feat(issue54): add multi-seed MoT routing stability and architecture controls

[PinkTulips139]

Summary

This PR implements and documents the Issue #54 study of multi-seed MoT routing stability, reproducibility, interpretability, and architecture controls on VisDrone2019-DET.

The highest-level experimental unit is an independently trained seed. Images and tokens are not treated as independent training repetitions.

Related to #54.

Main changes

• Added experiment manifest schemas, registries, and checkpoint SHA256 validation.
• Added safeguards preventing one checkpoint from being represented as multiple independent seeds.
• Added MoT routing export and cross-seed alignment by image, layer, and expert.
• Added routing metrics including:
  • dominant expert agreement;
  • token-level top-1 agreement;
  • Jensen–Shannon divergence;
  • routing entropy;
  • expert utilization;
  • between-seed variation.
• Fixed the MoT sparse-fusion AMP Float/Half dtype conflict.
• Added precision calibration and recoverable experiment controllers.
• Added isolated Phase 3 runners and strict serial queues.
• Added tests, formal reports, manifests, CSV summaries, and integrity checks.
• Added architecture controls for EsMoE and MoA without fabricating MoT routing records.

Formal experiment scope

All formal runs used VisDrone2019-DET, 30 epochs, batch size 8, and image size 640.

Architecture	Independent seeds	Precision
MoT (v10_mot)	5	FP32
EsMoE (v10)	3	AMP
MoA (v10_moa)	1	AMP

All nine formal runs completed successfully. Their checkpoint SHA256 values were verified and are mutually distinct.

Detection results

Architecture	mAP50	mAP50-95
MoT, 5 seeds	0.16037 ± 0.00293	0.08311 ± 0.00183
EsMoE, 3 seeds	0.16001 ± 0.00350	0.08368 ± 0.00200
MoA, seed 0	0.15844	0.08164

_Architecture controls under the shared VisDrone2019-DET protocol. Error bars show sample standard deviation where multiple independent seeds are available; MoA is represented by one descriptive seed._

MoA currently has only one independent seed, so no between-seed variance estimate is reported.

MoT routing results

Across the five independently trained MoT seeds:

• mean dominant-expert agreement: approximately 0.5260;
• mean token-level top-1 agreement: approximately 0.4354;
• routing stability differs substantially across the six captured MoT layers;
• mean route entropy is close to the three-expert theoretical maximum;
• repeated export from the same checkpoint produced token-level agreement of 1.0.

_Cross-seed routing agreement across the six captured MoT layers. Entropy and routing agreement measure different properties; high entropy is not interpreted as high routing stability._

The evidence therefore indicates relatively stable detection performance alongside only moderate or low internal routing agreement, with strong layer-level differences.

Evidence boundaries

This PR does not claim that:

• MoT is universally superior to EsMoE or MoA;
• MoA has demonstrated cross-seed stability;
• the architecture differences are statistically significant;
• routing instability necessarily reduces detection performance;
• a specific expert has a fixed semantic role;
• high routing entropy implies high routing stability;
• images or tokens can replace independent training seeds.

The architecture comparison is descriptive and specific to the current protocol.

Integrity and reproducibility

• All nine formal runs have status=passed.
• All formal runs completed 30 epochs.
• Checkpoint SHA256 values match their manifests.
• All nine formal checkpoint hashes are distinct.
• EsMoE and MoA contain no fabricated MoT routing artifacts.
• Formal report files include SHA256 integrity records.
• Large checkpoint binaries and datasets are not included in this PR.

Final validation

• CI run #234 passed at commit
  cdffaaf3e30d1ff742c0cd32f7ce7c295f7a0ade.
• Tests passed on Ubuntu, macOS, Windows, and ARM with Python 3.13.
• The minimum compatibility shard passed with Python 3.8 and Torch 1.8.
• MixtureDDP, MixtureP0Regression, IsolatedExports, Benchmarks, and the
  required Summary gate passed.
• All nine formal runs completed 30 epochs and have mutually distinct,
  manifest-matching checkpoint SHA256 values.
• Checkpoints, datasets, raw logs, and the large raw cross-seed JSON are not
  distributed in this PR.
• The omitted raw JSON remains identified by its filename, size, SHA256,
  formal provenance, compact global summary, and formally derived CSV reports.
• The public evidence was reduced from approximately 103,979 added lines to
  approximately 21,916 added lines without changing the formal results.

Branch status

The earlier note that this branch was behind upstream main is no longer
current. The reviewed upstream commit has been integrated through the normal
merge commit above. No rebase, force push, or history rewrite was used.

Review status

This pull request is ready for maintainer review.