fix(ci): harden release and rollout checks

[tlgimenes]

Summary

• classify release changes from a NUL-delimited git diff file instead of exporting the full PR file list through one environment variable
• keep queued release jobs safe by diffing the triggering merge commit while applying bumps on current main
• bound silent recovered runs inside the multi-pod rollout-churn polling window

Testing

• bun run fmt:check
• bun run lint
• bun run check
• bun run knip
• bun test scripts/release-changes.test.ts
• bun x tsc -p scripts/tsconfig.json
• actionlint .github/workflows/release-tagging.yaml
• replayed the real refactor(studio): split API and web workspaces #5120 first-parent diff: 5,349 NUL-delimited paths / 267,761 bytes, yielding the expected 12 manifests and deploy scope both
• full unit suite: all files before the unchanged monitoring SQL test passed; Bun 1.3.14 then reproducibly segfaulted on macOS at process startup for that test, while the same test passes 45/45 in isolation. Subsequent files were continued separately; two load-sensitive timer/process tests passed in isolation. Authoritative Linux suite runs in CI.
• multi-pod suite: not run locally because Docker is unavailable; authoritative suite runs in CI

Impact

No runtime production behavior changes. The release workflow no longer crosses the Linux per-environment-string limit on large PRs, and the multi-pod test environment now resolves truly silent recovered runs before its terminal poll expires.

---

Summary by cubic

Hardened release tagging to classify changes from a NUL-delimited git diff of the merge commit and compute deploy scope reliably. No production runtime changes.

• Bug Fixes
  • Use git diff merge^1..merge (NUL-delimited) and scripts/release-changes.ts to avoid env size limits and flaky file lists.
  • Verify the event SHA is the merge commit on main and an ancestor of origin/main, and apply bumps on current main for queued runs.
  • Derive deploy scope from paths (server/web/both) and only bump manifests with a string version.
  • In multi-pod tests, set RUN_IDLE_TIMEOUT_MS=120000 to bound silent recovered runs within the rollout polling window.

^{Written for commit ff81bbf. Summary will update on new commits.}