65_spike-wasm-embedded-inference.md

id	65
title	Spike WASM-Embedded Weasel Inference
status	✅

Spike WASM-Embedded Weasel Inference

Goal

Evaluate a WASM-based inference path that can be embedded in mdsmith with no runtime dynamic library dependency.

Tasks

1. Choose one Go-hosted WASM runtime candidate (wazero v1.11.0 with the default wazevo optimizing compiler) and define module loading strategy.
2. Build a minimal proof of concept that embeds a .wasm artifact with go:embed and runs inference in-process. The guest reuses the plan-64 classifier package verbatim, compiled with GOOS=wasip1 GOARCH=wasm -buildmode=c-shared.
3. Verify deterministic output behavior on a fixed corpus and fixed model parameters. Five in-process and five cross-process runs produced the same digest.
4. Measure CPU latency and memory overhead versus current MDS029 heuristic, pure-Go spike, and yzma spike baselines.
5. Measure binary-size impact with embedded .wasm artifact.
6. Define artifact update workflow and integrity checks (SHA256 pin on embedded wasm bytes, wazero validates on load).
7. Document fallback boundaries when WASM init or inference fails.

Findings and recommendation live in the spike README.

Acceptance Criteria

• Prototype runs with no YZMA_LIB and no external dynamic libs.
• Embedded WASM artifact loads via go:embed.
• Deterministic behavior is confirmed across repeat runs.
• CPU latency and memory metrics are captured.
• Binary-size impact is measured and documented.
• Recommendation is made: reject this path at the current classifier size. Keep the pure-Go classifier (plan 64) as the CPU fallback for MDS029.