ci: run the multi-GiB slow tests in one job, not four

[joseph-isaacs]

Rationale for this change

The three multi-GiB regression tests — build_views_offsets_overflow_i32, compress_large_int and fsst_compress_offsets_overflow_i32 — were gated on #[test_with::env(CI)], and GitHub Actions sets CI in every job, so they ran four times per CI run: Rust tests (coverage), (windows-x64), (linux-arm64) and (linux-musl).

The coverage job is the worst host for them: -Cinstrument-coverage on an unoptimized build makes the multi-GiB allocation and fill loops far slower, and the coverage they contribute is negligible.

Two latent problems turned up along the way:

• test_with::env is evaluated at compile time, and the variable is not tracked by cargo or sccache — setting it without touching the source leaves the test ignored. Whether a job ran the slow tests therefore depended on the build environment rather than the run environment.
• The nextest override meant to protect them was a no-op: filter = 'test(compress_large_int | fsst_compress_offsets_overflow_i32)' matches nothing (| is not regex syntax inside test()), and the override carried no settings. They were running on the default 150s kill budget, which fsst_compress_offsets_overflow_i32 (~80s on arm64) was uncomfortably close to.

What changes are included in this PR?

Run them once, everywhere they used to run

• Gate the three with plain #[ignore], so every job skips them by default — a run-time decision rather than a compile-time one. #[test_with::no_env(VORTEX_SKIP_SLOW_TESTS)] stays alongside it, so setting that variable at build time still drops them from the binary, which is how the two sanitizer jobs avoid compiling those bodies under asan/tsan.
• Run them from a second step in Rust tests (linux-musl) with --run-ignored only, naming the three explicitly. That job is the one that covers every event: Rust tests (linux-arm64) is gated on github.repository == 'vortex-data/vortex' and is skipped on pushes to develop along with the windows and coverage jobs, so gating them there would leave fork PRs and develop untested.
• The step's cargo arguments match the step above it, so nothing is rebuilt. Naming the tests also matters: a bare --run-ignored all un-ignores the CUDA tests that vortex-cuda's test macro marks #[ignore] when the runner has no GPU (568 failures, seen on this branch).
• Fix the nextest override: a correct filterset, a slow-multi-gib test group (max-threads = 1) so they never overlap — each keeps several GiB live — and a 20-minute budget instead of 150s.

Make the remaining run cheaper

• Build fsst-rs optimized under the ci profile. Its compression kernel dominated fsst_compress_offsets_overflow_i32 unoptimized: 79.7s → 7.9s on arm64, no longer even flagged SLOW.

Stop depending on scale for coverage

• codes_offsets_i64_path_roundtrips drives the generic compress_views with the i64 offset type directly, on three short strings (10ms), so that branch is covered without the ~3.2 GiB test. That size is not padding — FSST escape coding is exactly 2×, so crossing i32::MAX needs ~1.07 GiB of input and the existing test sits within 1% of the minimum.
• alp_patches_are_chunk_indexed replaces what compress_large_int was reaching by accident. Dumping the encoding tree by size shows the 50M-row array is the point where the sampled ALP exponents stop fitting and the result gains patch_indices / patch_values / patch_chunk_offsets; below ~5M rows there are no patches at all. Sprinkling values the sampled exponents can't represent reproduces that structure at 200K rows, and varying the density pins the patch_chunk_offsets width across u8, u16 and u32 — 1.3s for three cases that assert the shape and round-trip the values, against 40s for one that asserted neither.

Local repro for a slow test:

cargo test --release -p vortex-fsst fsst_compress_offsets -- --ignored

Checks

• yamllint --strict -c .yamllint.yaml on both changed workflows — clean.
• cargo clippy -p vortex-array -p vortex-btrblocks -p vortex-fsst --all-targets --all-features — clean.
• CI on this head: all test jobs green; the musl job's slow step reports 3 tests run: 3 passed in 26.2s with no rebuild.
• Could not run cargo +nightly fmt (no nightly toolchain in this environment); stable cargo fmt --check reports no diff in any touched file.

What APIs are changed? Are there any user-facing changes?

None. CI configuration and test gating only.

[codspeed-hq]

Merging this PR will degrade performance by 0.84%

⚠️ 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.

⚡ 1 improved benchmark
❌ 1 regressed benchmark
✅ 1954 untouched benchmarks
⏩ 442 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	WallTime	words_gather_scalar[65536]	8.2 µs	9.4 µs	-12.26%
⚡	Simulation	compact[(2048, 90)]	1.8 µs	1.6 µs	+12.07%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing claude/slow-ci-tests-coverage-d3dq68 (5fa8c2a) with develop (4080251)}

Footnotes

1. 442 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