push-guard 0.5.0

• New: with Baseline installed, a push is blocked only by test failures it caused. push-guard runs
  baseline check --json --base <B> -- npm test, <B> being where HEAD left its upstream (or the
  remote's default branch). NEW failures, NEW load errors included, block the push, each listed with
  its evidence. Failures that were already there (pre-existing, pre-existing with a changed error,
  flaky, environment) let it through, with a short toast, 0 new, 3 still failing (not yours) above
  the prompt, and a note to Claude on the push's result.
• Whenever Baseline can't answer clearly (exit 2 or 3, UNKNOWN, output that isn't its JSON report, a
  crash or timeout, a runner it can't read per test, a verdict without the runs to show for it, nothing
  new to push, a first push), push-guard runs npm test as before and blocks on any failure, adding a
  line saying why Baseline wasn't used.
• New: the baseline option, auto (default) or off. Baseline is found once per session through
  BASELINE_BIN, $PATH, or an installed Baseline plugin; without it nothing changes.
• The hold for pushes that change existing tests still applies, also when the only failures are old.
• New: projects with no npm test script are no longer pushed unchecked. Their typecheck,
  type-check, tsc, lint, check and format:check scripts run instead (npm run <script>, in
  that order). Without Baseline, any failing check blocks the push, with the last lines of its output.
  With Baseline, each runs as baseline check --json --base <B> -- npm run <script> and only errors new
  with this push block it. A check Baseline can't read error by error runs as is, and any failure blocks.
  A missing script never blocks. With none of these scripts, the push goes ahead loudly: a toast
  (push-guard: nothing was checked: ...), a note to Claude and a line above the prompt. The new option
  checks is auto (default), all (also build) or off (as before: nothing run). The line above the
  prompt names the checks (push-guard: typecheck, lint passing). Projects with a test script are
  unchanged.
• Limits: Baseline judges per test only runners it reads (vitest, jest including react-scripts test
  and craco test, node --test, pytest and others). A chained script is judged part by part; a part
  with ||, a pipe, $(...) or a redirect is not judged, and then push-guard runs the script and blocks
  on any failure. Other runners (mocha, tap, ava, make test) get a whole-command answer, used only when
  the command passed. Flaky and environment failures go through. See "With Baseline" in the README.
• Fixed: a push could hang for 20 minutes. npm test in Create React App (and jest/vitest watch scripts)
  waits for input unless CI=true, and push-guard waited 10 minutes for Baseline and then 10 more for
  npm test. Every command push-guard runs (tests, checks, Baseline) now gets CI=true and an empty,
  closed stdin. New option timeoutMinutes (default 5, 1 to 15; above 15 counts as 15, between 0 and 1
  as 1, 0 or less as 5): one budget per script for Baseline and the plain run together. Baseline gets
  --timeout just under it. A script that doesn't finish blocks
  the push with push-guard: <command> did not finish within N minutes (watch mode? a hung test?), so the push was blocked. Run it yourself to see why, or raise the timeoutMinutes option.; if Baseline used up
  the time, the script is not run a second time. On a project whose npm test starts watch mode, with
  CI unset, push-guard before the fix hung (Baseline's 10-minute limit ran out, then a second 10-minute
  npm test; the session was cut off after about 13 minutes); after the fix it blocked in 86s with
  "26 failures are new with this push" (Baseline took 70.1s of that).
• Fixed: a push that broke only code npm test doesn't cover went through silently. With the new option
  suites at auto (default), push-guard runs test:all when package.json has one, instead of test;
  otherwise test plus the test:* scripts that look like suites (not watch, e2e, UI, coverage, snapshot,
  device or benchmark ones, nor ones test already runs); with npm workspaces also
  npm test --workspaces --if-present. With Baseline, each runs as its own baseline check. A folder
  one or two levels down (api/, packages/*, apps/*) with a test script nothing ran is named in a
  toast, a note to Claude, the deny and the line above the prompt:
  push-guard ran npm test; these test scripts were NOT run: api (api/package.json: test). suites: root
  runs only npm test and names the suites it left out.
• Baseline reports are trusted less: a verdict that lets a push through must be about one test, load
  error or (for a parsed check) check error. Whole-command, (command) and (test session) verdicts,
  units of the wrong kind for the runner, and reports whose summary counts more failures on HEAD
  than they give verdicts for make push-guard run the script itself (blocking on any failure). A
  whole-command report that says the command passed is taken as a pass, so it isn't run twice. The rules
  are in "How it decides" in the README.
• Fixed: a chained test:all (react-scripts test && jest --config jest.api.config.js && npm run check:x)
  could be pushed with a break in a later part when Baseline reported only the first part clean. push-guard
  now reads Baseline's chain report (runner.name chain, runner.parts, part on each result) part by
  part with the same rules as any report: a NEW failure in any trusted part blocks (listed with its part),
  and the chain is let through only when every part was judged and every part's report holds up;
  otherwise the script runs as is and blocks on any failure. On a push that broke only the api code of a
  Create React App project (local bare remote, real headless Claude session), push-guard before the fix
  pushed silently; after the fix it blocked with "part 2 ... passed 5/5 runs on , fails 5/5 runs on
  HEAD", the remote unchanged, and Baseline answered in 23.9s.
• Messages name every script that ran and how it came out (push-guard ran: npm test passed, npm run test:api failed.); the nothing-checked toast names the test:* scripts it found and did not run.
• Tests now run with claude plugin test (165 pass, up from 157, with 8 new for chained scripts in
  hooks/baseline.test.ts); three "no test script" tests had a fixture that still gave them a test script.