Flaky test: persona hook 'no persona for an unsafe name: zz.test@setkernel.com' failed once in CI #54

The persona hook's unsafe-name test failed once in CI and passed on a re-run. A flaky test in check can hold up any agent-plugins PR, so this needs its cause found.

What happened

• On run 40641544025, attempt 1 (PR Org rules: on an issue with an allow-test-changes line, comment progress instead of ticking boxes #53, head 91c617f7, 06:03Z on 28 Oct), tests/run.sh reported 775/776 passed. The failing test was:

  not ok 727 - no persona for an unsafe name: zz.test@setkernel.com (exit 1, want 0; output must match: ^silent$)
  

• Attempt 2 of the same run, on the same head, passed. PR Org rules: on an issue with an allow-test-changes line, comment progress instead of ticking boxes #53 doesn't touch the persona hook or its tests; it changes managed/settings.json, rules/org-rules.md and one marker list in tests/run.sh.
• The test is at tests/run.sh:853-856. silent prints silent only when persona_hook exits 0 and prints nothing, stderr included. So on attempt 1 the hook either exited non-zero or printed something. The test output doesn't say which.

Do

• Find why the hook can fail or print for zz.test@setkernel.com some of the time. For example: a race on the cache under CLAUDE_PLUGIN_DATA left by the previous case, $tmp/gh-calls timing, or the sh_cmd variant in use.
• Remove the race. Don't add a sleep or a retry (org rule: "Never add sleeps or retries to hide a race").
• Make silent failures show what the hook printed and its exit code, so the next failure explains itself.
• Run the suite in a loop (100 runs, or as many as fit in a few minutes) and paste the pass count.

Risk: R1 (tests). Found by Claude Lead while checking PR #53's CI.

— Claude Lead