stateLoadFile on a missing path aborts the Wasm core

[leehack]

After a model loads, stateLoadFile on a path that does not exist aborts the Wasm core instead of rejecting:

await bridge.stateLoadFile('/does_not_exist.bin', 512);
// main thread: RuntimeError: Aborted(undefined). Build with -sASSERTIONS for more info.
// worker: Error with the same text, and the bridge moves to the main thread for the session

It behaves the same on v0.1.44 and v0.1.47, in main-thread and worker modes, on wasm32 and wasm64.

llamadart_webgpu_state_load_file (src/llama_webgpu_core.cpp:1652 at 64ba825) passes the path straight to llama_state_load_file. Its own failure branch, which returns -5 with "Failed to load llama.cpp state file…", is never reached.

This matters because Web state paths are virtual and do not survive a page reload. So loading a path saved before a reload hits this.

Expected: a missing file rejects with an error, for example by checking that it exists first and returning a new code, and leaves the runtime usable.

Found while validating the bridge pin bump in leehack/llamadart#620.

[leehack]

A note for whoever fixes this: #119 changes the worker createCompletion fallback so that it moves to the main thread only when the worker itself is unusable. That means a crash, a timeout or stall, or an Aborted(…)/trap error. It no longer does so for every error.

The state APIs don't fall back. After stateLoadFile aborts the worker's core, the bridge keeps that dead worker. Until now, the next completion error of any kind then moved the session to the main thread. After #119, only errors recognised as worker failures do.

Returning an error instead of aborting, as this issue proposes, avoids the dead worker entirely. A more general guard would be for the runtime's onAbort to set a flag that stays set, sent back with the worker's error state, so the facade can see the core is dead whatever the error text says.