perf: rhealpixdggs 0.8.5 hierarchical polyfill, with a per-geometry min_res fix

[alpha-beta-soup]

Summary

Bumps rHP's rhealpixdggs dependency to ^0.8.5, which rewrote polyfill as a hierarchical descent from a min_res parameter (default 0). Adopting it naively is dramatic in both directions:

• Large/complex polygons: 7x-118x faster than the pre-0.8.5 algorithm on rhealpixdggs's own NZ-scale reference polygon, and reaches resolutions (3.4M cells in 6.6s) the old algorithm couldn't practically reach.
• Typical real-world small polygons: a ~900x regression at the library's default min_res=0 — tests/data/chathams-pannz-2014.gpkg (159 real features, mostly small parcels) went from 0.38s to 345s for identical output.

The fix

RHPVectorIndexer._fill_min_res chooses min_res per geometry from its own bounding-box extent, reusing the existing DGGS_CELL_AREA_M2_BY_RES table (rHP cells are equal-area squares, so width = √area — confirmed identical to WGS84_003.cell_width() to the last decimal). With this, the same 159-feature fixture at resolution 9 runs in 1.05s — faster than the old algorithm's 1.68s, output bit-for-bit identical.

min_res is not a correctness parameter (rhealpixdggs's own tests assert output is min_res-invariant), so no fudge-factor margin is used. The one exception: a geometry touching a pole or spanning all 360° of longitude gets the whole polar square as its candidate region regardless of size — confirmed a tiny pole-touching bbox produces a ~42 million cell starting frontier at min_res=8 — so those two cases are explicitly clamped to min_res=0.

Also in this PR

• Switched _polyfill_polygon to polyfill_array, which also fixes a latent nondeterminism: the old set()-based path's row order depended on Python's string-hash randomisation across runs; polyfill_array's output is documented-sorted.
• Fixed the compact_cells import, which moved from rhealpixdggs.rhp_wrappers to rhealpixdggs.conversion between 0.8.3 and 0.8.5 (rhp_wrappers no longer re-exports it) — a genuine break unrelated to polyfill, found while testing.

New tests (tests/classes/rhp_fill.py)

• Unit tests for _fill_min_res, including the pole/antimeridian clamp
• Correctness/equivalence tests against min_res=0 and min_res=resolution (the pre-0.8.5 algorithm's own semantics)
• Performance-regression guards: a mocked "min_res is actually passed" check (zero timing flakiness) plus generous wall-clock budgets using the real fixture

Testing

This sandbox can't install packages, so verification used PYTHONPATH-shadowing the pure-Python 0.8.5 source (no compiled extensions, confirmed loads cleanly ahead of the locally-installed 0.8.3):

• black, ruff check ., mypy vector2dggs/ — all clean
• Full suite passes under the shadow (only pre-existing, unrelated psycopg2-missing failures)
• Confirmed the same suite gracefully skips (not crashes) rHP tests against the still-locked 0.8.3, proving the version bump is load-bearing
• poetry.lock relocked successfully; the 0.8.5 wheel hash matches raster2dggs's own lock for the same version independently

Please run the full suite once more against the real installed 0.8.5 wheel (poetry install -E all --with dev) before merging, as final confirmation against the published artifact rather than the source-shadowed checkout.

🤖 Generated with Claude Code