[michel2323]

Refs #91

Task: CI — select the test element types per run (issue #91, owner decision: fewer element types per CI run, no CPU-specific kernel specialisation tricks). Tests and CI configuration only, no library code.

What was built:

• test/backends.jl: SDS_TEST_ELTYPES (comma-separated, default all four, at least one real type) read once into SELECTED_ELTYPES (selected_eltypes), shown in the backend banner.
• test/utils.jl: ELTYPES/REAL_ELTYPES/COMPLEX_ELTYPES derive from it; eltypes_among(types) / eltypes_among(Complex) for testsets that run for some types only (notes a skip when the selection has none of them). backends.jl is now included before utils.jl.
• test/runtests.jl: SPLIT_FILES split over the selected types only; RUN_SHARED goes to the Float64 part, or the first selected type's when Float64 is left out.
• Audit: the two complex-only testsets (test_numeric_ldlt, test_reference_ldlt) and the (Float64, ComplexF32) FGMRES batch testset use eltypes_among; no testset indexes ELTYPES positionally; the RUN_SHARED testsets that pin Float64 (refinement to 1e-14, Schur/FGMRES edge cases) are precision-specific and stay Float64. No assertion changed. test_helpers tests the parser.
• .github/workflows/ci.yml: workflow-level SDS_TEST_ELTYPES = Float64,ComplexF32 on pull_request/push, all four on the weekly schedule, and a new workflow_dispatch input (default all four). Job names unchanged; the CUDA job keeps SDS_TEST_CPU=0, SDS_TEST_SKIP=test_aqua.
• AGENTS.md: test-conventions bullet, the ParallelTestRunner paragraph and the Commands block; README.md: SDS_TEST_ELTYPES in "Running the tests" (review round 1).

Tests: Both selections pass in full, with no failures or broken tests. Locally (CPU backend, SDS_TEST_GPU=0, 4 cores, 4 workers, Julia 1.13.1): all four types 71462/71462, SDS_TEST_ELTYPES=Float64,ComplexF32 41076/41076. CI on cc98eaf with Float64,ComplexF32: CPU 41076/41076, CUDA 44215/44215.

Wall time per job (before: run 37693884475 of #100, all four types; after: this PR on cc98eaf, Float64+ComplexF32):

job	before	after	summed part time before → after
CUDA (test-gpu, 12 workers), test step	18m20s (job 19 min)	19m03s (job 20 min)	12428 → 7370 s (−41%)
CPU (test-github-cpuonly, kkt, 32 workers), test step	21m40s (job 22 min)	20m59s (job 22 min)	32532 → 20707 s (−36%)
CPU local, 4 workers	21m14s	11m53s	4858 → 2702 s (−44%)

The CI jobs do not get shorter. With 12 or 32 workers, the slowest single part sets the wall time, and the slowest part is a complex part kept in the selection. On CUDA that is test_numeric_ldlt[ComplexF32] (1132 s now; before, [ComplexF64] 1089 s and [ComplexF32] 1014 s). On kkt every complex part takes about 1100 s, whichever complex type it is. The selection cuts total work (and runner occupancy) by about 40%. It only shortens the wall time where the run is throughput-bound, i.e. locally and on small runners. To shorten the CI jobs, the heavy files have to be split below the element type: test_numeric_ldlt, test_solve, test_numeric_cholesky_c and test_numeric_cholesky_a on CUDA. That is not done here.

Per file, seconds summed over the parts (each worker's cold compilation is charged to the first file it runs, so single rows are noisy)

file	CPU local, 4 types (s)	CPU local, 2 types (s)	CUDA CI, 4 types (s)	CUDA CI, 2 types (s)
test_api	147	17	236	201
test_bench_compare	6	6	6	6
test_bench_smoke	6	5	5	11
test_dense	179	96	311	182
test_fgmres	195	37	280	91
test_helpers	13	7	18	11
test_matching	270	134	281	142
test_matrix	6	4	10	9
test_numeric_cholesky_a	336	198	696	787
test_numeric_cholesky_b	488	248	258	58
test_numeric_cholesky_c	121	115	982	965
test_numeric_ldlt	256	132	3514	1897
test_numeric_lu	77	89	617	400
test_options	6	6	6	5
test_ported	1226	622	748	430
test_reference_cholesky	12	16	9	25
test_reference_ldlt	38	22	94	63
test_refinement	59	93	653	339
test_schur	625	488	867	360
test_solve	616	297	2509	1242
test_symbolic_etree	8	6	8	9
test_symbolic_pairs	15	20	62	12
test_symbolic_schedule	13	15	20	20
test_symbolic_supernodes	3	3	6	5
test_ubatch	137	26	232	100
sum	4858	2702	12428	7370

Deviations from PLAN.md / TASKS.md: none (PLAN.md untouched).

Follow-up issues opened: none. #91 stays open for lever 3 (folding the small Val flags into run-time branches, T25 work).

This PR changes ci.yml, which was expected to make the Claude reviewer skip it; the reviewer ran anyway (VERDICT: APPROVE on 2a6bf54). The owner merges it by hand once CI is green.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AvwMxoGmJm3eSmpzxAy72A

---

Generated by Claude Code

[claude]

VERDICT: APPROVE (head cc98eaf)

Infrastructure PR for issue #91 (not a TASKS.md task, so no Report block; the PR body serves as the report). Reviewed the full diff against AGENTS.md, the issue and the test conventions.

