perf(inference): add latency benchmarking, edge caps, and INT8 quantization

[ayush-kr-repo]

Closes #3676

Summary

The README advertises a <200ms p99 budget (89ms for HTGNN) but nothing measures it, and no optimization is applied to the served model. This PR adds the measurement harness plus two opt-in optimizations, and reports what they actually do.

Changes

• New src/inference/optimization.py
  • benchmark_latency → LatencyReport (p50/p95/p99/mean/min/max, meets_budget()); warm-up runs discarded; injectable timer so tests are deterministic
  • percentile — nearest-rank in integer arithmetic (floats would report the max as p99, see docstring)
  • cap_subgraph_edges — per-transaction edge budget; node set intact so no re-indexing; input never mutated
  • quantize_model_dynamic — INT8 dynamic quantization of Linear layers, degrading to the original model on failure
• ProductionRiskScorer — opt-in max_subgraph_edges (applied before tensor work) and quantize; both off by default

Measured results

Edge cap, driven through the real scorer with a 20k-edge hub account:

	p50	p99
baseline	0.53ms	0.74ms
capped (500 edges)	0.26ms	0.41ms

~45% off p99.

INT8 quantization, 3-layer MLP (p99) — included because the result is not the expected one:

width	fp32	int8	change
64	0.09ms	0.44ms	374% slower
256	1.37ms	0.97ms	29% faster
512	4.17ms	1.54ms	63% faster

Quantization overhead is fixed per call while savings scale with matmul size, so narrow layers regress. Quantized Linear also requires rank ≥ 2 inputs — a model feeding 1-D tensors works in FP32 and raises once quantized. Both are documented in the docstring; quantization is off by default and ships beside the benchmark harness so it can be verified per model.

Note: the actual HTGAT was not benchmarked as no checkpoint is committed to the repo, so this PR provides the harness and the size threshold rather than claiming a speedup for the served model.

Testing

• 34 new tests in tests/test_inference_optimization.py (RUN_TORCH_TESTS-gated): exact percentiles from a scripted fake clock, nearest-rank correctness (p99 of 1..100 is 99, not 100), warm-up excluded from measurement, budget checks, quantization output-shape/precision/fallback, and edge-cap invariants (aligned tensor truncation, edge_index within node range, no mutation, under-budget and None passthrough)
• Regression: test_production_scorer.py, test_mc_dropout_uncertainty.py, test_performance.py — 76 tests total passing; flake8 critical check clean

[github-actions]

🌟 Welcome to AegisGraph-Sentinel-2.0!

Thank you for submitting your pull request, @ayush-kr-repo! 🙌 We truly appreciate your time, effort, and dedication to improving our platform.

---

📌 Next Steps & Guidelines:

• Review Process: Our Project administrators will review your changes as soon as possible.
• Feedback: If there are any specific modifications, edge cases, or code improvements needed, we will provide detailed feedback directly in this pull request thread.
• Continuous Integration: Ensure that all automated pipeline checks and status workflows pass successfully.

---

Thank you once again for your valuable contribution to Puneet04-tech/AegisGraph-Sentinel-2.0! Let's build something amazing together. Happy coding! 🚀

[github-actions]

🎉 Contribution Merged!

Thank you so much, @ayush-kr-repo, for your contribution! Your pull request #3677 has been successfully merged.

We truly appreciate your time and effort in making this project better for the community. Keep up the amazing work! 🚀✨