Chm02 CUDA: mainline correctness fixes, tests, and timing surface

[tpn]

Closes #79

Summary

This PR mainlines the validated Chm02 CUDA bring-up work as a focused correctness-first integration slice.

It promotes the legacy Chm02 CUDA path from a CPU-assisted bring-up flow toward a first-class correctness path for single-graph runs by moving the major solve phases onto the GPU, fixing Linux compatibility issues, adding focused regression coverage, and exposing explicit CUDA timing fields.

Included

• Graph.cu / GraphCu.c fixes for CUDA add-keys, peel/order capture, assignment, and verify.
• Linux Chm02Compat / file-work fixes needed for no-file-io and file-io parity.
• Focused regression coverage for CUDA Chm02:
  • Hologram known-seed no-file-io
  • Hologram known-seed file-io
  • generated non-Assigned16 case
  • timing-field presence
• Explicit per-phase CUDA timing fields in CSV output.
• docs/chm02-cuda-mainline.md describing scope, non-goals, compatibility notes, and staged acceptance.

Non-goals

• The standalone batched GPU peeling POC.
• Large-scale GPU constructor/performance work.
• Throughput-optimized Chm02 CUDA kernels.

Validation

Ran:

• cmake --build build-cuda -j2
• ctest --test-dir build-cuda --output-on-failure -R 'perfecthash\.cuda\.chm02'

Current focused CUDA Chm02 test result:

• 4/4 passed

Notes

This remains a correctness-first CUDA path. The serial CUDA kernels are intentionally bring-up oriented and should not be read as a throughput claim.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 302f13b036

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".