fix: pre-flight disk space + cap GPU offload so small-GPU runs don't hang

[rjckkkkk]

Problem

An open-source user ran aima run qwen3.5-9b (llamacpp/k3s) on a laptop with a 4 GB RTX 2050 and got Timed out waiting for deployment to be ready, stuck at scheduling 2% for hours. Two independent product bugs combine here:

1. No disk pre-check before model download. The Qwen3.5-9B-IQ4_NL.gguf (~5 GB) streamed straight to disk with no free-space check, filling the filesystem past k3s's disk-pressure eviction threshold. The node got tainted node.kubernetes.io/disk-pressure:NoSchedule, so the pod never scheduled — the only symptom was the generic timeout. (kubectl describe pod → FailedScheduling ... untolerated taint {disk-pressure}.)
2. Full GPU offload forced on a GPU too small. The generated command was llama-server ... --n-gpu-layers 999 (offload all layers to GPU). 999 is a catalog default that assumes the model fits VRAM; on a 4 GB discrete card a 9B model can't, so even after the disk issue it would CUDA-OOM at load and time out again.

Fixes (two commits)

1. fix(resolver) — VRAM-aware offload. When a discrete GPU can't hold the model's footprint (ram_min_mib proxy > per-GPU VRAM), downgrade the engine-declared offload_config_key to CPU and surface a clear warning. It still deploys and works — just on CPU — matching AIMA's graceful-degradation model. Driven entirely by catalog YAML (offload_config_key), so no engine name / config key is hardcoded (INV-1/2). Unified-memory hosts (APU / GB10 / Apple) are untouched — the GPU shares system RAM there, so n_gpu_layers=999 is correct (this is exactly the verified Strix Halo config). Explicit user overrides are respected.

2. fix(run) — disk pre-flight. The moment a download's total size is known (from the existing downloader progress — reported before transfer for HF/ModelScope), abort if it won't fit while keeping a disk-pressure reserve (10% of the volume, 2 GiB floor), returning a clear fatal error (needs ~X MiB … only Y MiB free … short by ~Z MiB; free up space or choose a smaller model) instead of a disk-filling hang. Adds a small cross-platform hal.DiskUsage wrapper. No new download code (reuses the size the downloader already computes).

Tests & verification

• Unit tests: internal/knowledge/offload_test.go (downgrade / unified-memory left alone / fitting-GPU left alone / user-override respected / no-offload-key noop / warnings surfaced through CheckFit), cmd/aima/diskcheck_test.go (fit / disk-pressure-trip / reserve floor / unknown-size no-gate).
• Verified end-to-end against the real embedded catalog + actual qwen3.5-9b.yaml: on a simulated 4 GB discrete NVIDIA → engine=llamacpp n_gpu_layers=0 + warning; on unified-memory GB10 → n_gpu_layers=999 preserved.
• go build ./..., go vet, and full go test ./... all pass.

Not in scope

• Precise partial GPU offload (e.g. "12 of 40 layers") — the catalog carries no per-model layer count, so any nonzero guess risks OOM; CPU fallback is the safe floor. Adding layer-count metadata could enable this later.
• Engine-image pull disk checks (this addresses the reported model-download case).