T-G1 — VRAM estimation before launch

[lightningpixel]

Epic: Epic G — VRAM pre-flight & job queue
Priority: P2

Context

Likely the #1 support ticket (IDEAS.md) — OOM crashes. Becomes more pressing in 0.5 since Epic A adds heavy ops (remesh, booleans in Epic D) that compete for the GPU with the agent (Ollama) and the generators. Verified state: vram_gb is declared statically per generator in the manifest (generator_registry.py:765) and is only used for display today — nothing checks it before launch.

Scope

• Before a generation or a heavy registry op (remesh, booleans), compare the declared vram_gb to measured free VRAM.
• If insufficient: show an actionable message ("this model needs 10 GB, you have 6 GB free — unload X first?") instead of letting it crash.

Acceptance criteria

Launching a too-heavy model with no free space shows the estimate before any crash.

---

From ROADMAP_0.5.md.