Add reproducible dual-queue DQA/BP research pipeline

[Archi-Hogler]

Summary

• Add the dual-queue runtime, dynamic edge capacity/task arrival/channel state support, queue-aware reward tooling, smoke controllers, and analysis coverage.
• Add DQA-D3QN and BP-DQRG-D3QN, including the registered causal-ablation matrix, strict analyzers, diagnostics, and source-provenance hardening.
• Package the previously local-only runtime/test dependencies required to reproduce the branch from Git alone.
• Pin the validated environment: Python 3.9, NumPy 1.21.6, Matplotlib 3.2.2, TensorFlow 2.8.0, and protobuf 3.20.1.
• Keep generated results, outputs, models, snapshots, and research data out of the PR.

Scientific outcome

The fixed 3-seed x 3-mode causal ablation is UNSTABLE under the preregistered rules:

• Historical-information Drop improvement: mean 0.439321%, strict wins 2/3 -> Gate FAIL.
• Adaptive-gate Drop improvement: mean 0.478471%, strict wins 2/3 -> Gate FAIL.
• Positive mean effects with inconsistent seed direction do not support significance or universality claims.

Verification

• Clean-tree Python compilation: passed.
• Ordinary environment: 88 tests, 86 passed, 2 permitted TensorFlow-only skips.
• TensorFlow 2.8 environment: full tests/ discovery, 332/332 passed.
• Dependency pins match the validated runtime exactly.
• git fsck --no-progress: passed.
• Independent scoped code review: APPROVED, no open Critical/Important/Minor findings.

TensorFlow ran on CPU and emitted the existing CUDA-unavailable and TF1 compatibility warnings; they did not produce test failures or skips in the TensorFlow suite.