fix(tts): stop x64 startup crash in ONNX NCHWc graph transformer

[julia-kafarska]

Fixes #36, fixes #37, fixes #38 — all three are the same bug.

Root cause

onnxruntime-node 1.23.2 segfaults inside NchwcTransformer while applying Level-3 (layout) graph optimisations to model_q8f16.onnx. From the Intel Mac crash report in #38:

Thread 27 Crashed:: WorkerThread
0  onnxruntime::graph_utils::IsSupportedOptypeVersionAndDomain(...)
1  onnxruntime::NchwcTransformerImpl::Transform(Node&)
2  onnxruntime::NchwcTransformer::ApplyImpl(Graph&, ...)
3  onnxruntime::GraphTransformer::Apply(...)
4  onnxruntime::GraphTransformerManager::ApplyTransformers(...)
5  onnxruntime::InferenceSession::TransformGraph(...)
6  onnxruntime::InferenceSession::Initialize()
9  InferenceSessionWrap::LoadModel(...)

It's a hard native abort (EXC_BAD_ACCESS / 0xC0000005) inside InferenceSession.create, so:

• the JS try/catch → CPU fallback in createSessionWithFallback never gets to run,
• crashpad is never connected, so nothing is logged and %APPDATA%\Out Loud is never created,
• the app simply vanishes 2–3s after launch.

NCHWc is an x86-only transformer. That's why this reproduces on every x64 platform and never on Apple Silicon:

Issue	Platform	Symptom
#38	Intel Mac, macOS 15.7.7	crash report above
#37	Pop OS 24.04, i5-4690K, NVIDIA 970	log ends at Loaded embedded model: …model_q8f16.onnx → Segmentation fault (core dumped)
#36	Windows 11 x64, AMD and NVIDIA, plus Linux Mint	reproduced standalone in Node with no Electron, ["cpu"] EP

GPU vendor, DirectComposition, VC++ redistributable and Wayland/EGL are all red herrings — the crash also hits the CPU-only preload path, so --disable-gpu and the acceleration setting make no difference.

Fix

Cap graphOptimizationLevel at "basic" on x64, skipping the Level-3 pass entirely. arm64 keeps ORT's default.

Gating on process.arch === "x64" rather than process.platform === "win32": the crash is architecture-scoped, not OS-scoped, so a win32 gate would leave Intel Mac (#38) and Linux x64 (#37) broken.

Graph optimisation only changes how the math is evaluated, not what it computes — worst case is last-decimal differences in output samples, nothing audible.

Verification

• Session creation smoke-tested on arm64 at both "basic" and "all": both succeed, init 591ms vs 613ms, so the arm64 path is unaffected and the option costs nothing measurable.
• x64 verification comes from @Blue4Life90 in Windows 11 (AMD GPU): app initializes then exits ~2s into startup — no window, crashpad not connected #36, who root-caused this and tested the matrix on Windows 11 / AMD: default → crash, "disabled" → OK, "basic" → OK, mem-arena-off / single-thread → still crash. With "basic" the app reaches ONNX session providers: cpu → Model preloaded successfully and stays running.

Credit for the root cause and the fix goes to @Blue4Life90, who diagnosed it down to the transformer and verified it on affected hardware.

Possible follow-up

"extended" (Level 2) should also avoid NCHWc — the layout transforms are Level 3, and this ORT version lists 'layout' as a distinct level above 'extended' — which would retain more optimisation than "basic". Nobody has verified that on x86 hardware yet, so this PR ships the tested value.

Separately, this class of failure is invisible by design right now: three users lost days to a crash that writes no log. An early main-process log file (before userData exists) plus crashReporter on Windows would make the next one self-reporting.

🤖 Generated with Claude Code