[daniele21]

Problem and outcome

MLXEngine.close() was a no-op: even when Korgis logically unloaded a runtime, its backend could keep model/tokenizer references alive. Correct that root cause in the canonical MLX backend, ensuring idempotent close, closed-runtime inference rejection, and model-isolated teardown under lease-safe manager semantics.

Invariants

• No forceful removal during active leases.
• Do not invoke global Metal allocator/cache cleanups that could interfere with other resident MLX runtimes.
• Reference release is not proof of reclaimed RAM/Metal unified memory; physical measurements remain necessary.
• No backend-specific process or new resource manager created.

Classification

PRODUCT_LOCAL / INTEGRATION / STRONG (native runtime lifecycle and resources) / REMOTE_AUTOMATED deterministic, REAL_ENVIRONMENT 32-GiB Apple Silicon reclamation pending #239 and #244.

Evidence

Deterministic tests verify weakref release, repeated close, closed stream, busy-lease protection and unrelated MLX model retention. Required CI and base revalidation before merge.

Exact-head validation

• Candidate 82466a4a815f8e1f8967cf1a55dccd765c9c5e83 / converged dev base eb8cfdbe5eb32a72102d26189f60555cf3f80728.
• CI INTEGRATION/STRONG PASS: 3 Python versions, backend weakref/lease tests, lint, repository selectors.
• Repository Health PASS.
• No browser E2E required by selector for backend-only behavior. Physical Metal memory reclamation, active multi-model workload safety and Mac 32-GiB proof remain REAL_ENVIRONMENT pending (Validate macOS backend crash cleanup and host-memory guard on 32 GB Apple Silicon #239, Mac memory safety: native pressure monitoring and evidence-gated resource relief #244).