docs(pr-test-sync): diagnostic helpers, rate-limit budget, early readonly merge, same-step failures

[saqibmanan]

Updates .claude/skills/pr-test-sync/SKILL.md with lessons from fixing provider smoke (#128). One commit per section:

• Reuse the diagnostic helpers; never hardcode admin-managed taxonomy. Adds a table of wait_for_option, click_until, sector_name, backend_enum_labels and save_failure_artifacts, and what each one prevents. Hardcoding the sector "Budgets" put CI provider-smoke red on every PR after the 2026-09-24 dev refresh.
• Budget browser runs against 1000 POST/hour. A 429 doesn't skip a browser test; the test hangs after login and dies as a blank timeout. About three full provider runs from one machine spent the hour, twice. Explains the new "backend rate limit" report section.
• Readonly tests merged before the feature reached prod. What to tell the user, and why any DataSpaceBackend main merge (including dependabot) becomes a rollback risk until the feature ships. This happened with test: embedded use case dashboards (DataSpaceFrontend #476) #122 / DataSpaceFrontend#476.
• Gotchas:
  • A failure at the same step on every rerun points to data or a locator, not contention. This is the counterweight to the existing "change -n first" rule.
  • Check CI push runs before investigating red checks on tests you didn't touch.
  • Read XPASS with -rxX (Chart tests: test_prv_008 XPASSes while test_prv_004 XFAILs under the same xfail #132).
  • The zsh $FILES word-splitting trap.

sector_name, wait_for_option and the 429 section come from #128. The skill says to check that #128 is in the base branch.