Reduce Cargo build artifact size

.github/workflows/ci.yml

@@ -38,10 +38,16 @@ jobs:
       - uses: Swatinem/rust-cache@v2
 
       - name: Run tests
-        run: cargo test
+        run: cargo test --profile ci
 
       - name: Run facade README doctests
-        run: cargo test -p strided-rs --doc
+        run: cargo test --profile ci -p strided-rs --doc
+
+      - name: Report Cargo artifact usage
+        if: always()
+        run: |
+          du -sh target
+          df -h target
 
   scripts:
     name: maintenance scripts

Cargo.toml

@@ -47,3 +47,18 @@ rand_distr = "0.4"
 rayon = "1.10"
 smallvec = "1"
 thiserror = "1.0"
+
+# Workspace tests instantiate large generic kernels in many executables.
+# Omitting debuginfo keeps those artifacts small while preserving debug
+# assertions and overflow checks from Cargo's default development profiles.
+[profile.dev]
+debug = 0
+
+[profile.test]
+debug = 0
+
+# Hosted CI starts from cold caches, where incremental state has no value.
+[profile.ci]
+inherits = "test"
+incremental = false
+strip = "symbols"

docs/2026-08-09-cargo-artifact-reduction.md

@@ -0,0 +1,102 @@
+# Cargo Build-Artifact Reduction
+
+## Test-profile debuginfo
+
+On 2026-08-09, a cold `cargo test --workspace --no-run` comparison measured
+the effect of disabling debuginfo in the test profile. Both builds used fresh
+target directories, four Cargo jobs, disabled incremental compilation, and an
+empty `RUSTC_WRAPPER`.
+
+Environment:
+
+- macOS Darwin 25.5.0 on Apple Silicon (`arm64`)
+- `rustc 1.97.0 (2d8144b78 2026-07-07)`
+- `cargo 1.97.0 (c980f4866 2026-06-30)`
+
+Commands:
+
+```bash
+env CARGO_TARGET_DIR=<fresh-target> CARGO_BUILD_JOBS=4 \
+  CARGO_INCREMENTAL=0 RUSTC_WRAPPER= \
+  cargo test --workspace --no-run
+
+env CARGO_TARGET_DIR=<fresh-target> CARGO_BUILD_JOBS=4 \
+  CARGO_INCREMENTAL=0 CARGO_PROFILE_TEST_DEBUG=0 RUSTC_WRAPPER= \
+  cargo test --workspace --no-run
+```
+
+Results use allocated size from `du -sk`:
+
+| Configuration | Target size | Cold build time |
+|---|---:|---:|
+| Cargo default test profile | 1,717,360 KiB | 36.62 s |
+| Test profile with `debug = 0` | 824,372 KiB | 30.61 s |
+
+Disabling test debuginfo reduced allocated target size by 892,988 KiB
+(52.0%) and reduced this cold build sample by 6.01 seconds (16.4%). The test
+profile therefore sets `debug = 0`; debug assertions and overflow checks retain
+their Cargo defaults.
+
+The build produced 33 test executables. The hosted-CI profile and
+integration-test consolidation were measured independently below.
+
+## Development-profile debuginfo
+
+The same method was applied independently to `cargo build --workspace`, using
+`CARGO_PROFILE_DEV_DEBUG=0` for the experimental build.
+
+| Configuration | Target size | Cold build time |
+|---|---:|---:|
+| Cargo default development profile | 1,082,416 KiB | 26.18 s |
+| Development profile with `debug = 0` | 497,244 KiB | 23.59 s |
+
+Disabling development-profile debuginfo reduced allocated target size by
+585,172 KiB (54.1%) and reduced this cold build sample by 2.59 seconds (9.9%).
+The development profile therefore also sets `debug = 0`. Incremental
+compilation, debug assertions, and overflow checks retain their Cargo defaults.
+
+## Hosted-CI profile
+
+A named `ci` profile was then measured against the debug-free test profile:
+
+```toml
+[profile.ci]
+inherits = "test"
+incremental = false
+strip = "symbols"
+```
+
+Both test builds used fresh target directories and produced the same 33 test
+executables.
+
+| Configuration | Target size | Cold build time |
+|---|---:|---:|
+| Test profile with `debug = 0` | 824,372 KiB | 30.61 s |
+| CI profile with stripped symbols | 703,520 KiB | 30.98 s |
+
+The CI profile reduced allocated size by another 120,852 KiB (14.7%). The
+0.37-second time difference is noise for a single cold sample. Hosted test jobs
+use this profile and report `du` and `df` after testing so Linux and macOS runner
+usage remains visible. Coverage keeps its instrumentation-owned profile and is
+not stripped.
+
+## Rejected integration-test consolidation
+
+The `strided-kernel` crate has 18 integration-test executables. An experiment
+consolidated eleven general suites into one explicit harness while retaining
+the issue-specific Miri/source-contract targets and two process-isolated
+suites. This would have reduced the workspace from 33 test executables to 23.
+A fresh CI-profile comparison measured:
+
+| Configuration | Target size | Cold build time |
+|---|---:|---:|
+| Separate general integration targets | 703,520 KiB | 30.98 s |
+| Consolidated general integration target | 678,756 KiB | 29.93 s |
+
+The 24,764 KiB (3.5%) reduction did not justify changing focused `--test`
+commands and increasing harness coupling. The experiment also demonstrated
+two required isolation boundaries: `copy_plan_alloc` uses a process-wide
+allocation counter, while `execution_policy` tests Rayon and thread-local
+policy isolation. Both failed when unrelated tests ran concurrently in the
+same process. Integration-test consolidation was therefore rejected and the
+existing targets were retained.