[fpedd]

Stacking around pins was quadratic once anything was pinned, and stayed quadratic in the number of gaps between pins. A max tree over the gaps' free widths now finds each free item's first-fit gap in O(log gaps): identical results on 50k fuzzed layouts, 20k items with 10% pins from 9.2 s to 26 ms, and 10k items over 10k gaps from 3 s to 22 ms, with the pin-free path unchanged. BaseAllocator.allocate returns its result in input order as a documented contract and checks all postconditions in one pass, the raw-sequence path no longer wraps the input in temporary pools, and ensure_unique_ids takes a set-size fast path that keeps its TypeError for entities without an id.

An empty pinned pool no longer counts as overlapping another pool (validate, Memory and Pool.overlaps share one half-open overlap helper). Memory.is_allocated now also requires every pool base, so it agrees with Memory.extent; plot_allocation still draws packed pools without a base where Memory.allocate would put them, and still draws pinned pools that overlap (only Memory.allocate rejects those). Saving a pool whose ids collide as text raises, and every pool is checked before any file is written. Unknown registry names raise ValueError from both get and resolve. The matplotlib fallback stubs in visualize are gone, and the primitive tests are flattened into parametrized tables.

Kept on purpose: Memory.used_size (ALTO reads it), Memory/System.any_allocated, Pool.overlaps, available_cores, the unit constants and require_optional's install_extra. Only the old lowest_gap helper is dropped.