fix: audit v2 P0/P1 fixes (D1/D2/D3/D7/D9/D11/D14)

fix: distillation script uses FusionModel + add GRPO training script

fix: comprehensive audit fixes (ONNX/export, benchmark, deprecated APIs, attention unification)

fix: replace random embeddings in MoverScore with deterministic hash

fix: move FusionMini @classmethods from FusionMiniConfig to FusionMini class

fix: test suite + bilingual filter + TensorRT/OpenVINO export rewrite

refactor: unify Thinking Dial to architecture-level only, deprecate text injection

fix: N9 position_ids signatures + M2 THINK_END token split

chore: remove temp verification scripts

fix: comprehensive audit fixes - Thinking Dial unification, deployment scripts, README, bilingual filter, data download

fix: from_pretrained weight loading (HF 5.x compat) + _init_weights + tied_weights_keys

fix: CausalLMOutputWithPast return + audit verification

feat: setup.py, quick-start guide, architecture doc, model registry

fix: short-sequence benchmarks + ThinkingDial vs vanilla experiments

feat: GSM8K reward function + GRPO evaluation pipeline (v2.1.1)

v2.2.0: end-to-end pipeline validation + GRPO no-tokenizer fallback

v2.1.0: fix N25 off-by-one, S2 reward_fn string safety

v20: fix N23/N24/M7/S2 - critical regression fixes from v1.9.0 audit

v19: fix N23-N25 + S1 reward functions + F1/F2/F3 documentation

v18: fix N18/N19/N20/N21/N22 Thinking Dial pipeline bugs

v17: fix Thinking Dial training pipeline (N10/N11/N12/N13/N14/N15/N16/N17)

v16-hotfix: fix kv_head_dim bug, add incremental gen test

v16-docs: fix README accuracy, add F3/S2 disclaimer to distillation script

v16: fix critical audit defects (N6/N7/N8/N9 + S3/S4/M2/M5)

v15: fix remaining audit issues (F1/F2/S1/M3/M8/doc)

v14 fix: F1 ThinkingDialModel thinking_embedding/gate 实际使用，F5 top_p NaN guard，M1 RoPE position_ids，M6 position_ids 传递链，DPO tokenizer 属性，forward() position_ids 参数

fix: 修复 F-NEW-13 GQA双重扩展 + S-NEW-13 GQA KV维度 + LOW清理

fix: 修复所有测试警告 + 清理调试脚本

fix: 修复 SBLAttention forward_with_qkv() 形状错误和测试文件结构

Fix tests: FusionMiniLayer Q/K/V reshape + logging import + pytest compat

fix: external audit issues from commit 08f2cd2

fix: SBLAttention forward() guard for removed projections

fix: v13 external audit - CRITICAL/HIGH/MEDIUM fixes

fix: comprehensive audit fixes - CRITICAL/HIGH/MEDIUM/LOW

Release: v1.2.0 - Add BLEU/ROUGE/METEOR, TensorRT/OpenVINO, AMP training, LIME/SHAP

Feat: Add LIME and SHAP model interpretability tools

Feat: Add AMP trainer and gradient accumulation for training optimization

Feat: Add TensorRT and OpenVINO deployment options

Feat: Add BLEU, ROUGE, METEOR evaluation metrics

Feat: Add GGML/ONNX deployment and optimized SBLA attention (0.49ms)

Release: v1.1.0 - Add BERTScore, MoverScore, visualization, deployment options, and optimize SBLA attention

Feat: Add graphical visualization tools (attention heatmap, loss curve, architecture)

Feat: Add BERTScore & MoverScore evaluation metrics

Feat: Add model visualization tools (attention, loss curve, architecture)

Feat: Add advanced evaluation metrics (Distinct-n, Repetition Rate, Token Entropy)

Test: Add SBLA attention performance test (2.07ms avg)

Release: v1.0.0 - First official release (complete documentation + tests + changelog)

Test: Add edge cases tests (4/4 passed)

Feat: Add actual training (100 steps) + documentation (tutorial + API)

Train: Add 10-step training test (loss decreases steadily)