ADR 0004 - creation_from_tensor test category + tracking-scope broadening

rstsr implements the array-joining / construction-from-tensor functions (concat/concatenate, stack, hstack, vstack, diag, meshgrid, unstack) in rstsr-core/src/tensor/creation_from_tensor.rs. The prior NumPy-parity audit (ADR-0003) covered only the manipulation/ surface (reshape/transpose/swapaxes/moveaxis/squeeze/expand_dims/flip/broadcast/to_contig/ to_layout); concat/stack/diag were not covered by that audit.

This ADR decides where parity tests for those functions live, and how functions rstsr does not yet implement are tracked.

1. Decision

1. Parity tests for creation_from_tensor.rs functions live in a dedicated creation_from_tensor/ category under tests/core_func/ (mirrored under doc_draft/ for doc tests). The category name mirrors the source module src/tensor/creation_from_tensor.rs, consistent with how manipulation/ mirrors src/tensor/manipulation/ and linalg/ mirrors src/tensor/linalg/.
2. The tracking scope (ADR-0003 artifact (b) numpy_coverage.csv and the sync_numpy.py SURFACE) is the manipulation + creation_from_tensor surface, not manipulation alone. Functions rstsr does not yet implement (split/array_split/hsplit/vsplit/dsplit, dstack, column_stack, block, tile) are recorded as todo rows - one representative NumPy test method per class, with the full method count cited in the note.

2. Considered

• Fold join functions into manipulation/ - rejected. rstsr's source groups them in creation_from_tensor.rs, not manipulation/; the test category mirrors the source module so the mapping is mechanical.
• Per-function categories (concat/, stack/, ...) - rejected; too granular. One category per source module, holding all its parity tests, matches the existing manipulation/ granularity.

3. Consequences

• The category mirrors the source module, so a function newly added to creation_from_tensor.rs maps to a parity test in creation_from_tensor/ mechanically - no per-function category decision.
• todo rows make unimplemented functions visible rather than silently absent: a future implementation or a NumPy version bump surfaces them as tracked candidates.
• Divergences for this surface (intentional or since-fixed) live in the ADR-0003 (c) reports (numpy_differences.md / numpy_differences_resolved.md), not duplicated here.