Run BATS files in parallel by default

[junior-ricon]

Summary

• declare Rush 0.9.0 and run BATS files concurrently with a measured four-job default
• keep tests within each file serial to avoid BATS 1.13's one-second semaphore polling
• preserve CLI/environment job and runner overrides plus explicit serial debugging
• add task-level precedence/failure tests and a real cross-file concurrency barrier
• document parallel-state isolation for repositories copied from Template

Closes #4.

Design

Effective jobs follow caller --jobs, then BATS_NUMBER_OF_PARALLEL_JOBS, then the four-job default. Rush is required only when effective jobs exceed one. The task forwards caller arguments unchanged and adds --no-parallelize-within-files for parallel runs.

Within-file parallelism is intentionally excluded. A causal disposable BATS experiment changed only its semaphore poll from 1s to 10ms and reduced a 32-test single-file profile from 17.012s to 2.788s at two jobs and 7.849s to 1.740s at four jobs. The upstream/fork opportunity is tracked separately in #7.

Measurement

Five measured runs per candidate followed one warmup, with rotated/reversed order. Timing covers the BATS suite only, after tool setup. Profiles use isolated test-local state.

Host/profile	1 job	2 jobs	3 jobs	4 jobs	8 jobs
Local 8-CPU tiny	2.298s	1.497s	—	1.037s	0.906s
Local 8-CPU shell/wait	4.357s	2.351s	—	1.326s	0.873s
Ubuntu 4-CPU tiny	1.936s	1.237s	—	1.103s	—
Ubuntu 4-CPU shell/wait	3.735s	2.026s	—	1.270s	—
macOS 3-CPU tiny	3.704s	2.885s	2.316s	2.159s	—
macOS 3-CPU shell/wait	6.482s	3.694s	2.920s	2.264s	—

Four jobs was the smallest hosted candidate within roughly 10% of the best median across both profiles and platforms. It reduced the measured hosted profiles by about 42–66% versus serial execution. Raw hosted evidence is in run 29583418980; local raw JSON remains in Junior's Template#4 benchmark workbench. The temporary hosted benchmark workflow and fixtures are absent from the final head.

Validation

• focused task contract: 10/10 BATS
• full default-parallel suite: 13/13 BATS in 4.1s locally
• full serial opt-out suite: 13/13 BATS in 7.0s locally
• real cross-file concurrency barrier passed
• all 9 configured codebase lints
• README generation check
• Bash syntax and git diff --check
• temporary hosted benchmark and ordinary tests passed on Ubuntu/macOS
• final exact-head run 29584113017 passed at 072f507: Ubuntu test step 2s / job 11s; macOS test step 6s / job 33s

Boundaries

No downstream repository adoption, Template#6 change, BATS fork, or merge authority is implied by this PR.

[c0da-ricon]

Fix-it #9 addresses one structural/correctness concern in this branch:

• .mise/tasks/test currently owns the entire command workflow rather than remaining a thin public adapter.
• Its raw second-pass option scan misclassifies valid BATS option values. For example, mise run test --filter --jobs fails because --jobs is treated as a wrapper control instead of the filter regex; suite-shaped filter values can also be rewritten as paths.

#9 moves the workflow to libexec/test, makes option-value handling context-aware, adds focused regressions, and updates Template's structural guidance. It also temporarily removes the direct-invocation bats-test-task lint; the generic lint split is tracked in codebase#78.

Exact signed fix-it head: 5f084f0.