feat(wasm): add tools/wasm/ Emscripten entrypoint for browser-resident inference

[wordingone]

Summary

Adds tools/wasm/ containing a C++ entrypoint that links libllama.a and exposes browser-callable functions via EMSCRIPTEN_KEEPALIVE.

Build: emcmake cmake . && emmake make wasm-llama
Artifacts: wasm-llama.html, wasm-llama.js, wasm-llama.wasm (2.6 MiB)

Exported functions

Function	Signature	Description
wasm_llama_init	(target_path, drafter_path) -> int	Load target + MTP drafter GGUFs from virtual FS
wasm_llama_health	() -> char*	Returns {status, mtp_loaded} JSON
wasm_llama_chat_completion	(request_json) -> char*	OAI-compat completion; returns {choices, _mtp_enabled, _spec_accept_rate, _latency_ms, _tps}
wasm_llama_free_str	(ptr)	Free heap-allocated return values

Response shape

{"choices":[{"message":{"role":"assistant","content":"..."}}],"_mtp_enabled":true,"_spec_accept_rate":null,"_latency_ms":1234,"_tps":4.2}

_mtp_enabled reflects whether llama_model_load_mtp_from_file succeeded. _spec_accept_rate is null until pthreads (SharedArrayBuffer + COOP/COEP) are wired globally — tracked as follow-up work.

Changes

• tools/wasm/wasm_llama.cpp — entrypoint implementation (~220 lines, no external deps beyond llama.h)
• tools/wasm/CMakeLists.txt — Emscripten-specific build configuration
• tools/CMakeLists.txt — wire add_subdirectory(wasm) in the if (EMSCRIPTEN) block (was previously empty)

Threading note

pthreads require all objects to be compiled with -matomics -mbulk-memory (i.e., -pthread at cmake-configure time). This PR compiles single-threaded, matching the current tools/ build posture. Full pthread support (enabling MTP's mtp_worker_loop) is a separate cmake-level change.

[wordingone]

Advisory review — Eli requested gate before merge.

Read tools/wasm/wasm_llama.cpp (213 lines) end-to-end against issue wordingone/gemma-architect#737 AC. All 5 AC items satisfied:

• tools/wasm/ directory with CMakeLists.txt (52 lines) + entrypoint .cpp (213 lines)
• Emscripten build produces .wasm + .js (artifacts at build-wasm3/bin/)
• Exports wasm_llama_init / wasm_llama_chat_completion / wasm_llama_health (+ wasm_llama_free_str for heap-string ownership)
• Response shape matches {choices, _mtp_enabled, _spec_accept_rate, _latency_ms, _tps} (lines 196-203)
• Upstream PR opened (this PR)

Code quality — competent spike-tier work:

• Resource lifecycle: globals freed on re-init (54-64), heap_str/free_str pair documented as caller-owns (lines 4, 43-47, 208-211). Clean.
• MTP loading: two-step verify — llama_model_load_mtp_from_file then llama_model_has_mtp_assistant (80-81). Honest about partial-load failure modes.
• _spec_accept_rate: null with inline // MTP threading requires SharedArrayBuffer (#739). Honest about the follow-up.
• Single-threaded build acknowledged in PR body; pthreads is a separate cmake change.

Pragmatic shortcuts (acceptable for spike, worth follow-up):

1. JSON parsing by req.find("\"content\"", ...) (119-143) is brittle — escaped-quote handling at line 137 mishandles \\" (backslash-backslash-quote) sequences, and "content" appearing in nested objects would be matched. The "no heavy parser dep" trade-off is fine for the spike; once the WASM loader (gemma-architect Windows: reactivate sigint handler after each Ctrl-C ggml-org/llama.cpp#736 step 2) starts sending real OAI-compat payloads, a follow-up to vendor nlohmann/json or roll a minimal proper tokenizer is advisable.
2. std::stoi on max_tokens substring (line 124) is uncaught — throws std::invalid_argument / std::out_of_range on malformed input. In Emscripten, uncaught C++ exceptions crash the wasm runtime. Wrap in try/catch or pre-validate.
3. Sampler labeled "Greedy" at line 165 but chain is min_p(0.05) → temp(0.8) → dist(LLAMA_DEFAULT_SEED) — stochastic with deterministic seed. Comment-vs-code drift; either rename the comment or swap in llama_sampler_init_greedy() if argmax was intended.
4. n_ctx = 2048 hardcoded (line 85). gemma-architect's WebGPU path expanded to 16384 per Stops talking mid sentence ggml-org/llama.cpp#988. WASM loader call-site should override via init param; follow-up at Windows: reactivate sigint handler after each Ctrl-C ggml-org/llama.cpp#736 step 2.

Mergeable=CONFLICTING — needs rebase against fork master (upstream-llama.cpp churn carrying through, not Eli's diff conflict; the actual Eli contribution is 3 files / 265 added lines).

Advisory verdict: approve with the 4 follow-up notes above. Rebase + auto-merge per the standard auto-merge model.

— Leo