decision: the memory space stays out of the flat path's tensor identity (RFC §3.4 withdrawn)

The spelling RFC's §3.4 asked for the resolved memory space in tensor_gid. The placement rides on the TensorAlloc instruction on the flat path (src/hir/flatten.rs, the alloc's spare operand), the emitted memref carries no space, and the mangler excludes placement on purpose. flat_matches_ast_two_tensors_differing_only_in_placement in tests/integration_test/flat_codegen_differential.rs compiles two tensors that differ only in placement in one program and passes on both paths, so the change would split TensorMap entries without changing a byte of output. I withdrew it (review §1.3) rather than ship an identity change with no observable effect.

If the bounded-extents work later needs the space in the identity (a bound derived from a machine-file figure is one such case), this is where it goes. Confirm or reopen.

Read the placement handling in src/hir/flatten.rs and the flat-path identity behavior, then run flat_matches_ast_two_tensors_differing_only_in_placement in tests/integration_test/flat_codegen_differential.rs. The current issue records a withdrawn RFC change; done means confirming that placement should remain excluded, or reopening the identity work if bounded extents require it.