4.0.0 — 2026-08-07

Release Notes

Configured release units were not in the dependency graph.

A [release_unit.<name>] block — explicit or glob-form — was registered as a
graph node and nothing else. The only two places that draw dependency edges ran
for [cascade_inputs] and for auto-discovered units, so in a repo where
both sides of a dependency are configured there was no edge between them at
all. A fix in a shared library reached no service, and prepare said
nothing_to_do — no error, because there was no unresolved dependency to
complain about. Nothing was missing; nothing had ever been drawn.

Everything below follows from that one hole, and from the three other things
that were quietly compensating for it.

Fixed

• Configured units get their dependency edges. Discovery now receives a
  map of the paths each [release_unit.X] covers, and routes the edges of any
  package inside them onto the owning unit. A hexagonal service spanning a bin
  crate plus a tree of satellites collects the dependencies of all of them;
  edges between its own crates collapse to nothing. Edges bind to a concrete
  unit id rather than a name, so a cargo crate can no longer bind to a
  same-named npm package, and an ambiguous name is an error instead of a
  first-match guess.
• satellites attribute their commits. The key is documented as "paths
  whose commits attribute to this unit" and reached the ownership map and the
  drift detector, but never the unit's path matcher. For a hexagonal service —
  where the released crates/bin is a thin wrapper and the work lands in
  crates/api, crates/core, crates/infrastructure — that means essentially
  every real change attributed to no unit at all.
• A configured unit is no longer also auto-discovered. Workspace protocols
  enumerate every member from one call on the workspace root, which the
  path-level skip-list could not filter. Both nodes landed on the same
  directory, and when the unit name equalled the package name the two were
  indistinguishable: multiple projects with same name.
• Cargo.lock is refreshed and committed. Configured units write their
  versions through a different rewriter than auto-discovered ones, and that
  one never went near the lockfile — in a repo where every cargo unit is
  configured, the lock was never updated at all. The sync now runs once per
  affected workspace after the rewrites, from what actually changed, and the
  lockfile is added to the release commit. A release commit carrying bumped
  manifests against a stale lock is not untidy, it is terminal: the next
  checkout regenerates the lock, and prepare refuses to run in a dirty tree.
• A failed lockfile update is an error. It used to be a warn! nobody
  reads, which is how the above went unnoticed for days.
• The Scoop bucket updates itself. update-scoop.yml listened on
  release: published, which never fires for a release created with the
  workflow's own GITHUB_TOKEN — GitHub's recursion guard. Two releases in a
  row shipped a stale manifest and were fixed by hand. It is now a job in
  release.yml, on the same tag push that already works.

Changed

• A cargo workspace counts as one project only if every member actually
  inherits its version. The test was the mere presence of
  [workspace.package].version, which is version inheritance being offered,
  not taken — practically every modern workspace sets it. That collapsed
  ordinary multi-crate repos into a single unit owning no [package], named
  after the repo directory, which nothing releases and which trips the
  untagged-unit guard.
• A gitignored Cargo.lock is left alone — not refreshed, not committed.

Migration

Both changes below surface config that was only ever describing a bug.

• If you added a [release_unit.<name>] block to suppress a unit named after
  your repo directory, delete it. That unit no longer exists, and a
  partial-override block matching nothing is a hard error.
• Expect more units to bump. Cascades that silently did nothing now work,
  including [cascade_inputs] reaching units transitively through a configured
  library. Run belaf status --ci before the first prepare to see the new
  set. If a unit bumps that you never intended to release, declare it:
  kind = "ignore" keeps it out of the release set entirely, kind = "internal" keeps it as a graph node that cascades but is never tagged.

Install belaf 4.0.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/ilblu/belaf/releases/download/v4.0.0/belaf-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/ilblu/belaf/releases/download/v4.0.0/belaf-installer.ps1 | iex"

Install prebuilt binaries via Homebrew

brew install ilblu/tap/belaf

Download belaf 4.0.0

File	Platform	Checksum
belaf-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
belaf-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
belaf-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
belaf-x86_64-pc-windows-msvc.msi	x64 Windows	checksum
belaf-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
belaf-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum