Clone on Linear, Conv2d and Embedding shares parameters with the original

[vyncint]

Summary

Param is an Arc<RwLock<Tensor>> handle and the layers derive Clone, so a cloned layer trains together with its original. LayerNorm, BatchNorm2d, Dropout and Sequential are not Clone, so the surface is also inconsistent.

Where (main at 853a224)

• crates/oxmera-nn/src/param.rs:12-15 — #[derive(Debug, Clone)] pub struct Param { inner: Arc<RwLock<Tensor>> }
• crates/oxmera-nn/src/linear.rs:11-15, conv.rs:14, embedding.rs:11 — #[derive(Debug, Clone)] on the layer.

Reproduction

let a = Linear::new(2, 2, 1);
let b = a.clone();
a.weight().set(Tensor::zeros([2, 2]));
assert!(b.weight().value().to_vec_f32()?.iter().all(|&x| x == 0.0)); // holds: b changed too

Fix

Either remove Clone from the layers (a declared break in the 0.x window) and offer fn deep_clone(&self) -> Self / Param::detached_copy(), or keep Clone and document the aliasing on every layer and on Param. Weight tying should be a visible choice, not a side effect of .clone().

Done when

Cloning a layer either copies its parameters or is impossible; the chosen behaviour is documented on Param and tested.

[mikemikimike]

Implemented in #79.

The fix makes parameter copying explicit and prevents copied Linear, Conv2d, and Embedding layers from silently sharing trainable storage. It preserves shared Param handles for optimizer and module parameter-list semantics, while deep_clone() creates independent weights and optional biases.

Regression tests cover independent weight and bias storage for all three affected layer types. The PR's GitHub checks all passed, including both main CI platforms, CPU-only coverage, MSRV, dependency policy, research checks, CUDA convergence checks, stress tests, and DCO/attribution validation. The PR post-review also passed the workspace test suite and documentation build in the isolated runner after preserving the repository's LF-sensitive CUDA source files.

PR: #79