MarkdownDocumentReader emits boundary markers (<!-- HEADING_START:H2 -->,
<!-- TABLE_START -->, <!-- DOCUMENT_IMAGES_START -->, etc.) that are consumed
internally by boundary detection but leaked into final chunk Content, forcing
every consumer (search display, embedding input, LLM context) to strip them.

FluxCuratorChunkAdapter now removes all HTML comments at the single common
chunk-conversion point (after boundary detection, decoupled from it) and lifts
the heading level into Props[ChunkPropsKeys.HierarchyHeadingLevel] before
stripping, so no structural information is lost.

- Add ChunkPropsKeys.HierarchyHeadingLevel ("hierarchy.headingLevel")
- StripInternalMarkers + TryExtractHeadingLevel in FluxCuratorChunkAdapter
- 11 unit tests (structural + multimodal markers, level lift, blank-line cleanup)
- README: document clean chunk content + known split-marker limitation