[rubas-agent]

Before

The NIF builds whisper.cpp v1.9.4 from the vendor/whisper-rs-0.16.0-patched branch of this repo. ggml-org/whisper.cpp released v1.9.5 on 2026-10-06.

Problem

v1.9.4 ignores the abort callback during the language auto-detection, and it reads out of bounds on a model file with an invalid tensor type or a VAD model with an encoder layer count other than 4. transcribe/3 and load_model/2 reach all three.

After

The vendor branch carries whisper.cpp v1.9.5 (ggml 0.26.0), and both [patch.crates-io] entries point at that commit.

Vendor commit, a fast-forward that keeps the old rev reachable: 9a4232f...1f3c84e

	Before	After
whisper.cpp	v1.9.4 (927cfce3)	v1.9.5 (d1be6fde)
ggml	0.23.0	0.26.0
Vendor rev	9a4232f	1f3c84e
whisper-rs / whisper-rs-sys	patched 0.16.0 / 0.15.0	the same, kept (issue #26)

What we read, what we took, and the checks

The vendor commit moves the sys/whisper.cpp submodule to v1.9.5 and regenerates sys/src/bindings.rs, the fallback for a failed bindgen run, with bindgen 0.72.1 and libclang 23. whisper.h did not change. ggml adds GGML_PREC_* values and four functions that the NIF does not call. sys/build.rs needs no change.

Upstream changes in src/ and include/ that reach the NIF:

• whisper : run abort_callback on the language auto-detect path ggml-org/whisper.cpp#4077: whisper_full_with_state passes the abort callback to the language auto-detection. The NIF runs the detection inside state.full when the language is nil or "auto" on a multilingual model, so an :abort_handle now stops it too. An abort there makes full() fail, and transcribe.rs already returns the partial result when the flag is set.
• whisper : add validation for ttype on model load ggml-org/whisper.cpp#4091: the model and VAD loaders reject a tensor type outside 0..GGML_TYPE_COUNT. Before, a malformed file indexed type_traits[] out of bounds in load_model/2.
• vad : reject n_encoder_layers other than 4 in model load ggml-org/whisper.cpp#4064: the VAD loader rejects n_encoder_layers other than 4. Before, a VAD model with fewer layers read past a heap array through :vad_model_path.
• whisper : fix int overflow in whisper_full_parallel chunk offsets ggml-org/whisper.cpp#4044 fixes an overflow in whisper_full_parallel, which the NIF does not call.
• whisper : optional ANEForge encoder backend (Apple Neural Engine) ggml-org/whisper.cpp#3905 adds an ANEForge encoder. It is a stub on Linux and Intel Macs. On Apple Silicon, it runs only when the environment sets ANEFORGE_ENCODER, and it loads its dylib with dlopen, so the link inputs of the Metal build do not change. The changelog says so.

The rest of v1.9.5 is ggml backend work (CUDA, Vulkan, Metal, SYCL, WebGPU). The CPU baseline flags are the same: Adding CPU backend variant ggml-cpu: -msse4.2;-mf16c;-mfma;-mbmi2;-mavx;-mavx2.

Skipped: no refactor applies. The NIF never calls the language-detection API directly, and whisper-rs gets no new API.

Upstream whisper-rs is still at 0.16.0 of 2026-03-12, with #277, #278, and #271 open, so issue #26 stays open.

Only the CPU variant was built locally. The CUDA, HIP, and Metal builds at v1.9.5 run in the release matrix, which is their gate.

Checks with WHISPER_CPP_BUILD=1 and GGML_NATIVE=OFF, against the pushed vendor rev: task check (format, compile, Credo, Clippy, 62 Elixir tests, 33 Rust tests, zizmor) and task test:integration (86 tests with real inference on ggml-tiny.en and ggml-tiny) pass.