v2.9.0 — Batch Discipline

Scaling up SmartWorker batch and unattended work safely: a tier must earn
each work type before it runs at scale, parallel dispatch gets hard rules,
and no subtask ever runs twice.

Added

• SmartWorker calibration gate — tier / capability labels are now
  framed as execution slots, not permanent truths. Before a tier is
  trusted with a new class of work (especially batch/unattended runs), a
  small calibration pilot must pass the owner's quality threshold, recorded
  in the spec's new calibration field and re-run when the underlying model
  changes. The capability twin of the mechanical-gate rule: a tier is not
  suitable until a pilot shows it passing.
• SmartWorker parallel fan-out — the Main Agent may dispatch multiple
  workers concurrently for many same-shaped independent subtasks (new
  optional concurrency spec field) — star topology under the single Main,
  never a peer mesh. Five rules: calibrate first; owner-set concurrency cap
  • provider rate limits; disjoint write targets per worker; a per-batch
    (not per-worker) budget ceiling; and single assignment + idempotent
    outputs (dispatch manifest, retry only confirmed-failed items, reconcile
    against real artifacts on resume, prefer overwrite-by-key over appends —
    no subtask runs twice). Applies equally when a Loop Contract's run fans
    out internally.
• Loop Contract duplicate-work guard — the cross-run state anchors are
  now explicitly the duplicate-work guard: a worker checks actual artifact
  state before doing an item and never re-does what the anchors show as
  done; the definition-of-done check must be idempotent so a retried or
  re-entered run converges instead of repeating work.

Changed

• Loop Contract template: Contract Restate moved into the run report —
  the kickoff prompt's "restate before doing anything" instruction was
  skipped in live trial runs (chat-level asks are decorative); the restate
  is now a required opening section of the run report itself, making it
  part of the machine-checkable done condition.

Notes

Every feature in this release was motivated and validated by live
downstream usage before shipping — real batch knowledge-work runs and the
governed autonomous-run trials behind v2.8.0.

Upgrading

Use the AI-driven upgrade prompt in the README ("Upgrading an existing
workspace") — preserve-first: your filled context files and records are
never overwritten.

Full changelog: CHANGELOG.md · v2.8.0...v2.9.0