A grammar-constrained completion can abort the Wasm core #115

A grammar-constrained createCompletion can abort the Wasm core instead of returning text or an error:

Aborted(undefined). Build with -sASSERTIONS for more info.

In worker mode the abort also moves the bridge to the main thread for the rest of the session (__llamadartBridgeWorkerFallbackReason holds the abort text).

Repro (same on v0.1.44 and v0.1.47), Qwen3.5-0.8B Q4_K_M on the CPU:

await bridge.loadModelFromUrl(modelUrl, { nCtx: 1024, nThreads: 2, nGpuLayers: 0 });
await bridge.createCompletion(
  '<|im_start|>user\nIs the sky blue? Answer yes or no.<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n',
  { nPredict: 8, temp: 0, topK: 1, seed: 1, grammar: 'root ::= "yes" | "no"' },
);

• The same grammar also aborts with topK: 40, temp: 0.8.
• A small JSON grammar aborts with topK: 1 and succeeds with topK: 40.
• stories15M with the yes/no grammar and topK: 1 aborts too.

Likely cause: create_sampler (src/llama_webgpu_core.cpp:807-822 at 64ba825) adds the grammar sampler after top-k and top-p. When truncation keeps only tokens the grammar rejects, no candidate is left. llama.cpp's common_sampler avoids this: it checks the sampled token against the grammar and, if rejected, resamples with the grammar applied first.

Expected: a grammar-constrained completion returns text the grammar accepts, or rejects with an error the caller can handle.

Found while validating the bridge pin bump in leehack/llamadart#620. llamadart passes GenerationParams.grammar and topK through unchanged.