Add AGENTS paper registry and backend-selectable estimation path (CPU/CUDA/MLX/OpenCL)

[Copilot]

This PR addresses two requests: persist the AGENTS.md paper set in-repo, and extend the current estimation workflow with selectable GPU-capable backends while preserving the existing simple-structure MLS2PLM contract. It introduces backend-aware execution wiring for supported models/optimizers and CLI-level device selection.

• Paper registry (AGENTS.md traceability)

  • Added docs/agents_papers.md with the full AGENTS.md reference list and DOI links.
  • Provides a stable in-repo source for formula/diagnostic grounding work.

• Backend selection in fit configuration

  • Added compute_backend to FitConfig with allowed values: cpu, cuda, mlx, opencl.
  • Added backend normalization and availability checks in python/fast_mlsirm/backend.py:
    • CUDA: validates cupy
    • MLX: validates mlx.core
    • OpenCL: validates pyopencl and platform presence

• Backend-aware estimation/objective plumbing

  • Threaded backend selection through fit() and objective evaluation (neg_loglik_and_grad, linear_predictor paths).
  • Kept model algebra/parameterization unchanged; backend selection changes execution path, not formula semantics.
  • FitResult now carries compute_backend; fit summary serialization includes it.

• CLI/API surface updates

  • Added --device {cpu,cuda,mlx,opencl} to:
    • fast-mlsirm fit
    • fast-mlsirm diagnose-dimensions
  • CLI JSON output now reports selected backend.

• Dependency metadata and coverage updates

  • Added optional gpu extras in pyproject.toml for CUDA/MLX/OpenCL ecosystems.
  • Updated tests for backend config validation and CLI output fields.

fast-mlsirm fit \
  --responses runs/sim_001/responses.npy \
  --factors runs/sim_001/item_factor.csv \
  --model MLS2PLM \
  --optimizer adam_lbfgs \
  --device cuda \
  --out runs/fit_cuda

[Copilot]

Pull request overview

This PR adds an in-repo paper registry for traceability and introduces a backend-selectable estimation path (CPU/CUDA/MLX/OpenCL) by threading a compute_backend choice through configuration, CLI, fit execution, and result serialization.

Changes:

• Added docs/agents_papers.md as an in-repo registry of the AGENTS.md paper list.
• Introduced FitConfig.compute_backend plus CLI --device {cpu,cuda,mlx,opencl}, and persisted the chosen backend into FitResult and JSON outputs.
• Implemented backend-aware objective/linear predictor code paths and backend availability checks (new python/fast_mlsirm/backend.py), with tests updated accordingly.

Reviewed changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/test_fit_pipeline.py	Asserts FitResult.compute_backend is reported (CPU in smoke test).
tests/test_config.py	Adds validation test for invalid compute_backend values.
tests/test_cli.py	Exercises --device and asserts JSON payload includes compute_backend.
README.md	Documents backend selection and includes --device in example command.
python/fast_mlsirm/types.py	Extends FitResult to carry compute_backend.
python/fast_mlsirm/objective.py	Adds backend-aware execution paths for linear_predictor and neg_loglik_and_grad.
python/fast_mlsirm/io.py	Persists compute_backend into fit_summary.json.
python/fast_mlsirm/fit.py	Validates backend availability once per fit and threads selection through optimization.
python/fast_mlsirm/config.py	Adds compute_backend to FitConfig and validates allowed values.
python/fast_mlsirm/cli.py	Adds --device flag to relevant commands and reports backend in JSON output.
python/fast_mlsirm/backend.py	New module for backend normalization and availability checks.
pyproject.toml	Adds optional gpu extras for CUDA/MLX/OpenCL dependencies.
docs/agents_papers.md	New paper registry document.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[seonghobae]

Superseded by #109. Per maintainer direction, GPU support is implemented as GPGPU inside the Rust core (wgpu, MIT/Apache-2.0) exposed as a device sub-option of the existing rust backend (rust_device = cpu|gpu|auto), keeping the single backend axis {numpy, rust, auto}. This avoids the competing Python compute_backend axis and the cupy/mlx reimplementation of the objective, which collided with the backend abstraction in backend.py/config.py/types.py/objective.py/fit.py. The GPGPU path has a runtime CPU fallback (no GPU required in CI) and numerical-parity tests against the NumPy reference.

[seonghobae]

Closing as superseded and stale.

GPU/backend half → superseded by #109. Per maintainer direction (see the note already on this PR), GPU support is being implemented as GPGPU inside the Rust core (wgpu, MIT/Apache-2.0) exposed as a device sub-option of the existing rust backend, keeping the single backend axis {numpy, rust, auto}. Since this PR was opened, main independently grew its own python/fast_mlsirm/backend.py defining normalize_backend/resolve_backend for {numpy, rust, auto} — this PR's backend.py is a new file redefining the same names with a conflicting {cpu, cuda, mlx, opencl} meaning, and its compute_backend axis collides with that abstraction across config.py/types.py/objective.py/fit.py (hence the DIRTY/CONFLICTING state). The added backend.py only validates cupy/mlx/pyopencl availability; it implements no actual GPU compute ("changes execution path, not formula semantics"), so there is no unique working GPU code to salvage — #109's GPGPU-in-Rust path (with CPU fallback + numerical-parity tests) replaces it cleanly.

Paper-registry half → already in-repo. The full AGENTS.md reference list (Kang & Jeon 2025, Jeon et al. 2021, Molenaar & Jeon 2026, etc., with DOIs) already lives in AGENTS.md on main, which is the stable in-repo grounding source. The separate docs/agents_papers.md is redundant, and its AGENTS.md link used an absolute CI-runner path (/home/runner/...) flagged in review.

No unique un-merged value remains. Tracking GPU work in #109.