S2.64: SAM 2.1 Small, Base+ and Large, each checked against PyTorch

Hiera configs and size detection from the weights, registry entries pinned to
commits with verified SHA-256, the loader picks the size by checksum. The GPU
encoder submits once per block so Large fits in memory, and uploads no longer
panic on a lost device. Reference tests run per size on the CPU and GPU.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

crates/engine/src/analysis/models.rs

@@ -336,7 +336,7 @@ fn load_file(path: &Path, id: &str) -> std::result::Result<Loaded, String> {
     let bytes = std::fs::read(path).map_err(|e| format!("cannot read {}: {e}", path.display()))?;
     // SAM 2 runs on candle and is native only; every other model loads through the registry.
     #[cfg(not(target_arch = "wasm32"))]
-    if id == seg::SAM2_TINY.id {
+    if effectcraft_sam2::mask::SIZES.iter().any(|m| m.id == id) {
         return effectcraft_sam2::mask::load(&bytes).map(|m| Loaded::Mask(m));
     }
     seg::load(id, &bytes)