Single-pass patch vectorization (fixes OOM on large-extent rasters)

[NewGraphEnvironment]

Summary

• Fix OOM in dft_transition_vectors() on large-extent rasters: the per-class loop allocated full-grid vectors per class and two full-grid logicals per patch (ncell × n_patches scaling — ~18+ GB extrapolated on the 102.6M-cell UFRA floodplain). Replaced with a single terra::patches(x, directions = 8, values = TRUE) pass plus a sparse patch→label map that touches only the ~2% non-NA cells.
• Output is identical to the old implementation — pinned by a regression test whose reference numbers (185 patches / 123.11 ha / 57 at patch_area_min = 1000) were captured from the old code before the rewrite. Only patch_id numbering / row order changes (now raster scan order).
• terra floored at >= 1.8-10: older versions had an edge-wraparound bug in patches(values = TRUE) that silently merged patches touching opposite raster edges.
• Benchmark at 24M cells: 1.9 s for a 4,799-patch raster vs 122 s for the old code on a milder 1,232-patch raster. Reporter can drop the column-tiled stop-gap from the per-area pipeline driver.
• Follow-up filed as dft_rast_transition builds 6+ full-grid vectors (incl. two character) — same OOM class as #27 #28: dft_rast_transition() has the same disease (6+ full-grid vectors incl. two character vectors) and needs its own refactor.

Fixes #27

Related Issues

• Relates to NewGraphEnvironment/sred-2025-2026#16
• Relates to dft_rast_transition builds 6+ full-grid vectors (incl. two character) — same OOM class as #27 #28 (follow-up: producer-side memory refactor)

Test plan

• devtools::test(): 211 pass, 0 fail — all 10 pre-existing tests unchanged (behavior contract), plus new brute-force decomposition test (engineered topology: diagonal merges, class boundaries, crossed diagonals, code-0 class, holes), all-NA edge case, and fixture regression guard
• lintr clean on touched files (one pre-existing vignette lint, untouched)
• 24M-cell synthetic benchmark against the final implementation: 1.9 s
• /code-check clean on each commit

Notes

Approach was selected empirically during planning — the issue's tiling and sparse-window options benchmarked worse (sparse-window degenerates on mainstem-shaped data where every class's bounding window spans the grid) and as.polygons+explode was disqualified (GDAL polygonize is 4-connected, silently splits diagonal joins). Full analysis in planning/archive/2026-07-issue-27-transition-vectors-oom/.

🤖 Generated with Claude Code