Concerns: pushword/core
The move, the cache purge and the preview now run once the row is committed, so a flush that throws afterwards can no longer leave the database naming a file that was already renamed away. Nothing to do.
On the gd driver only: a master whose bitmap (width × height × 4 bytes) does not fit in what is left of memory_limit no longer gets its dimensions and main color read at rename time — that allocation is a fatal error, not an exception. The background pw:image:cache fills them in.
Affected: sites on gd with masters above ~20 Mpx and a memory_limit under 512M. If that background task does not run on your host, run pw:image:cache to fill the metadata, or raise memory_limit.
pw:image:cache no longer counts your non-images as processedThe sequential run derived its processed count as total - skipped - errored, and a non-image media falls in none of those three: it only gets its public symlink, it never reaches the cache generator. Every PDF, GPX, XML and SVG in the library was therefore charged to processed on every run — a converged media library reported the same non-zero figure each night while regenerating nothing. The count is now tallied from the derivatives actually generated, in both the human summary and the --format=agent JSON (processed, generated).
Affected: nothing to do, but expect the number to drop. If you graph or alert on it, rebaseline: the run that used to say 52 processed on a stable library now says 0. The parallel run (-p above 1) was already correct and is unchanged.