A position: fixed box normally uses the viewport as its containing block. CSS Positioned Layout lets an ancestor take that role by establishing a fixed positioning containing block, and CSS Transforms says any transform other than none does exactly that: the element becomes the containing block for all descendants, absolute and fixed alike, using its padding box.
So the "fixed" header is now positioned against the transformed ancestor. It scrolls with that ancestor, width: 100% means the ancestor's width, and since a transform also creates a stacking context, its z-index is trapped there and its ancestors' overflow can clip it.
The same happens with:
translate,rotate,scale(any value butnone, eventranslate: 0),perspective,transform-style: preserve-3dfilterandbackdrop-filterother thannone, except on the root elementcontain: layout,paint,strictorcontent, andcontent-visibility: auto, which turns on layout and paint containmentwill-changenaming any property above, even before it changes
Not on the list: opacity, isolation, mix-blend-mode, clip-path, overflow, z-index, and container-type (it applies size and style containment, not layout).
The spec states the rule without a rationale, but one way to see it: a transform maps its whole subtree into a new local coordinate system. A descendant drawn in that system can't also stay pinned to the untransformed viewport, so the transformed element becomes its reference.
Fixes: move the fixed element out of the transformed subtree (a portal to <body>), remove the transform once an animation finishes (transform: none releases it), or render it in the top layer as a popover or modal dialog, whose containing block is the viewport regardless of ancestors.