v0.4.0 — ship vLLM plugin for inference-time probe scoring

[caiovicentino]

Listed in Q3 2026 roadmap. ~5ms overhead target. Plugin should:

1. Hook into vLLM's LLM.generate() postprocess
2. Forward residual stream to loaded probe(s) at named layer
3. Return score alongside text in response

Reference: vLLM's extension API + agent-probe-guard sklearn probe interface.