[sathariels]

Summary

Opens the Phase 2 timing-chart slice: a reproducible C++ native vs Python reference sample_neighbors benchmark with measured stored evidence.

• python/sagerec_sampler_benchmark.py builds a deterministic synthetic CSR, verifies ADR-005 without-replacement parity, then times graph_sampler.BipartiteCSR.sample_neighbors against sagerec_reference_sampler with warm-up and multiple measured repetitions.
• Headline latency, throughput, and speedup use the median of measured repetitions (mean/min/max are also recorded). Speedup is reference_median_seconds / native_median_seconds from measured seconds only.
• Provenance includes graph size, node-query workload, k, replacement policy, seed, build type, compiler, Python, CPU, and the timing statistic.
• scripts/run_sampler_timing.py writes:
  • results/sampler_timing.json
  • results/sampler_timing.md
  • results/sampler_timing.svg
• Default source is synthetic so CI never downloads MovieLens. --source movielens-100k uses existing train-only processed/ artifacts and never downloads.
• Unittests cover parity-before-timing (timer is not called on mismatch) and the required JSON schema. CTest python_binding_smoke discovers the new tests.

Docs no longer mark Phase 2 timing charts as closed.

Does not invent ML quality metrics, retcon MF/GraphSAGE JSON, add MovieLens 1M, GCN, PyG NeighborLoader, multi-seed leaderboards, or Rust.

Measured locally (Release, this agent environment)

Copied from the committed JSON (synthetic 512×1024 graph, degree 32, k=10, 4000 queries, seed 7, 3 warm-up + 9 measured reps, GNU 13.3.0, Python 3.12.3):

Sampler	Median seconds	Median queries/s
native graph_sampler	0.003911	1,022,638
Python reference	0.434863	9,198

Median speedup (reference / native): 111.177×. One-machine sampler evidence, not a production latency or SOTA claim.

How to verify

cmake -S cpp -B build -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_COMPILER=g++
cmake --build build --config Release
ctest --test-dir build --output-on-failure --build-config Release
PYTHONPATH=build:python python3 -m unittest python/tests/test_sampler_benchmark.py -v
PYTHONPATH=build:python python3 scripts/run_sampler_timing.py

Locally: Release CTest 17/17 passed, including python_binding_smoke. Default CI must stay on the synthetic path.