Revalidate the current ONNX model on Qualcomm QNN

[triasha72]

Summary

• rerun the repository's current ONNX model through Qualcomm AI Hub on Snapdragon 8 Elite QRD
• compile batch 1/32/256 graphs, link one QNN context binary, and validate held-out inference
• update the tracked QNN report and portfolio matrix with matching model provenance
• add a resumable authenticated rerun script and reproducibility documentation

Physical-device results

Batch	Latency	Throughput	Peak memory	Placement
1	0.038 ms	26,315.8 samples/s	122,855,424 B	NPU × 9
32	0.040 ms	800,000.0 samples/s	122,880,000 B	NPU × 9
256	0.047 ms	5,446,808.5 samples/s	122,896,384 B	NPU × 9

Source ONNX SHA-256: 191927e05b5f82a539f0ad35c78dafe7f969cb4e57e9c76556cb9f35053e658e

Compile jobs: jpyxnmd05, jp0jk610g, jp8x813qg
Link job: jgk4d8lvp
Profile jobs: j567vw3np, jpyxnmvr5, jglx7lllg
Inference jobs: jg9m8d3q5, jpez2yl7p, jgk4d82op

Verification

• ruff format --check .
• ruff check .
• 26 Qualcomm/QNN and acceptance tests passed
• regenerated acceptance outputs match tracked files byte-for-byte

The full local pytest collection was blocked by a stale macOS scikit-learn binary in the existing virtual environment; GitHub's clean Linux CI remains the authoritative full-suite run.

Power remains unmeasured, and AI Hub model-profile latency is not presented as Android application end-to-end latency.