Checked

• Scope: tests, test helpers, ci.yml and AGENTS.md only; no library code, PLAN.md and TASKS.md untouched, no Manifest.toml, no public names involved.
• test/backends.jl: selected_eltypes parses the comma list in ALL_ELTYPES order, rejects unknown names, duplicates collapse, and a selection without a real type is an error (needed by the shared testsets). SELECTED_ELTYPES is read once per process and shown in the banner.
• test/utils.jl: ELTYPES/REAL_ELTYPES/COMPLEX_ELTYPES derive from the selection; eltypes_among(types) and eltypes_among(S::Type) filter the running part and log the skip only from the part that owns the shared testsets. Include order backends.jl → utils.jl → matrices.jl is sound: backends.jl has no top-level use of the other helpers, and utils.jl needs SELECTED_ELTYPES.
• test/runtests.jl: the banner module now runs before the suite is built so a bad SDS_TEST_ELTYPES fails early; SPLIT_FILES split over the selected types only; SHARED_ELTYPE falls back to the first selected type when Float64 is left out; test names unchanged (SDS_TEST_ONLY/SKIP prefix matching and the runner history still work). Workers read the same environment as the main process, as SDS_TEST_CPU already does.
• ci.yml: workflow-level env with inputs is valid; pull_request/push get Float64,ComplexF32, schedule an empty string (all four), workflow_dispatch its input (default all four). Job ids, check names and the GPU job's SDS_TEST_CPU=0 SDS_TEST_SKIP=test_aqua unchanged.
• Tests: no assertion weakened; the three testsets moved to eltypes_among are equivalent to the old filters; test_helpers covers the parser (@test_logs match_mode = :any with no pattern passes with or without the skip note and returns the value). No ELTYPES[k] or first(ELTYPES) anywhere in test/.
• CI: all checks pending at review time (the red fix job is the implementer workflow, not a test). I could not run Julia in this session, so the behaviour under a non-default selection rests on reading plus the pending CI.

Blocking findings: none.

Non-blocking (one inline comment on AGENTS.md:195):

• test_numeric_cholesky_a.jl:63-69 still compiles a Float32 regime-A Cholesky in the Float64 part regardless of the selection; test_symbolic_schedule.jl:340 hard-codes (Float32, ComplexF64) (host-only, cheap, but the new "never" in AGENTS.md should be qualified to numeric testsets).
• README.md "Running the tests" does not list SDS_TEST_ELTYPES.
• The PR body's Tests section says "the table follows"; please fill in the local CPU run with both selections (and the CUDA numbers once CI is green) before the owner merges, so the body matches what was verified.

[claude]

VERDICT: APPROVE (head 2a6bf54)

Re-review after the push that answered the previous round. Infrastructure PR for issue #91 (no TASKS.md task, so the PR body stands in for the Report). The head differs from the approved cc98eaf only in AGENTS.md (the "numeric testset … never over a hard-coded list" wording, with the host-only symbolic elsize checks exempted) and one README.md line adding SDS_TEST_ELTYPES to the "Running the tests" block, both as the owner said on the earlier thread. The Float32 factorization in the shared regime-A elsize check stays by the owner's decision.

Checked on this head

• Scope: test/, ci.yml, AGENTS.md, README.md only; no library code, PLAN.md, TASKS.md and Manifest.toml untouched, no public names involved.
• test/backends.jl: selected_eltypes keeps ALL_ELTYPES order, collapses duplicates and blanks, rejects unknown names and a selection with no real type; SELECTED_ELTYPES is read once per process and printed in the banner. No top-level dependency on utils.jl, so the new include order (backends.jl → utils.jl → matrices.jl) is sound; nothing outside test/runtests.jl includes utils.jl or backends.jl.
• test/utils.jl: ELTYPES/REAL_ELTYPES/COMPLEX_ELTYPES derive from the selection or the part; eltypes_among filters the running part and logs the skip only from the part that owns the shared testsets.
• test/runtests.jl: the banner module runs before the suite is built, so a bad SDS_TEST_ELTYPES fails early; split files run over the selected types only; SHARED_ELTYPE falls back to the first selected type (always real, given the parser) when Float64 is left out; test names and SDS_TEST_ONLY/SKIP matching unchanged.
• Audit: test/ has no ELTYPES[k]; the three testsets moved to eltypes_among are equivalent to the old filters, no assertion changed. test_helpers covers the parser and the helpers (@test_logs match_mode = :any with no pattern passes regardless of the skip note and returns the value).
• ci.yml: workflow-level env gives pull_request/push Float64,ComplexF32, schedule the empty string (all four), workflow_dispatch its input; job ids, check names and the GPU job's SDS_TEST_CPU=0 SDS_TEST_SKIP=test_aqua unchanged. The workflow parsed and queued on this head.
• CI: build green; Run tests CPU and CUDA jobs pending at review time. I could not run Julia in this session; the parser and @test_logs behaviour rest on reading.

Blocking findings: none.

Non-blocking: the PR body's Tests section still says "the table follows". Since the owner merges this by hand once CI is green, please fill in the local run with both selections and the CI numbers before merging so the body records what was verified.