[PERF IMPROVEMENT] func_broad_phase: hoist loop-invariants, cache i_pair, swap-and-pop, gate dead writes

[lohiaj]

Summary

Four orthogonal cleanups in genesis/engine/solvers/rigid/collider/broadphase.py, all in the hot func_broad_phase_*_kernel_2_range_for. They preserve the SAP set-of-overlapping-pairs invariant.

1. Hoist loop-invariant sort_buffer reads out of the inner per-active sweep loop. The outer i and batch i_b are constant across the inner for j in range(n_active), but the DSL/compiler can't safely CSE through the calls to func_check_collision_valid and func_is_geom_aabbs_overlap, so sort_buffer.is_max[i, i_b] and sort_buffer.i_g[i, i_b] were re-loaded once per inner iteration. Now hoisted into is_max_i and i_g_i once per outer step. Same fix in both the no-hibernation and hibernation branches.

2. Cache collision_pair_idx[i_ga, i_gb] once per pair. It was loaded once inside func_check_collision_valid (validity check) and again, separately, on the no-overlap fall-through that clears contact_cache.normal — two indirect loads of the same index with a function call in between. Now loaded once before the validity call and threaded through as i_pair; func_check_collision_valid's signature gains the i_pair parameter and drops the now-unused collider_info. All three call sites updated.

3. Swap-and-pop active-buffer removal. The original linear search + linear shift to preserve insertion order cost O(n_active) per removed geom. SAP only reads active_buffer as a set in the inner pair check, so swap-and-pop preserves correctness and reduces each removal to O(1). The inner for k in range(j, n_active - 1) shift is gone. Applied to all three removal sites (active_buffer, active_buffer_hib, active_buffer_awake).

4. Gate dead min_buffer_idx / max_buffer_idx writes in the first-time init path on qd.static(use_hibernation). Those two arrays are only read inside hibernation-gated branches, so without hibernation the writes are dead stores; the static gate elides them at compile time.

Correctness

Differential test vs pristine amd-integration HEAD: position and velocity statistics match to within floating-point noise; contact counts identical.

Scope

Single file, +40 / -36 = +4 LoC net. No changes outside broadphase.py.

E2E at 8192 envs:

throughput: 709,923.7 env*steps/s
wall_time: 5.770 s

[yaoliu13]

This PR is not based on the latest amd-integration: perf/jlohia/broadphase-cleanup...ROCm:Genesis:amd-integration

[lohiaj]

Closing this in favor of the monolith-focused PR #34. The broadphase cleanup is small and no longer clears the current release-baseline bar for E2E impact.