fix(engine): mtlBuf's documented copy fallback was an infinite recursion

[penta2himajin]

Closes #177.

if let b = a.asMTLBuffer(device: device, noCopy: true) { return b }
return mtlBuf(a, device)          // same arguments -> infinite recursion

The docstring promises a copy fallback ("不可なら copy"). There was none: the fallback called itself with identical arguments, so a nil from the no-copy wrap would recurse until the stack overflowed. It now passes noCopy: false — what the comment already described.

Latent, not live

mlx-swift currently copies internally rather than returning nil, so today the branch is unreachable. It is latent across 267 call sites, and whether it stays unreachable is a property of mlx-swift's implementation, not of qwisp. A version that returns nil instead of copying would turn every weight-buffer wrap into a crash rather than a slow path.

No test, deliberately

The branch cannot be reached without forcing asMTLBuffer to return nil, and there is no hook for that. A test that only exercises the happy path would raise the locked-test count without guarding anything — the opposite of what that counter is for.

Found while

Attributing the 8GB tier's memory. nonMLXmetal is metal - active - cache with metal = MTLDevice.currentAllocatedSize, and mtlBuf wrappers alias existing pages: they inflate that counter without costing resident memory. This repo had already documented exactly that in #150 (TellRuntime.swift:640-644) — reading mtlBuf to confirm it was a view and not a copy is what surfaced the recursion.

Gates

RAWTESTS 100/100 · BENCHBATCHTEST PASS · COMPTEST 97/97 · CBGUARD PASS

🤖 Generated with Claude Code