v0.26.0

Added

• An input check can cover the non-text parts of a user turn — an instruction written inside an attached image is judged
  like the same words typed. GuardContext.Parts carries the turn's non-text parts (GuardInputPart: media type +
  bytes). A part is judged by the text a consumer-registered IInputPartTextExtractor reads from it (OCR, or the
  consumer's own vision model with a transcribe-only prompt; FluxGuardBuilder.WithPartTextExtractor, or an
  IInputPartTextExtractor service for AddFluxGuard), with the same input pipeline as the text. The most severe decision
  wins, TriggeredGuard.Part names the part that triggered (1-based; null = the text), and the block reason starts with
  Part N (media type):. A part no extractor reads is counted in the new GuardResult.UncheckedParts instead of passing
  silently; a failed extraction blocks in FailMode.Closed and counts as unchecked in Open. Checking stops at the first
  block. FluxGuard ships no extractor and takes no new dependency.

Changed

• Breaking (FluxGuard.SDK) — FluxGuardChatClient guards the images of the last user message too, and a user
  message with images and no text is now checked. It passes the message's inline images (DataContent) as
  GuardContext.Parts and calls IFluxGuard.CheckInputAsync(GuardContext) instead of CheckInputAsync(string, …); an image
  referenced by URI only (UriContent) is not fetched and not passed. Migration: an IFluxGuard test double behind the
  middleware sets up CheckInputAsync(Arg.Any<GuardContext>()); register an IInputPartTextExtractor to have the images
  judged (without one, the result's UncheckedParts counts them and the text verdict stands).

Dependencies

• Microsoft.Extensions.AI 10.10.0 -> 10.11.0. Test tooling: xunit.v3 4.0.2, MTP HangDump 2.5.1, coverlet.MTP 10.1.0.