ci: gate the public API against the published release, and stop the pins drifting

.github/workflows/ci.yml

@@ -44,6 +44,54 @@ jobs:
           name: termlens-report-ci-${{ matrix.os }}
           cli-version: "0.10.1" # checked against the manifest by check-skill-version.sh
 
+  # The public API against the last published release.
+  #
+  # `--release-type` is FORCED, never inferred, and the reason is measured
+  # rather than assumed. Removing `Config::kernel` from launchbound-space
+  # while bumping the manifest to 3.0.0 in the same PR:
+  #
+  #   inferred        ->  "0 checks: 0 pass, 254 skip", exit 0
+  #   --release-type patch ->  "1 major and 0 minor checks failed", exit 100
+  #
+  # cargo-semver-checks compares the manifest version to the baseline and
+  # runs only the lints that bump would not already excuse. Declare the major
+  # bump in the same commit as the break and it has nothing left to say. That
+  # is the hole: the gate agrees with whatever the PR claims about itself.
+  #
+  # `patch` means "no API change of any kind is excused", so every lint
+  # fires. It does NOT forbid additions -- cargo-semver-checks only reports
+  # breaking changes, so a new `pub fn` passes (verified: 223 checks, 223
+  # pass, with a `pub use` added).
+  #
+  # A deliberate break carries the `breaking` label, which switches this to
+  # `major`. Three of the eleven published crates are binary-only and have no
+  # library API; the eight that do are all covered by --workspace.
+  #
+  # `baseline-version` is a literal, moved in the release PR after publish --
+  # left at the old release it would compare against a version nobody can
+  # install, and moved before publish it would compare against one that does
+  # not exist yet.
+  semver:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Install pinned toolchain from rust-toolchain.toml
+        run: rustup show
+      # A prebuilt binary, downloaded every run: never built from source, and
+      # never restored from a cache. A gate that reports on whatever a cache
+      # held the day it was filled is not reporting on this PR.
+      - uses: taiki-e/install-action@v2
+        with:
+          tool: cargo-semver-checks
+      - name: Compare the public API against the published baseline
+        env:
+          RELEASE_TYPE: ${{ contains(github.event.pull_request.labels.*.name, 'breaking') && 'major' || 'patch' }}
+        run: |
+          echo "release-type: $RELEASE_TYPE"
+          cargo semver-checks --workspace \
+            --baseline-version 2.1.0 \
+            --release-type "$RELEASE_TYPE"
+
   # MSRV applies to the crates that do not need the pinned nightly
   # (CONTRIBUTING.md); checked against the committed lockfile.
   msrv:

Cargo.toml

@@ -33,14 +33,19 @@ keywords = ["cuda", "gpu", "autotuner", "convergence", "kernel"]
 categories = ["development-tools", "development-tools::profiling"]
 
 [workspace.dependencies]
-launchbound-space = { path = "crates/launchbound-space", version = "2.0.0" }
-launchbound-prune = { path = "crates/launchbound-prune", version = "2.0.0" }
-launchbound-build = { path = "crates/launchbound-build", version = "2.0.0" }
-launchbound-bench = { path = "crates/launchbound-bench", version = "2.0.0" }
-launchbound-report = { path = "crates/launchbound-report", version = "2.0.0" }
-launchbound-search = { path = "crates/launchbound-search", version = "2.0.0" }
-launchbound-model = { path = "crates/launchbound-model", version = "2.0.0" }
-launchbound-metal = { path = "crates/launchbound-metal", version = "2.0.0" }
+# These `version =` pins must equal `workspace.package.version` above. They
+# are what a crates.io consumer resolves against, so a stale pin says a
+# published crate is compatible with a release it was never built against --
+# harmless for a path build, wrong as a record. They sat at 2.0.0 through the
+# whole 2.1.0 line; `just versions` (and a test) now refuse the drift.
+launchbound-space = { path = "crates/launchbound-space", version = "2.1.0" }
+launchbound-prune = { path = "crates/launchbound-prune", version = "2.1.0" }
+launchbound-build = { path = "crates/launchbound-build", version = "2.1.0" }
+launchbound-bench = { path = "crates/launchbound-bench", version = "2.1.0" }
+launchbound-report = { path = "crates/launchbound-report", version = "2.1.0" }
+launchbound-search = { path = "crates/launchbound-search", version = "2.1.0" }
+launchbound-model = { path = "crates/launchbound-model", version = "2.1.0" }
+launchbound-metal = { path = "crates/launchbound-metal", version = "2.1.0" }
 anyhow = "1"
 clap = { version = "4", features = ["derive"] }
 serde = { version = "1", features = ["derive"] }

docs/RELEASING.md

@@ -22,9 +22,13 @@ maintainer moving between them is not relearning the process.
 gh workflow run stress.yml -f iterations=100
 gh run watch                    # ten shards, both OSes
 
-# 1. Bump the version. It appears once per crate plus the workspace pins.
-$EDITOR Cargo.toml              # version = "X.Y.Z"
+# 1. Bump the version. `workspace.package.version` AND the internal
+#    `version =` pins in [workspace.dependencies] -- both, or `just
+#    versions` fails. They are not cosmetic: pins left behind a major
+#    bump make `cargo metadata` refuse to resolve the workspace at all.
+$EDITOR Cargo.toml              # version = "X.Y.Z", and the eight pins
 cargo check --workspace         # refreshes Cargo.lock
+just versions                   # the two agree
 
