submit marks existing draft PRs ready for review without --open, firing review requests

首先运行报告中的两遍 gh stack submit --auto 复现，并检查 submit 命令的更新路径，尤其是处理现有 pull request 的位置。将该路径与 --open 的行为进行比较，并使用 gh pr view A --json isDraft 进行验证；完成标准是：除非传入 --open，否则重复提交不会改变现有 draft，并且测试覆盖该回归问题。

Environment

• gh 2.96.0 (2026-07-02), gh-stack v0.0.4, git 2.46.1, macOS 26.6

Summary

gh stack submit --auto marked two already-open draft PRs as ready for review, which fired review requests to five CODEOWNERS. I did not pass --open.

The help text is explicit that this should not happen:

New PRs are created as drafts by default. Use --open to mark them as ready for review.

Repro

gh stack submit --auto        # creates PR A and PR B as drafts
gh pr view A --json isDraft   # => true, as documented
# ... make no change to the branches ...
gh stack submit --auto        # second run, still no --open
gh pr view A --json isDraft   # => false

Second run's output, which gives no indication it changed draft state:

Checking stack state...
Pushing to origin...
PR #A for branch-a is up to date
PR #B for branch-b is up to date
✓ Stack updated on GitHub with 2 PRs
✓ Pushed and synced 2 branches

The PR timeline confirms it:

<t>  ready_for_review   by=<me>
<t>  review_requested   by=<me>   x5

Those timestamps match the second submit to the second. I made no other change in between, and did not touch the web UI.

Why it matters

Marking a PR ready is not a local operation. It notifies every CODEOWNER, and on a repo with required reviewers it moves the PR into other people's queues. Doing it on an update pass, from a flag whose documented meaning is "use auto-generated titles without prompting", is surprising in a way that cannot be undone: gh pr ready --undo restores draft state but the notifications have already gone out.

It also makes submit unusable as an idempotent "sync my stack" command, which is what its output implies it is when everything is already up to date.

Expected

submit should only change draft state when --open is passed. If the current behaviour is deliberate for update passes, the help text should say so, and the output should name the change instead of printing is up to date.