bench: recapture the laptop-test pipeline on the CSC DE route

[nick-youngblut]

Summary

pipeline_ooc_constrained's numbers and floors were captured on pyscx 0.18.0, before the benchmark fixtures carried CSC sidecars. Its DE stage now takes cpu_csc_nnz instead of cpu_csr, so the pipeline's published numbers and floors described a path that no longer runs. This PR re-captures the benchmark, recalibrates its floors, and rewrites the doc section.

Changes

• Recapture (13 raw results).

  • Covers the 8 pyscx cells and the 5 scanpy cells that complete, run one SLURM job at a time so every SCX-vs-scanpy ratio comes from a single capture.
  • The 3 scanpy OOM cells keep their results, since nothing in them depends on pyscx.

  dataset	pyscx 16 GB, before → after	DE stage	scanpy 32 GB
  pbmc10k	61 s → 66 s	46 → 49 s (cpu_csr, no sidecar)	58 s
  tabula_sapiens_100k	267 s / 4,259 MB → 113 s / 2,587 MB	179 → 19 s	247 s / 5,005 MB
  census_500k	1,494 s / 6,336 MB → 548 s / 3,496 MB	992 → 54 s	2,547 s / 17,828 MB
  census_1m	3,186 s / 8,819 MB → 1,227 s / 4,332 MB	2,022 → 88 s	OOM

• Floors (thresholds.yaml).

  • total_wall_s and pipeline_peak_rss_mb are recalibrated at 1.4–1.5× the new medians. The old census_1m wall bound (5,500 s) would have passed a full fallback to cpu_csr.
  • New stage_wall_s__rank_genes_groups floors at census_500k (85 s) and census_1m (140 s) cover a fallback from the exact-nnz kernel to the densify kernel. de_route_is_csc reads 1.0 on both kernels, and the pipeline total absorbs that fallback within node noise.
  • Checked against saved densify results (DE 114 s / 228 s): they fail the new floors while passing total_wall_s.

• Docs.

  • docs/performance.md, § The laptop test: new table and stage breakdown. DE is now 7 % of the census_1m run instead of 63 %.
  • The "where SCX loses" paragraph is scoped to files without a sidecar: pbmc10k is below the auto rule's 50,000 cells and still loses DE about 6×.
  • Source and manifest notes are updated.
  • docs/sharding.md gains the nnz figure beside the densify one.

Notes

• The other three community benchmarks are not re-run.
  • Their paths do not read a sidecar, and check runs at census_1m reproduce the baseline's peak RSS (QC 2,197 vs 2,196 MB; score_genes 805 vs 821 MB), still bit-exact against scanpy.
  • The multimodal atlases carry no sidecar.
• The raw JSONs record library_versions.pyscx = 0.18.0. That is the distribution metadata scx-bench has installed. Each job asserted the repository's release build by path and build profile before timing anything.
• No baseline is promoted. v0.18.0-community-benchmarks still holds the pre-sidecar pipeline rows, and LATEST is untouched.

🤖 Generated with Claude Code

https://claude.ai/code/session_012LmrECbnnqEEHQtURZRn84

[gemini-code-assist]

Code Review

This pull request updates the benchmark results and documentation to reflect the performance improvements from the new cpu_csc_nnz kernel and the use of CSC sidecars. It includes updates to several JSON benchmark files, the thresholds.yaml configuration, and the performance.md and sharding.md documentation. The reviewer correctly identified that the JSON field name should be de_route rather than route in the documentation.

[gemini-code-assist]

For accuracy, the recorded field in the JSON results is de_route, not route. It would be clearer to refer to it as de_route="cpu_csc_nnz".