Compact permuted_dense (no inverse-permutation arrays) for kron-path blocks

[Transurgeon]

Every permuted_dense allocated col_inv (global n ints) and row_inv (global m ints) at construction. On the dense left_matmul kron path this metadata dominates: p independent blocks each carry a full n_vars-sized col_inv, O(p*n_vars + p^2*m0) ints per node against exact-nnz values — 5.9 GB engine peak on OptimalAdvertising where true nnz needs ~5 MB.

Change

This introduces a compact PD variant — storage proportional to the block, not the global shape — scoped to the kron-csc family:

• new_permuted_dense_compact leaves both inverse arrays NULL; the default constructor is untouched, so every non-kron PD keeps today's layout and code paths.
• BA_pd_csc_alloc builds its output compact; compactness propagates through copy_sparsity / transpose / index chains of compact sources.
• Init-path membership tests (idxs_hits_set callers, index_pd_alloc) gate on NULL and fall back to binary-search scans of the sorted perms (sorted_pos / sorted_hits).
• Eval-path consumers materialize the operand-side array at alloc time only when the product is non-empty (permuted_dense_ensure_col_inv / _row_inv), so every fill kernel is untouched — BA_pd_csc_fill_values asserts the contract (inv != NULL || C->n0 == 0); index_pd_fill_values of a compact source reads a source-row map precomputed into kernel_iwork.
• The mutable kron scratch keeps its arrays (its kernels write col_inv).

Verification

• All 421 tests pass (Debug, ASan+UBSan, leaks clean). New peak-memory regression test (row-sum + col-sum Jacobian at 128x128): init peak 1324 -> 292 bytes/var (21.7 MB -> 4.8 MB), asserted < 400 bytes/var.
• Backend correctness vs CPP (get_problem_data, CLARABEL): CVaR, SDP, Huber, TvInpainting, QuantumHilbertMatrix MATCH exactly; OptimalAdvertising matches except one pre-existing NaN-constant-folding difference (identical on the baseline engine).
• OptimalAdvertising engine-internal peak: 5901 MB -> 621 MB; the residual is the per-branch CSC cache (documented follow-up). Cold extraction: 5.12 s -> 2.33 s.
• CVaR (8.9 GB) and SDP (941 MB) peaks are unchanged — they sit on the kron pd/spd paths (BA_pd_pd_alloc / BTA_pd_spd_alloc outputs are still full by scope). Extending compactness there is the natural follow-up.

Before merge

• Full 25-problem timing + memory suite gate (V6 of the plan) still to run.
• Bump the SparseDiffPy submodule after this lands.

🤖 Generated with Claude Code