[kvlonge]

Why

Every release adds about 230 MB to PyPI, which hit its 25 GB limit with 0.195.0. The native extension is almost the whole wheel (43 of 45 MB compressed); it is already stripped, so the bulk is generic code that the default 16 parallel codegen units each instantiate.

Changes

• New dist cargo profile: release with codegen-units = 1.
• Build Distributions takes a cargo-profile input, defaulting to dist, for both native builds. Release Distributions and Publish therefore build released wheels with it unchanged.
• Verify's pull request wheel job keeps release, so its build time does not double.
• Wheel job timeout 75 to 120 minutes: cold Windows and macOS jobs take 17-31 minutes today, and dist about doubles the native builds.
• Release Performance builds its candidate with dist, like the published baseline.

Rejected: fat LTO with 1 unit gave 40.1 MB at 3.1x build time and 7.9 GB peak; fat LTO with 16 units saved 0.4%.

DuckShop: not applicable. This is a build profile; compiled output is identical.

Verification

Same orb, rustc 1.99.1, linux x86_64, cold builds:

	release	dist
sqlbuild wheel	45.03 MB	39.29 MB (-12.7%)
sqlbuild-cli wheel	14.55 MB	12.16 MB (-16.4%)
extension build wall	442 s	917 s
extension build user CPU	2471 s	2004 s
peak build RSS	5.5 GB	5.9 GB

• Dense 3,000-model compile, 5 alternating runs, medians: cold wall 8.16 to 8.12 s (CPU 0.996x), cold writing the cache 1.002x, warm 0.25 s both. Compiled files and compile JSON excluding timings are identical.
• Test-built wheels in a clean Python 3.12 venv: native import, sqb playground pg compile and build PASS=99, sqb server serving /, deep links, /login.
• uv build with MATURIN_PEP517_ARGS=--profile dist runs maturin pep517 build-wheel --profile dist and builds into target/dist.
• The first Release Distributions run measures dist on every platform.
• make check-ci, public-repository scan.