[asjad2401]

Stacked on #3 — merge that first (this PR's base will retarget to main).

Summary

Every failure path used to be an .expect() / panic!, which reaches Python as pyo3_runtime.PanicException (a BaseException, so except Exception doesn't catch it). Now:

Situation	Before	After
float64 / list input	pyo3 conversion error	TypeError: input 'x': expected a float32 array, got float64 (use .astype(np.float32))
wrong rank / static dim	panic deep in an op	ValueError: input 'x': expected shape [?, 3], got shape [2, 4]
op or backend failure	PanicException + Rust stderr dump	RuntimeError: node 'gemm1' (Gemm): matmul: inner dimensions must match ...
missing output	panic	RuntimeError: model output 'y' was never computed

How:

• Ops return OpResult; ExecutionContext::require replaces .expect() on inputs.
• Each node runs under catch_unwind, so panics inside Burn are converted too, with node name and op type attached.
• The default panic hook is silenced (the message is already in the exception) unless RUST_BACKTRACE is set.

Bug fix found along the way

Fortran-ordered inputs such as x.T were read in memory order, silently scrambling the data — rust-numpy's as_slice() succeeds for F-contiguous arrays too. Only C-contiguous arrays now take the fast path; everything else is gathered in logical order.

Testing

tests/test_ops.py gains error-path cases (dtype, non-array, rank, static dim, input count, backend panic → RuntimeError) and non-contiguous inputs (transposed, strided, negative stride). All 24 checks pass, plus test_bridge.py, compare_ort.py, cargo fmt --check, and clippy -D warnings.

Note

Node names in errors are onnx-ir's generated names (gemm1), not the original names in the ONNX file — onnx-ir renames nodes during parsing.

🤖 Generated with Claude Code