chore: release v0.4.22

[MagicalTux]

🤖 New release

• fstool: 0.4.21 -> 0.4.22

Changelog

0.4.22 - 2026-08-17

Added

• (littlefs) read, write and in-place edits for lfs2 images
• (web) create, edit and download images in the browser
• (memedit) in-memory authoring — blank filesystems, partitioned disks

Fixed

• (ext) honour journal revoke records across transactions
• (ext) decode HTree directory roots and honour journal revoke blocks (#32)

Other

• (littlefs) redraw the tag and skip-list diagrams in our own form

Added

• (littlefs) new backend for the embedded-flash filesystem (lfs2, disk
  versions 2.0 and 2.1): read, write, and in-place edits. Metadata pairs are
  replayed from their CRC-committed logs (tags, splices, tails, global-state
  deltas, lfs2.1 forward-CRCs) and written back as compactions; files live
  inline in metadata or in CTZ skip-lists, which are rebuilt only from the
  first changed block onwards so a partial write leaves earlier blocks
  untouched. Block allocation reconstructs the in-use map by traversing the
  volume, as littlefs itself does. Wired into create -t littlefs, the TOML
  spec (type = "littlefs", with block_size / block_count / prog_size /
  version / name_max / inline_max options), repack --fs-type littlefs,
  info, add/rm, the in-memory authoring surface, and the browser build.
  littlefs user attributes surface as user.littlefs.<type> extended
  attributes; the format has no symlinks, device nodes or POSIX metadata, so
  those are refused rather than faked. Cross-validated in both directions
  against the reference C implementation via littlefs-python, including
  handing an image back and forth mid-edit.

• (memedit) new in-memory authoring surface: Workspace formats a blank
  filesystem or lays out a partitioned disk (MBR/GPT), takes files and
  directories, and hands back the image bytes at any point.
  creatable_filesystems() advertises 15 types with their real minimum sizes.

• (wasm) Workspace and creatable_filesystems() bindings, so the browser
  build can author images as well as read them.

• (web) "Create a new image" mode: pick a filesystem and size, or build a
  partitioned disk with a filesystem per partition, then add files, delete
  them, browse directories, and download the image — repeatedly, while
  continuing to edit. An uploaded image can also be switched into edit mode.

Fixed

• (ext) journal recovery now honours revoke records across transactions.
  Replay was single-pass and scoped each revoke to the transaction that
  carried it, so a block revoked in transaction N+1 was still replayed from
  transaction N — it had already been written by the time the revoke record
  was read. That is exactly the case revoke records exist to prevent (a
  metadata block freed and reused as file data), so stale metadata could land
  on live data. Recovery now runs the two passes JBD2 requires: a scan that
  collects committed transactions and builds a revoke table keyed by the
  highest revoking transaction id, then a replay that skips any block whose
  revoke id is at or after the transaction replaying it (the kernel's
  jbd2_journal_test_revoke rule, wrap-safe). An uncommitted tail
  transaction's revoke records are discarded along with its writes.
• (ext) parse_descriptor_tags walked the tag array in 8-byte steps
  regardless of the journal's real tag size, so a trailing slot too short for
  a 16-byte checksum-v3 tag failed the whole replay instead of ending the
  array; it also read into the 4-byte checksum tail that checksum-v2/v3
  descriptor blocks carry.

---

This PR was generated with release-plz.