Backpropagate v1.7.0

Added

• Card-aware full fine-tuning ceiling. mode="full" no longer caps at a hard 4B. The ceiling is derived from the 4-addend training-memory arithmetic against detected VRAM — 16 GB → 4B, 24 GB → 5B, 32 GB → 6B pure-GPU (_full_ft_ceiling_for_vram). Override explicitly with --full-ft-ceiling-billions / Trainer(full_ft_ceiling_billions=...).
• FSDP2 CPU-offload full fine-tuning — single-card 7B-class. --full-ft-offload / Trainer(full_ft_offload=True) configures FSDP2 fully_shard + CPUOffloadPolicy + activation checkpointing + bf16 to spill params + optimizer state into host RAM, lifting the full-FT ceiling to ~8B / 7B-class on a 32 GB card (+~64 GB host RAM). It's a documented escape hatch: slower (PCIe/CPU-bandwidth-bound) and needs an NCCL backend (Linux/WSL2). On a host without NCCL (e.g. Windows-native) it fails fast with the new DEP_FSDP_UNAVAILABLE rather than silently running without offload and OOMing. The trainer initializes a single-process group automatically (no accelerate launch / torchrun needed).
• 24–34B QLoRA model presets. New presets for the 32 GB envelope: qwen2.5-14b (~8.5 GB, the daily-driver sweet spot), mistral-small-24b (~18 GB), qwen2.5-32b (~26 GB, just fits at max_len 2048), and llama-3.1-8b. rank=alpha per size, paged 8-bit AdamW.
• 32 / 48 GB batch + VRAM tiers. Trainer._detect_batch_size and the estimate-vram tier table gained 32 GB (batch 6) and 48 GB (batch 8) tiers. estimate_vram(..., offload=True) models the FSDP2 offload path and reports a new host_ram_gb.
• Contrastive full-FT error. RUNTIME_FULL_FT_MODEL_TOO_LARGE now names the right recovery for the situation — --full-ft-offload when the model fits the offload ceiling, LoRA/QLoRA when it's past even that.

Changed

• Repositioned for the 32 GB envelope. README, handbook (full-fine-tuning, cli-reference, error-codes, estimate-vram), and presets now lead with single-card 7B-class full fine-tuning + 24–34B QLoRA → Ollama, with an honest "scales down to 16 GB" note. The old "7B+ full FT → use HuggingFace/A100" anti-pitch is replaced by the --full-ft-offload path.
• MLX rail reframed to an explicit unverified preview. All "New in v1.5" version hype removed from the MLX surfaces (code docstrings, feature flags, --backend help, README, model cards). The rail is kept but labeled experimental / not dogfood-verified on real Apple Silicon / not part of the supported feature set. No functional change.