bump PR auto-closed it. The update path force-reset the PR branch to the default branch HEAD before committing the bumped files. For a moment the PR had zero commits and GitHub auto-closed it, while the script still reported UPDATED. In the last two runs 12 of 13 "updated" PRs ended up closed with a stale title, and the next run then opened duplicate PRs.

• Existing PRs are now brought up to date by merging the default branch into the PR branch (POST /repos/{repo}/merges), so the PR never loses its commits.
• On a merge conflict the branch is deleted and recreated and a fresh PR is opened.
• The script verifies the PR is still OPEN before reporting UPDATED.

2. No resilience to GitHub outages. During a GitHub incident (degraded Git operations) 22 repos failed in a row because Contents API writes were rejected, and the run had no way to recover.

• Failed repos are retried in passes with exponential backoff: up to 10 passes, starting at 2 s and doubling up to 5 min.
• Only failed repos are retried; files already bumped on the branch are not rewritten, so partial progress is kept.
• Tuning is read from repository variables NOTIFY_RETRY_MAX_ATTEMPTS, NOTIFY_RETRY_BASE_DELAY, NOTIFY_RETRY_MAX_DELAY (already set to the defaults), so it can be changed without a merge.

Also

• gh errors are no longer discarded; the last stderr line is printed next to each FAIL so logs show the actual API response.
• A failed gh pr create was reported as CREATED because stderr was captured into the URL variable.
• New NOTIFY_REPOS env var and repos input on workflow_dispatch to process an explicit list of repos, e.g. to re-run only the ones that failed.
• Script restructured into functions so the per-repo flow can be re-run and unit tested.

Test plan

• New BATS suite (13 tests) with a gh CLI mock: backoff schedule, create path, leftover-branch reuse, merge-based update without force-reset, conflict → recreate, closed PR → FAIL, no rewrite of already-bumped files, retry until success, give up after max attempts, only failed repos retried, self repo skipped, dry run. Full bats -r . passes (140 tests).
• actionlint on the workflow.
• --dry-run against two real consumer repos (one create, one update) produced the expected plan.
• After merge: re-run the workflow via workflow_dispatch with the repos input for the repos that failed in the 2.6.0 run.

🤖 Generated with Claude Code

https://claude.ai/code/session_015HbhiukLQv71hkruK31fyT