feat: save default reviewers globally with jury agents jury

[zzxwill]

Closes #98

What changes

Without --jury, jury review used every enabled reviewer, so a single missing CLI (e.g. grok) aborted every run unless --jury was passed each time or each repository's jury.config.json was edited.

This adds a saved global default reviewer list, mirroring jury agents judge:

• jury agents jury claude,droid,amp saves "reviewers" in ~/.jury/config.json, beside the global judge; opt-in agents are allowed. jury agents jury --reset removes it.
• jury agents jury with no arguments opens a checkbox picker on a TTY (install status, opt-in and judge marked, empty selection refused, esc cancels), built on Node's readline — no new dependency (lib/picker.js). Off a TTY it prints the current setting.
• Precedence: --reviewer/--jury → repository reviewer roles → saved default reviewers → built-in pool.
• A saved list disables automatic role assignment, as --jury does. The selected judge is excluded; a list containing only the judge is an error rather than an empty jury.
• A saved reviewer that is uninstalled, disabled by the repository, or no longer an agent stops the run with an error naming the saved setting instead of silently shrinking the jury.
• jury agents marks saved defaults, notes when a repository overrides them, and bases its exit status on the saved pool.
• README, docs/configuration.md, lib/help.js and CHANGELOG updated.

Also fixes a race found by this PR's first CI run (Node 24 / macos-latest, web-launch.test.js: "opened URL must select a run available in the API response"): writeRun truncated run.json and rewrote it in place, so a console read landing mid-write saw empty JSON and dropped the live run from /api/run. It now writes a temp file and renames it into place.

Validation

• New test/default-reviewers.test.js: save/reset round-trip and validation of the global file; precedence across repo roles, saved list and built-in pool; judge exclusion; disabled/unknown saved names; automatic roles turned off; CLI save/show/validate/reset leaving the file unchanged on bad input; jury agents markers; a --dry-run review using the saved list, with --jury and --judge overrides; picker key handling (move, toggle, empty-save refusal, save, esc/q/ctrl-c cancel).
• Drove the real picker through a pseudo-terminal: enter saves the selection, toggle-then-esc cancels without writing the file.
• writeRun race: a stress reproduction (rewrite run.json in a loop while listing runs) missed the live run on ~40% of reads before the fix and 0 of ~6,000 after. New runs.test.js case fails without the fix and passes with it.
• npm test: all pass except loop.test.js "--judge selects one judge…", which fails identically on unmodified master in the dev container because a real claude CLI on PATH triggers automatic role assignment; it passes in CI.
• npm run test:package passes.

Limitations

• Repository jury.config.json entries with "role": "reviewer" override the saved list entirely rather than merging with it.
• jury reply without --reviewer also resolves through the saved defaults, so a saved list that later breaks (e.g. names a now-disabled agent) errors there too.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YWwuLax3VBtTJuujgK748c