Add a row_reduce primitive and migrate sum onto it (lesson 3, M3)

[dance858]

Summary

Lesson 3, milestone 3. Stacked on #123 (the problem_jacobian fix), which its problem-level test depends on; retarget to main once #123 merges.

• New vtable pair row_reduce_alloc(A, group, m_out) / row_reduce_fill_values(A, C): C[j, :] is the sum of rows i of A with group[i] == j. The input-nnz to output-position map is bound to the result (bound_iwork), so the fill takes no map. Fills zero the output and accumulate, never bump values_version (the eval wrapper does), and are safe under the affine skip.
• sum builds group from its axis (0, i / d1, i % d1 over the child's column-major rows) and calls the primitive. sum_expr.idx_map and free_type_data are gone.
• Kernels: sparse builds each output row's sorted column union with a marker array (same shape as the old block-sum helper, generalised); permuted_dense keeps col_perm and collapses the hit groups into a sorted row_perm; stacked_pd reduces per block and goes through the existing coalesce-accumulate skeleton, keeping the raw spd on pre_coalesce.
• Intended behaviour change: sum of a stacked_pd child is stacked_pd for every axis. Previously axis -1 gave a 1-row pd and axes 0/1 fell back to CSR. Trade-off: the merged block takes the column union per row signature, so non-kron-shaped spd children can carry structural zeros where the CSR fallback had exact nnz (same trade-off as transpose_spd).
• Deleted: sum_row_partition_alloc (slot, typedef, doc), sparse_sum_row_partition_alloc, sum_all_rows/sum_block_of_rows/sum_evenly_spaced_rows_csr_alloc, the three pd helpers plus dispatcher, the spd fallback, and two dead includes in stacked_pd.c. accumulator, the trace helpers, sum_4_csr_alloc, compose_csr_idx_map_for_spd and sorted_union_int_arrays stay (other callers).
• tests/profiling/profile_log_reg.h: the manual "Path B" baseline now uses the primitive on the PD (note: it times the dense reduce rather than the CSR accumulator). Values still match Path A to 0.

Tests

• New tests/utils/test_row_reduce.h: sparse (column merge, empty output row, NaN-poisoned fill proving the memset, refill), pd (non-monotone collapsing group, bound_iwork, twin match), spd cross-block accumulate on a shared column, spd within-block routing, spd all-to-one, and an alloc-free spd fill under SP_TRACK_MEMORY.
• Atom level: sum of a 2-block spd child for axes -1/0/1 asserting is_stacked_pd plus numerical Jacobian check; axis 1 also checked exactly through the CSR view.
• Problem level: sum(A @ X, axis=1) as a constraint reaches the aggregated Jacobian in row order.
• Removed the three pd unit tests that inspected the caller's idx_map.

457 tests Release, 460 with SP_TRACK_MEMORY. UBSan, guard-malloc + leaks, PROFILE_ONLY (built and run), clang-format clean. Mutation checks, each rebuilt in isolation: dropping the sparse memset, reading the wrong pd source row, plain scatter instead of accumulate in the spd fill, and swapping the atom's axis formulas all fail the suite.

🤖 Generated with Claude Code

https://claude.ai/code/session_01B6Bs765i1HTbqx3LUu6LMP