v0.25.0

Flex geometry, corrected in four places where the engine measured items by their border box and CSS measures them by their outer box. All packages share the line: engine (crates.io forme-pdf), every @formepdf/* npm package, formepdf (PyPI), forme-go (tag v0.25.0), and the VS Code extension.

Both fixes came from a real document put side by side with another renderer, and both were measured against Chrome rather than argued from the spec alone.

Changed: these move existing documents

• An item taller than a fixed-height row overflows it equally, not downward. CSS Flexbox 8.3: with align-items: center, an item whose cross size exceeds the line's "will overflow equally in both directions". The single-line cross-size rule kept the larger of the container's inner size and the content's, so a container SMALLER than its content kept content size: the line grew past the box, the centring offset evaluated to zero, and the item sat at the top with every overflowing point going downward. A status pill with height set to align with its neighbours dropped its text onto the bottom border. Chrome overflows 1.26pt above and below on the reported geometry; Forme now splits it the same way. Centring in a box with room to spare is unchanged
• A flex item's margins are main-axis space no sibling can take. CSS Flexbox 9.7 resolves flexible lengths against each item's OUTER hypothetical main size. Four places used the border box instead, and each compounded the last: free space, justify-content slack, line packing (9.3, which collects items into lines by outer size), and the per-item cursor advance. A document header with a flex: 1 title beside a fixed logo with marginLeft pushed the logo clean out of the row by exactly that margin, over the page's content edge. Wrapping rows over-packed their lines: six 150pt boxes with 8pt margins fit three to a line by width alone but only two by outer size, which is what a browser does, and the over-packed line then handed justify-content negative slack to distribute

Fixed

• forme-go is on the shared version line. It previously ran its own v0.9.x tags and went three engine releases without one, because a separate line gives no signal that the SDK is stale. RELEASE.md says so now, along with the rule that main is pushed before the tag: branch protection routes a release through a PR, so tagging first points the tag at a commit the remote does not have

Internal

• scripts/verify-sdk-wasm.sh builds the engine WASM fresh and compares it against the copies the Python and Go SDKs embed. Confirming that a rebuild COMMIT exists proves nothing about when it was built: at 0.24.0 the Go SDK carried one dated the day before the release, from an engine missing two of that release's changes. It has since caught the same drift on every engine change
• Three byte-wall fixtures for shapes the wall could not see: fixed-height badges, flex margins, and wrapped float rows