build: build released wheels with one codegen unit per crate

.github/workflows/build-distributions.yml

@@ -34,6 +34,11 @@ on:
         required: false
         type: number
         default: 0
+      cargo-profile:
+        description: "Cargo profile of the native builds: dist for released wheels, release for quicker checks"
+        required: false
+        type: string
+        default: dist
     outputs:
       sha:
         description: "Commit the distributions were built from"
@@ -88,7 +93,8 @@ jobs:
     name: Build wheel (${{ matrix.platform.name }})
     needs: source
     runs-on: ${{ matrix.platform.runner }}
-    timeout-minutes: 75
+    # Cold dist builds take up to about 2x the 17-31 minute release-profile jobs on Windows and macOS.
+    timeout-minutes: 120
     strategy:
       fail-fast: false
       matrix:
@@ -141,7 +147,7 @@ jobs:
           command: build
           # One platform records a cargo build timing report for build time work.
           args: >-
-            --release --out dist --features extension-module
+            --profile ${{ inputs.cargo-profile }} --out dist --features extension-module
             ${{ matrix.platform.name == 'linux-x86_64' && '--timings' || '' }}
           target: ${{ matrix.platform.target }}
           manylinux: auto
@@ -159,7 +165,7 @@ jobs:
         with:
           command: build
           args: >-
-            --release --out dist --manifest-path crates/sqlbuild-cli/Cargo.toml --features ui
+            --profile ${{ inputs.cargo-profile }} --out dist --manifest-path crates/sqlbuild-cli/Cargo.toml --features ui
             ${{ matrix.platform.name == 'linux-x86_64' && '--timings' || '' }}
           target: ${{ matrix.platform.target }}
           manylinux: auto

.github/workflows/release-performance.yml

@@ -66,7 +66,10 @@ jobs:
         run: uv python install 3.12
       - name: Install dependencies
         run: uv sync --locked --all-groups --all-extras
+      # The candidate uses the released wheels' cargo profile, so both sides of the comparison match.
       - name: Build candidate wheel
+        env:
+          MATURIN_PEP517_ARGS: --profile dist
         run: |
           uv build --wheel --out-dir "${RUNNER_TEMP}/candidate"
           uv build --wheel crates/sqlbuild-cli --out-dir "${RUNNER_TEMP}/candidate"

.github/workflows/verify.yml

@@ -1099,6 +1099,8 @@ jobs:
       sdist: false
       upload: true
       retention-days: 3
+      # Release Distributions builds and smoke tests the slower dist profile for every release.
+      cargo-profile: release
 
   compose:
     name: Packaged Compose lifecycle

Cargo.toml

@@ -203,8 +203,14 @@ debug = false
 [profile.release]
 strip = "symbols"
 
+# Released wheels: one codegen unit per crate keeps one copy of generic code that parallel units
+# each instantiate, so the native binaries are about an eighth smaller and build in about 2x time.
+[profile.dist]
+inherits = "release"
+codegen-units = 1
+
 # Local iteration: release-level optimisation with incremental builds, so rebuilding after a
-# focused change does not recompile the whole binding crate. CI and wheels use `release`.
+# focused change does not recompile the whole binding crate. CI uses `release`.
 [profile.fast]
 inherits = "release"
 incremental = true