Assemble the far field on the device — the build is now faster than before WP2

[TobiBu]

Stacked on #65, which flagged this as its own main cost.

#65 made the update 34–40× faster and the build 21 % slower — 633.8 ms against 524.7, all of
it host-side — because the leaf-major partition does two orderings of 2.8 million entries where
the per-particle expansion did one np.repeat. Every input it needs is already a device array
when assemble_svgd_topology receives it, so none of that had to be on the host.

build at N = 10⁵, float64	device	host	total
per-particle expansion (before #65)	366.5 ms	158.3 ms	524.7 ms
leaf-major, host assembly (#65)	362.5 ms	271.3 ms	633.8 ms
leaf-major, device assembly	364.5 ms	69.6 ms	434.0 ms

The host half falls 3.9× and the whole build 1.46×, leaving it 90.7 ms faster than before
the leaf-major change rather than 109 ms slower. So "never materialise M" is a build-time win
after all — it just needed the assembly to follow the representation onto the device. The update
is untouched (φ 35.11 vs 34.83 ms, φ+∇ 121.60 vs 121.43): this is assembly only.

Two changes that made it simpler, not harder

• The monopole table is keyed by node id, not compacted to the distinct far source nodes.
  That deletes the dense remap, and it makes the table's shape num_nodes — fixed by
  (n, leaf_size) — instead of a data-dependent F, so it no longer churns between rebuilds,
  for the same reason WP6: static capacities — 16x on a sampler that rebuilds every step #64's static capacities mattered. Roughly twice as long, three scalars a row.
• Pairs the policy dropped get a leaf count of zero rather than a boolean filter, because a
  filter has a data-dependent length and would need its own synchronisation. They expand to
  nothing. Exactly one scalar now crosses to the host: the total entry count, which is a
  static shape everything downstream needs anyway.

Why the numpy path is still here

Below 16 384 accepted far pairs the dispatch and the synchronisation cost more than the work they
replace — the same shape of threshold the old expansion had. It is now semantically identical
to the device one, down to the node-id keying, so test_the_device_and_host_far_assemblies_agree
builds both and compares them rather than trusting each alone. Entry order within a leaf is free
(the entries are summed), so that comparison is per-leaf multiset.

test_the_transpose_csr_is_consistent_with_the_forward_one pins far_entry_perm directly — that
it is a permutation, that it groups the source list, and that it matches the node offsets. §6.3 of
the report records why: getting that map backwards leaves every value correct and the gradient
wrong by 88 %, and it slipped through once already.

One guard reads better as a result. With the table keyed by node id it is never empty, and a
capacity="pow2" padding entry names node 0, whose monopole exists and is masked by
far_leaf_live — so the guard is back on the entry list, and the empty-table trap it was working
around cannot recur in this form.

616 tests green. Evidence: reports/YGGDRAX_pallas_nearfield_report.md §6.5.

🤖 Generated with Claude Code