 # 2. Move the CHANGELOG section: [Unreleased] -> [X.Y.Z] - YYYY-MM-DD,
 #    leaving an empty [Unreleased] above it.
@@ -70,6 +74,20 @@ tag had to be pushed by hand.
   ```sh
   gh workflow run install.yml
   ```
+- **Move the semver baseline, in a PR of its own after the publish.**
+  `baseline-version` in `ci.yml`'s `semver` job is a literal. Left at the old
+  release it compares every PR against a version nobody can install any more,
+  and it would also carry this release's own breaks forward as if they were
+  new. Moved *before* the publish it names a version that does not exist yet
+  and the job cannot fetch it. So: publish, confirm the index has it, then
+  bump the literal.
+
+  ```sh
+  $EDITOR .github/workflows/ci.yml   # baseline-version: X.Y.Z
+  ```
+- **A break needs the `breaking` label on its PR**, which switches the semver
+  job from `patch` to `major`. Without it the job fails, which is the point;
+  with it, the release notes owe the reader a migration note.
 
 ## What a version number means here
 

justfile

@@ -6,7 +6,7 @@ default: ci
 # "all required jobs green" aggregator job in ci.yml — this recipe is the
 # aggregator, and the `ci` job runs it verbatim on both OSes — so a new gate
 # becomes required by being listed here.
-ci: fmt-check clippy test deny schemas pins skill
+ci: fmt-check clippy test deny schemas pins versions skill
 
 # Cargo errors on a memberless virtual workspace, so the cargo recipes no-op
 # until the first crate lands in S1. `grep -c` prints 1 when packages is empty.
@@ -48,6 +48,13 @@ prune cc="8.6":
 pins:
     ./scripts/check-pins.sh
 
+# The workspace version, every crate's version, and the internal `version =`
+# pins in [workspace.dependencies] agree. The pins sat at 2.0.0 through the
+# whole 2.1.0 line: harmless for a path build, wrong as a record, and fatal
+# to the next major bump (`cargo metadata` refuses to resolve).
+versions:
+    ./scripts/check-versions.sh
+
 # Golden + JSON Schema validation of report documents (S4).
 schemas:
     cargo test -p launchbound-report --test schema_and_golden

scripts/check-versions.sh

@@ -0,0 +1,67 @@
+#!/usr/bin/env bash
+# CI gate: the workspace version, every crate's version, and the internal
+# `version =` pins in [workspace.dependencies] all agree.
+#
+# The pins are what a crates.io consumer resolves against. `launchbound-prune`
+# 2.1.0 depending on `launchbound-space = "^2.0.0"` says it works with a
+# release it was never built or tested against -- harmless for a path build,
+# which is why it went unnoticed through the whole 2.1.0 line, and wrong as a
+# record. It is not only cosmetic: a `3.0.0` workspace bump with the pins left
+# at `2.0.0` fails `cargo metadata` outright ("candidate versions found which
+# didn't match: 3.0.0"), so the drift breaks the next major release rather
+# than the current one.
+#
+# Portability: macOS ships bash 3.2 -- no `declare -A`, no `local -n`, no
+# GNU-only sed. (scripts/check-pins.sh, which learned this the hard way.)
+set -euo pipefail
+
+ROOT=$(cd "$(dirname "$0")/.." && pwd)
+cd "$ROOT"
+
+status=0
+
+# `workspace.package.version` -- the one everything else must match. Anchored
+# to the section so a dependency's `version =` cannot be read by mistake.
+workspace_version=$(
+  sed -n '/^\[workspace\.package\]/,/^\[/s/^version = "\(.*\)"/\1/p' Cargo.toml | head -1
+)
+if [ -z "$workspace_version" ]; then
+  echo "VERSION GATE: could not read workspace.package.version from Cargo.toml" >&2
+  exit 1
+fi
+echo "workspace.package.version: $workspace_version"
+
+# The internal path pins.
+pins=$(sed -n 's/^\(launchbound-[a-z-]*\) = { path = "[^"]*", version = "\([^"]*\)".*/\1 \2/p' Cargo.toml)
+if [ -z "$pins" ]; then
+  echo "VERSION GATE: no internal version pins found -- has [workspace.dependencies] changed shape?" >&2
+  exit 1
+fi
+
+count=0
+while read -r name pin; do
+  [ -z "$name" ] && continue
+  count=$((count + 1))
+  if [ "$pin" != "$workspace_version" ]; then
+    echo "VERSION GATE: $name is pinned at $pin, workspace is $workspace_version" >&2
+    status=1
+  fi
+done <<EOF
+$pins
+EOF
+echo "  internal pins checked: $count"
+
+# Every member crate inherits the workspace version rather than setting its
+# own; a literal here would drift silently.
+for manifest in crates/*/Cargo.toml; do
+  own=$(sed -n '/^\[package\]/,/^\[/s/^version = "\(.*\)"/\1/p' "$manifest" | head -1)
+  if [ -n "$own" ]; then
+    echo "VERSION GATE: $manifest sets version = \"$own\" instead of version.workspace = true" >&2
+    status=1
+  fi
+done
+
+if [ "$status" -eq 0 ]; then
+  echo "every recorded version agrees"
+fi
+exit "$status"