engine-room v1.0.0

Recipe-driven provisioner for local AI engines — the action half of a knowledge/action split. Browse verified engine recipes, resolve one against your live GPU rig, then provision → launch → measure → roll back — reproducibly and validated.

First stable release, taken to a clean bill of health via a full dogfood swarm (audit → independent source-verify + severity-calibration → test-first amend → 3-lens adversarial verify → fix-up).

Highlights

• er CLI (stdlib-only Python, zero third-party deps): rig / list / show / preflight / provision / teardown / status.
• Safe provisioning: dry-run by default; --execute is the only rig-touching flag. Ledger-before-side-effects, newest-first identity-verified compensators (never kills an unrelated process, never claims rollback on a no-op), and a model+axis-keyed, bound-direction-aware measure gate that enforces a VRAM ceiling.
• Fail-fast preflight: capability/constraint dispatch with a legible note status (an unmodeled constraint is never silently passed), a resolvability floor, and a stable instance_id.
• Control Panel prototype (app/) — polymorphic per recipe kind, offline/read-only browse mode, loud contrastive ANDON.
• 97 stdlib unittest tests, SECURITY.md + threat model, CI, and a Starlight handbook.

Reproducibility is partial by design — content-addressed vendor pins are Goal 1; unpinned/placeholder shas are accepted today (disclosed in SECURITY.md).

See the CHANGELOG and the handbook.