fix: sync bug fixes from the WasmEdge MLX plugin - #7

Summary

Ports bug fixes made to plugins/wasi_nn/MLX in WasmEdge back to this repo. WasmEdge-specific changes (namespaces, license headers, log prefixes, sv suffixes, WASI VFS) are left out.

Changes

• whisper: remove shell command injection in loadAudio (cf. WasmEdge 4742d02fba)
  loadAudio joined the audio path into a command string and passed it to popen(), which runs it through /bin/sh. Shell metacharacters in the path (e.g. a.wav; rm -rf ~) ran arbitrary commands, and paths containing spaces broke.
  ffmpeg now runs via posix_spawnp with an argv array, so no shell is involved. Unlike WasmEdge, file-path input is kept because example/whisper.cpp and library users rely on it.

• vlm: stop generation at EOS
  generate() always ran to MaxTokens. This adds a per-model virtual eosTokenIds(), and Gemma3 returns <eos> (1) and <end_of_turn> (106).
  WasmEdge hard-codes Token == 1 || Token == 106 in the shared loop. That would break mllama, where token 1 is " in the Llama 3 vocab, so the stop IDs are per-model here.

• loadBytesFromFile: don't kill the process on failure (cf. WasmEdge 584437baa5)
  It now logs an error and returns an empty string instead of calling exit(1).

• Rename shadowed identifiers (cf. WasmEdge e271503856)
  Renames parameters and locals that shadowed members or outer names in Module, Embedding/Linear, pooling, RotatingKVCache, and the whisper decoding and tokenizer code. No behavior change.

• spdlog formatting (cf. WasmEdge 2b8429d303)
  Values are passed as {} format arguments instead of being concatenated into the format string.

• ci: run macOS build on macos-latest
  GitHub no longer provides the macos-14 runner (The configuration 'macos-14-arm64-us-default' is not supported), so the macOS job was cancelled before it started.

• build: drop unused fmt dependency
  CMakeLists.txt fetched and compiled fmt 10.2.1 but never linked or included it; spdlog uses its bundled copy. That fmt version fails to compile with the newer clang on macos-latest (consteval / FMT_STRING errors), so it is removed.

Not ported

• WasmEdge's switch from const T& back to by-value parameters. This repo intentionally moved to const references in f1fffaf.
• An unused flatten added to QuantizedEmbedding::forward in WasmEdge. It has no effect.
• The Gemma3 attention-mask take(...) difference. Both versions look inconsistent with the Python reference mask[..., -K:], so this needs a separate look.

Testing

• With clang + Ninja (the CI setup), the project sources compile: 29 of 30 translation units, including example/vlm.cpp, which includes the changed gemma3.h.
• model/gemma3/gemma3.cpp fails locally with the same error on main. It's a local clang / GCC 15 libstdc++ issue and unrelated to this PR.
• The full local build stops in third-party deps (sentencepiece, fmt) because the local compiler is newer, so the examples were not run end to end. Relying on CI for the full build.

🤖 Generated with Claude Code