[jeremybeard]

Cuts the e2e run from ~43 to ~13 minutes.

• Tests run in three chains per deployment kind (standard, hibernate), each chain on its own deployment, plus one deployment each for the preview and remote tests: 9 at once. Tests that compare a deployment's versions before and after still never share one concurrently.
• The deployments are created one at a time (creates return immediately), retrying rate-limited requests, and published as a slot→ID map; one job deletes them all, along with any deployment preview a failed or cancelled test left behind.
• A new input validation job runs the action against a stub astro that records its calls, so a rejection only passes if it came before the action touched Astro. A control case must be accepted.
• The diff-base patch now applies to every run shape, which needed the deploy-type step in action.yaml to read the before/after SHAs through its existing shell variables. No behavior change.
• Fixes a timestamp check that always passed and an output name typo that made one assertion compare against an empty value.
• Runs the e2e deployments on runtime 3.3-2, with 15-minute health waits.

[wendtek]

Nice!

[neel-astro]

LGTM, thanks for all the cleanup 🚀

[neel-astro]

I wonder if we can write some auto update CI job to keep these runtime versions to latest