rc/0.2.0: GPU device-efficiency profiling + fraction-to-boundary vectorization

[wahln]

Release candidate for 0.2.0. No version bump yet — opening for PR review first; the bump + changelog will follow once the review is settled.

Summary

A measure-first GPU optimization pass, validated on a local RTX 4070 (CuPy), plus the device-efficiency profiling harness it required.

Profiling harness (the deferred device_efficiency deliverable)

• Real device detection in capabilities() via the Array-API inspection API (__array_namespace_info__().devices()) — was hardcoded ("cpu",). Result.device added and shown in the tier-1 summary.
• benchmarks/harness: DeviceMetrics + measure_device_solve, with a host-sync counter (patches the array type's scalar dunders) and the CuPy GPU/CPU time split.
• benchmarks/runners/device_efficiency.py (GPU-gated → no-op on CI) and the previously-stubbed micro-benchmarks (matvec / dense solve / Newton step).

The optimization

• Finding: fraction_to_boundary looped over the array in Python calling float(v[idx]) per element — O(n) host↔device syncs, called 6× per iteration. On GPU this dominated: 5,510 syncs/iter at n=1000, scaling linearly to 27,422 at n=5000.
• Fix: vectorized to a single reduction (Wächter–Biegler 2006 eq. 15), Array-API pure (no concrete backend in core).
• Result on the 4070: krylov syncs/iter 27,422 → 133 (no longer scales with n); wall 9.04s → 0.39s (~23×) at n=5000; identical iteration counts and KKT error. The loop is now ~97% GPU-compute-bound (gpu_time ≈ wall), so further sync consolidation was deliberately not pursued (data showed <3% headroom).

Tests / gates

• Device-efficiency harness smoke test, an O(1)-sync regression guard for fraction_to_boundary, and a vectorization-equivalence test.
• All gates green: 799 passed, ruff/format/mypy/import-purity clean.

Notes for review

• Invariants held: core change uses only array_namespace/xp.*; backend-specific timers live in benchmarks/.
• torch/jax CUDA wheels are a separate install step; the harness is already backend-parametric for when they're added.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR prepares the 0.2.0 release candidate by adding a GPU/device-efficiency profiling harness under benchmarks/, surfacing device information in solver results/logging, and removing an O(n) host↔device sync bottleneck in the interior-point loop by vectorizing fraction_to_boundary.

Changes:

• Vectorize fraction_to_boundary to eliminate per-element scalar materializations (dramatically reducing GPU syncs/iter) and add unit + regression coverage.
• Add device reporting: detect available devices in backend capabilities and record the solution device on Result, including tier-1 logging output.
• Introduce the device-efficiency study harness/runner and revive micro-benchmarks for kernel-level timing across installed backends.

Reviewed changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
ipax/ipm/barrier.py	Vectorized fraction_to_boundary to avoid per-element host syncs.
tests/unit/test_barrier.py	Adds empty-vector and reference-equivalence tests for fraction_to_boundary.
tests/regression/test_fraction_to_boundary_no_elementwise_sync.py	Adds regression guard that scalar syncs during fraction_to_boundary stay O(1).
benchmarks/harness/__init__.py	Adds device-efficiency measurement primitives (sync counter, GPU timing/memory, report formatting).
benchmarks/runners/device_efficiency.py	Adds CLI runner to generate JSON/Markdown device-efficiency reports over backend/route/size grids.
tests/integration/test_device_efficiency.py	Smoke-tests the harness + runner on CPU backends and validates sync-counter restoration behavior.
ipax/backend/namespace.py	Implements device discovery via __array_namespace_info__().devices() and exposes it in Capabilities.
ipax/result.py	Adds Result.device field for reporting where the solve ran.
ipax/solve.py	Populates Result.device from the solution array’s .device attribute.
ipax/_logging.py	Includes device in the tier-1 format_result() output.
tests/unit/test_logging.py	Updates logging expectations to include the new device line.
tests/unit/test_backend_namespace.py	Adds tests for real device reporting and fallback behavior in capabilities().
benchmarks/runners/micro/bench_kernels.py	Implements micro-benchmarks for matvec/dense solve/Newton-step kernels across installed backends.

---

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

[codecov]

Welcome to Codecov 🎉

Once you merge this PR into your default branch, you're all set! Codecov will compare coverage reports and display results in all future pull requests.

ℹ️ You can also turn on project coverage checks and project coverage reporting on Pull Request comment

Thanks for integrating Codecov - We've got you covered ☂️