docs(lora): add audited Issue #50 strong-baseline research extension

[PinkTulips139]

Summary

Extends the existing Issue #50 LoRA rank-sweep work with an audited research benchmark covering:

• strong fine-tuning baselines;
• numerical-stability diagnostics;
• multi-seed evaluation;
• formal result-validity semantics;
• requested-versus-actual batch auditing; and
• dependency-free result validation.

This PR does not replace historical LoRA rank-sweep results or modify their recorded metrics.

What this adds

• Head-only, Neck + Head, Last Stage + Neck + Head, Full Fine-tuning, Stable LoRA, and Partial Fine-tuning + LoRA comparisons.
• Audited formal results containing only valid completed runs.
• Multi-seed means, sample standard deviations, and best-run values.
• Separation of requested batch, actual batch, queue execution status, and formal validity.
• Exclusion of duplicate runs, duplicate result directories, invalid OOM recovery entries, and implementation failures.
• A dependency-free validation script for CSV schema, unique seeds, sample statistics, and the audited four-objective non-dominated front.

Key results

• Brain Tumor: Last Stage + Neck + Head reached 0.39172 ± 0.02316 mAP50-95 across seeds 0/1/2, with a best run of 0.40944.
• VisDrone: Full Fine-tuning with actual batch 4 reached 0.15342 ± 0.00221, with a best run of 0.15527.
• Stable LoRA did not provide an absolute accuracy advantage under the audited protocols.
• Head-only remained parameter-efficient on Brain Tumor.

Scope and limitations

• VisDrone LoRA used fraction=0.2.
• Several comparison methods currently have only one valid seed.
• AMP-safe LoRA was not validated successfully and is excluded from formal results.
• Adapter LR 0.1× and Router LR 0.5× are protocol-specific diagnostic settings, not universal optimal values.
• The experiments were produced on the historical research branch.
• They include upstream fixes fix(moe): AMP-safe index_add_ dtype alignment #124 and fix(lora,ddp): prevent MoE control-path LoRA ready-twice crash #125 but predate fix(lora): honor fallback RS-LoRA scaling #170 and fix(lora): preserve alpha warmup across EMA lifecycle #177.
• Therefore, these results must not be interpreted as final RS-LoRA performance on the current upstream/main.

Provenance

• Audited Phase 2 source commit: bb20447
• Published research-extension commit: 33e62d8e8548b2690e6b1acb9fad9776cd94923a
• Upstream PR branch commit: 9a33ac3fd1300bd13693c265ef30d6b9810ca1a9

Validation

• ruff check examples/lora_examples/validate_yolo_master_issue50_research_extension.py
• python -m py_compile examples/lora_examples/validate_yolo_master_issue50_research_extension.py
• python examples/lora_examples/validate_yolo_master_issue50_research_extension.py
• git diff --check

The validator confirms 14 valid formal runs, 9 method summaries, unique configuration/seed combinations, sample-standard-deviation semantics, and the audited four-objective non-dominated front.

[PinkTulips139]

CI note: the only failing job is the legacy Python 3.8 / PyTorch 1.8.0 test environment. The three failures occur in existing Mixture/MoA/MoT tests because torch.autocast is unavailable in PyTorch 1.8. This PR changes only LoRA research documentation, audited CSV tables, and a dependency-free validation script. The modern tests, benchmarks, isolated exports, MixtureDDP, and Mixture regression checks passed.