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

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

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: SBLAttention forward() guard for removed projections

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

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

Test: Fix training basic test (torch.optim not torch.nn)

Test: Fix training basic test (remove DeepSpeed dependency)

Test: Fix and simplify inference basic test (remove non-existent imports)

Test: Fix GBK encoding error (remove emoji)

Test: Add tiny config test (reveals tests are slow not hanging)

Test: Add simple import test and minimal training test

Feat: Add model evaluation metrics and basic tests

fix(v12): resolve 6 defects from v11 audit (F-NEW-6, S-NEW-5, M-NEW-5/6, MI-NEW-4/5/6)

fix(v11): resolve remaining 5 defects (S-NEW-1/2, M-NEW-2/3, MI-NEW-1)

fix(v9): resolve 3 fatal + 1 serious + 1 moderate defects from external audit

fix(v8): S1 lora vocab sync, M1 test_sbla tuple unpack, N1 dead code removal

fix: v5 remove empty tokenizer.json, unify think tokens, integrate tokenizer into training pipeline

chore: replace author real name with GitHub username

fix: v3 comprehensive defect repair

fix: 修复 S1-S4/M2-M4 缺陷，规范化工程结构

feat: 完成 SBLA 注意力集成 + MVP 训练流程

feat: 初始化 Fusion 项目 - 六边形开源大模型