fix(ci): make Sync-from-Docs auto-merge actually work

[acedatacloud-dev]

Why

The Sync from Docs workflow has been silently broken since it was created — every docs-updated dispatch leaks one or more PRs (stale drafts or 0-commit zombies), causing the open-PR backlog you can see in the queue right now (APIs ~40, Clis ~30, SDK ~7).

The Skills repo had the exact same workflow with the exact same bugs and was fixed in AceDataCloud/Skills#181. This PR ports that fix verbatim. The fix has been running cleanly in Skills for several days.

Root cause — five compounding bugs

#	Bug	Effect
1	Agent-done detection used gh api .../actions/runs?event=dynamic filtered by workflow-run name. That query regularly returns 0 hits or a null conclusion, so the polling default "pending" sticks for the entire window.	30-min timeout reached without ever attempting a merge → PR left in DRAFT.
2	PR detector grabbed the first copilot/* branch it saw, not the one bound to the current issue.	Stale draft from a prior run was tracked instead of the new one; new agent's PR was ignored.
3	create-task closed prior issues but not their orphaned PRs (including 0-commit zombies from agent boot failures).	Drafts and zombies pile up indefinitely.
4	timeout-minutes: 35 plus 90×20 s polling = 30-min budget, shorter than typical Copilot sync runs.	Job exited before agent finished.
5	gh pr merge had no --delete-branch, no retry, and shared the polling job.	Even when the rare success path triggered, branches stayed; one API flake killed the whole job.

What this PR does

1. Pre-cleanup — create-task now closes every leftover open copilot/* PR at the start of each dispatch (with branch deletion), not just the prior auto-sync issues. A new Docs commit makes prior attempts stale by definition.
2. PR-to-issue binding — the wait loop locates the PR via #<issue> in PR body (Copilot's standard linking format), with a single-PR fallback.
3. Reliable completion signal — replaced the workflow_runs poll with the observable signal isDraft == false. Copilot itself marks PRs ready-for-review when its agent finishes successfully.
4. Zombie guard — any copilot/* PR open for 30 min with 0 commits is auto-closed (with branch deleted).
5. Bigger budget — timeout-minutes 35 → 60, polling 90×20 s → 110×30 s = 55 min.
6. Hardened merge step — new isolated step does gh pr ready (idempotent) → settle checks → gh pr merge --squash --admin --delete-branch with up to 3 retries. On final failure, comments on the PR.
7. Wider permissions — contents: write, pull-requests: write so the job can actually merge and delete branches under GITHUB_TOKEN (also helpful as fallback if a secret rotates).

Outcome contract

After this lands, every docs-updated dispatch resolves to one of three outcomes:

• ✅ Merged to main — happy path
• ✅ Issue closed by Copilot — no changes needed
• ❌ Surfaced failure — PR commented explaining what broke (zombie / merge-failed / checks-failed)

No more silent draft accumulation.

Backlog cleanup

The existing open copilot/* PRs in this repo will be closed in a follow-up bulk-cleanup pass after this PR lands (the new create-task step would close them on the next dispatch anyway, but doing it explicitly produces a cleaner audit trail).

Test plan

• YAML validated locally (yaml.safe_load).
• Same logic is already live in Skills repo with successful runs since 2026-05-04.
• Will be exercised by the next docs-updated dispatch from PlatformBackend.