[crvernon]

Stacked PR — part of the Demeter v3.0.0 upgrade. Base: v3/packaging-repo. Review and merge in order, bottom of the stack first.

Plan items: 3.4 (partial)

Benchmark: IDENTICAL. Global case 16.9 → 12.1 s (−28%), RSS −17%.

Commits

Run region x metric tasks serially; cap convolution threads at CPU count

The intensification and expansion tasks are Python bound, so the
ThreadPool only added GIL contention. Measured on the global case: 16.2 s
threaded, 11.1 s with serial passes. The pool was also sized to the
number of unique integers in the (region, metric) index pairs rather
than to the number of tasks. Each task modifies only its own region x
metric cells and target_change entries, and since 1.1 no shared input
is mutated, so serial execution gives bit-identical results.

The kernel density convolution keeps its thread pool, because
ndimage.convolve releases the GIL (all serial: 21.9 s). The pool is now
capped at the CPU count instead of one thread per land class.

Benchmark vs v3.0.0-b4 (all IDENTICAL):

• global_0p5deg: 16.9 s -> 12.1 s (-28%), peak RSS -17%.
• global_0p5deg_negcons: 17.1 s -> 11.7 s (-32%).
• texas_0p5deg: 0.22 s -> 0.17 s (-20%).
  The global case has gone from 34.0 s at b0 to 12.1 s (-64%).

🤖 Generated with Claude Code