Standardize five-minute parallel CI task plans

[Nickfost]

Outcome

Make the fleet consume independently runnable project tasks and deterministic shards so available workers reduce wall-clock time. Keep fast and full only as aggregate local interfaces.

flowchart LR
  P[project plan.json] --> M[GitHub matrix]
  M --> W1[worker: lint]
  M --> W2[worker: unit 1/N]
  M --> W3[worker: unit 2/N]
  M --> W4[worker: integration 1/N]

Loading

Implementation: draft PR #15
Validated run: Validate ci-fleet prototype #14

Required contract

• Every project declares named tasks and deterministic shard counts.
• Every matrix job has a hard five-minute timeout.
• Expected test payload targets four minutes or less, reserving startup/reporting time.
• scripts/ci/run.sh <task> --shard I/N is the fleet execution interface.
• scripts/ci/run.sh fast|full remains an aggregate local/developer interface.
• Plans reject duplicate task IDs, invalid groups, missing fast/full coverage, and estimates above policy.
• Shards use isolated Docker/Compose resources.
• Documentation explains that total test-minutes divided by workers is an ideal lower bound, not a guarantee.
• Existing projects migrate incrementally and retain their prior CI until equivalent parallel validation passes.

Scope

Update the normative project standard, compliance/migration instructions, copyable project example, organization configuration schema, initializer, validator, and regression tests. Do not register a runner, migrate a project, or deploy infrastructure in this change.