StreamingAiGuard: wraps Stream<String> from streaming LLM responses,
buffers chunks, scans at configurable boundaries (default newline),
yields GuardedChunk per segment. Terminates on block. PII rehydration
per-chunk. StageRun + runInputStage/runOutputStage exposed as public API.

GroundingScanner: keyword-overlap heuristic checking LLM output against
source context. Extracts content words (non-stop-words), computes overlap
ratio, flags below threshold. Default action: warn.