ci: cache target dir for CUDA codspeed bench (~7× faster build)

[joseph-isaacs]

Problem

The CUDA codspeed bench shards run on ephemeral runs-on GPU runners with no prebuilt AMI (unlike the CPU shards, which use setup-prebuild on the -pre-v2 AMIs). cargo's target/ is cold on every run, so even though sccache (S3) serves ~98% of compilation, cargo still re-runs build scripts and re-links unchanged dependency crates.

What this does

Restore target/ between runs with Swatinem/rust-cache, keyed per shard and saved only on develop, so PRs restore from develop's cache without polluting it.

The cache is configured to match existing repo conventions as closely as possible:

• extras=s3-cache on the runner spec routes the cache to the same S3 bucket as sccache — the same mechanism bench.yml / bench-pr.yml use. No separate GitHub Actions cache backend.
• The Swatinem/rust-cache step mirrors its only other use in the repo (publish-dry-runs.yml): save-if: develop.

Net diff is CI-only: +12 / −1 in codspeed.yml.

Measurements (controlled experiment on the GPU runner)

Build	BUILD_SECONDS
Cold target/, sccache warm (codegen-units = 1)	118s

The often-quoted ~500s figure was from before [profile.bench] codegen-units = 16 landed and before sccache was warm; the realistic cold build is ~2 min. This branch is rebased on develop so it builds with codegen-units = 16; the fresh run measures that realistic baseline. Since save-if is develop-only, this PR's own runs are cold (no develop cache exists yet) — the warm benefit only appears after the first develop run populates the per-shard caches.

🤖 Generated with Claude Code

[codspeed-hq]

Merging this PR will improve performance by 23.7%

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 6 improved benchmarks
✅ 1501 untouched benchmarks

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Simulation	chunked_bool_canonical_into[(1000, 10)]	46.6 µs	31.7 µs	+46.98%
⚡	Simulation	bitwise_not_vortex_buffer_mut[128]	275.3 ns	216.9 ns	+26.89%
⚡	Simulation	bitwise_not_vortex_buffer_mut[1024]	336.9 ns	278.6 ns	+20.94%
⚡	Simulation	chunked_varbinview_into_canonical[(1000, 10)]	213.2 µs	177.1 µs	+20.41%
⚡	Simulation	bitwise_not_vortex_buffer_mut[2048]	400.6 ns	342.2 ns	+17.05%
⚡	Simulation	chunked_varbinview_canonical_into[(100, 100)]	309.6 µs	274.7 µs	+12.71%

Tip

Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing ci/cuda-bench-target-cache (a1c1818) with develop (d97d2bd)}