feat(strict): flat-emission wavefront walk as an opt-in traversal for the fused lane

[TobiBu]

Stacked on #340 (merged into perf/small-leaves, the base). Depends on yggdrax TobiBu/yggdrax#74 (mac_type, peak_wavefront on dual_tree_walk_mutual) -- merged. Plan: the "tree walk" plan of 2026-09-10; record in docs/tree_walk_2026-09.md.

Why. At N=200k, leaf 64, theta 0.6 a strict_run_v2 step costs 177 ms with the CSR M2L lane on, ~100 ms of it in the traced dual-tree walk and its list emission: dense per-node output rows (820 MB at leaf 64) loop-carried and copied through lax.cond, a sort per round, and a post-loop flatten over total_nodes x K slots. Measured in isolation the traced dual walk costs 503 ms per call and yggdrax's flat-emission dual_tree_walk_mutual 39.9 ms (13.4 ms with int32 indices and a queue sized to the measured peak) for identical pair counts.

What. The strict streamed split build routes BY DEFAULT (JACCPOT_STATIC_STRICT_FUSED_FLAT_WALK=0 restores the dual walk) to dual_tree_walk_mutual fed the dual walk's own mac_extents and mac_type, so the far and near lists are the dual walk's as sets:

• far pairs un-mutualised and interleaved so the live pairs stay a prefix (every M2L consumer masks idx < far_pair_count); capacity = the compact far-pair cap;
• the leaf neighbour CSR from one stable argsort of the directed near pairs plus searchsorted offsets, width = the neighbour-edge cap, no per-leaf row buffer;
• eager: queue ladder on queue_overflow; a far/near overflow doubles a cap the caller did NOT name in the env (floors 131072 / 2^21 directed pairs, ceilings 2^26 / 2^28) and raises naming a cap the caller did name; the settled widths travel through _strict_fused_capacity_handoff as floors to the traced refresh and later eager prepares, so scan carry shapes match and never shrink; traced: any overflow saturates far_pair_count to the capacity, which trips strict_run_v2's existing saturation guard (documented at the guard);
• the capacity report is always emitted and carries peak_wavefront; _strict_fused_capacity_handoff sizes the traced queue as pow2(1.5 x peak) on this lane.
• Defaults (second commit). Both JACCPOT_STATIC_STRICT_FUSED_FLAT_WALK and JACCPOT_STATIC_STRICT_FUSED_M2L_CSR (feat(m2l): target-tiled CSR Pallas M2L kernel with on-chip rotations (opt-in lane) #340's CSR kernel, sm_80 gate unchanged) default to 1. A merely-defaulted flat-walk flag falls back to the dual walk QUIETLY when the configuration cannot take it (treecode walk requested, a MAC other than bh/dehnen, a solver-owned pair policy, the non-flat far-pair layout); an explicit 1 there still raises.
• tests/conftest.py honours YGGDRAX_WORKTREE (it put the sibling checkout, a paper branch, at the front of sys.path unconditionally).

Checks. test_flat_walk_default_dispatch.py (CPU, 9): default routing, =0, the fallback/raise pairs, named-vs-unnamed caps, the floor handoff. Seam test (12): set parity vs the dual walk for bh/dehnen at three thetas and a radius scale, prefix-live far pairs with both directions, valid CSR, no self/duplicates, traced call keeps capacity width, overflows raise naming the knob, odd caps rejected, queue ladder grows, unnamed caps grow. GPU wiring test (3): lane entered by default, =0 takes the dual walk, forces within 2e-5 of the dual lane, both flags refused, treecode flag alone wins. End to end at N=200k leaf 64: aggL2 vs fp64 direct 1.2011e-03 (identical to the dual lane), trajectory-recovered force vs eager at steps 1-2 < 7e-4. test_strict_run_v2_refresh_capacity parametrised over the flag. Per step, clean idle A100, N=200k p4 th0.6 (flat walk + CSR + int32): leaf 256 96.2 ms, 128 71.3, 64 63.3 (from 409.6 at the start of the small-leaves work, 176.7 with CSR alone), 32 82.5; th0.8: leaf 64 41.6, 32 53.6. Per-order sweep at leaf 64 monotone and identical to 4 digits to the dual lane; 300-step dynamics flag on/off: no bias (docs/tree_walk_2026-09.md).

🤖 Generated with Claude Code