Development records and generated results

User-facing documentation, small regression fixtures, reproduction scripts and curated guide figures stay in Git. Development logs, detailed qualification reports, raw measurements and historical plot collections live outside the checkout. Building these docs and running ordinary tests does not require the development archive.

Local archive

The default layout is a sibling of the source repository:

dev/
  warpSPHCore/
  warpSPHDevelopment/
    README.md
    core/
      README.md
      REPOSITORY_PLAN.md       # detailed plan at the time of migration
      docs/development/        # reports and generated figures
      docs/regression/         # raw results and workload fixtures
      experiments/            # experiment results and reports
      higherOrderSPH/          # generated harness reports
      logs/2026-10-16/          # captured development runs
      manifests/              # migration inventory and SHA-256 checksums

Other projects can use separate children such as plotting/ under the same warpSPHDevelopment root. Backup of that shared folder is manual; no cloud upload, synchronization or retention service is configured by this repository.

scripts/development_paths.py resolves the core artifact directory. To use a different location, set the shared root (not its core child):

export WARPSPH_DEVELOPMENT_ROOT=/path/to/warpSPHDevelopment

This is also useful for worktrees whose immediate parent is not dev/. The helper only resolves paths. Producers create directories when writing; imports and --help do not create an archive.

Regeneration

Run commands from the source checkout, with the appropriate package/runtime installed. Set output paths explicitly for tools requiring them:

export WARPSPH_DEVELOPMENT_ROOT="$(cd .. && pwd)/warpSPHDevelopment"
CORE_DEV="$WARPSPH_DEVELOPMENT_ROOT/core"

python scripts/diagnostics/qualify_mfm.py --suite wave --n 32 --device cpu \
  --precision float64 --backend warp --modes MFM MFV --closures project \
  --tend 0.001 --output "$CORE_DEV/docs/regression/mfm-short-new.json"

python scripts/diagnostics/qualify_reconstruction.py --help
python scripts/diagnostics/qualify_mfm_geometry.py --help
python scripts/benchmarks/plot_work_scaling.py --help

Development plotting/frontend tools with default result paths now use the external core directory. Explicit CLI output arguments remain caller-owned. Historical commands in archived reports preserve the paths used for their original runs; replace those output paths with the external location when rerunning them. Source hashes, configurations and numerical assumptions remain in the original records. Exact historical numbers require compatible sources and dependencies; regeneration on a newer revision creates new evidence. Source-hash guards are retained rather than silently relabeling old results.

The field reconstruction guide is a curated documentation artifact. Its small figures and example dataset intentionally remain in docs/docs/assets/reconstruction/, and its regeneration command continues to update those public assets.

Core capability qualification

The bounded core pass is recorded by source commit 2034a4e. Its final local checks passed:

Gate	Result
Full CPU suite	2,568 passed, 2,429 skipped; includes 44 isolated float64 derivative scripts
Real-CUDA gate, float32	4,143 passed, 163 skipped
Real-CUDA gate, float64	4,163 passed, 143 skipped
Targeted operation matrices	All five passed, including 1D, 2D/both precisions/light jitter and 3D CUDA
Public examples	Fourteen configurations per precision on CPU/CUDA
Isolated wheel	Same public examples passed on both devices and precisions
Documentation	Production build passed

The runtime was Python 3.13.14, Torch 2.13.0+cu130, Warp 1.17.0 and NumPy 2.5.2, with an RTX PRO 6000 Blackwell for CUDA. Python >=3.10 is the package metadata range; hosted CPU CI is configured for Python 3.12. Those declarations do not certify every dependency/OS/GPU combination. Skips and exclusions are described in the archived reports. Hosted CUDA CI still requires GPU infrastructure.

Detailed evidence is in core/docs/development/delivery-validation.md and core/docs/regression/capability-delivery-2026-10-16.json inside the shared development folder. Historical reconstruction/MFM reports and raw ladders are there too. The public reconstruction, MFM/MFV, autodiff and linalg pages retain their supported contracts and explicit limitations. The runnable examples are independent of the archive.