[greenclaw]

…keys

LazyHeap was a VecDeque fully re-sorted before front()/pop whenever anything had been inserted, so LFU eviction and VTTL insert/eviction cost O(n log n) per operation (filling an unbounded VTTLCache with 100k entries took ~15s).

Replace it with an indexed binary min-heap:

• each element stores its buffer position, an insertion sequence number (ties keep insertion order, like the old stable sort) and the priority it was placed with;
• push / pop_front / remove are O(log n);
• LFU get() increments the frequency in O(1) without moving the buffer; the root is repaired lazily in front(), which is valid because frequencies only grow;
• arbitrary key changes on replace are queued in O(1) and applied in front(), so replacing a value during iteration still does not move the buffer (generation-version semantics unchanged);
• ordered iteration sorts in place (a sorted buffer is a valid heap).

Bug fix: LFUCache.get() incremented the frequency without invalidating the sorted order, so popitem() could evict the most frequently used key:

c = LFUCache(10)
for k in "abcd": c[k] = k
c.popitem()
for _ in range(10): c.get("b")
c.popitem()  # -> ('b', 'b'), expected ('c', 'c')

Notes:

• VTTL ties (e.g. several entries without ttl) are evicted in insertion order; previously by buffer position.
• ~21 extra bytes per LFU/VTTL entry for heap bookkeeping. The frequency is stored as u64 internally; the public API and pickle format still use the same integers.

[greenclaw]

fixed tests for free-threaded 3.14t & 3.15t

[awolverp]

I will release a new version in about 2-3 hours