Replace LRU/MRU/FIFO/LFU with SIEVE eviction

[toloco]

Summary

• Unified eviction: Both in-process (memory) and cross-process (shared) backends now use SIEVE — a simple, scan-resistant algorithm with near-optimal hit rates
• Lock-free shared reads: All shared memory reads are now fully lock-free (previously only FIFO was lock-free; LRU/MRU/LFU required a write lock on every hit for ordering updates). SIEVE's visited bit is set via an idempotent store with no lock
• Simplified API: Removed the strategy parameter and Strategy enum — cache() now just works with SIEVE, no configuration needed
• 540 fewer lines: Deleted src/strategies/ module (5 files), simplified ordering.rs from multi-strategy dispatch to a single sieve_evict() function
• Layout VERSION 2→3: Forces recreation of old mmap files (field renames: _reserved→sieve_hand, frequency→visited)

Files changed (27 files, -540 net lines)

Area	Files	Change
Shared memory core	layout.rs, ordering.rs, mod.rs, region.rs	SIEVE eviction, lock-free reads
Memory core	store.rs, entry.rs, strategies/ (deleted)	Inline SIEVE, remove strategy dispatch
Python API	_decorator.py, _strategies.py, __init__.py, .pyi	Remove strategy param and Strategy enum
Tests	8 test files	Remove strategy params, add SIEVE-specific tests

Test plan

• cargo clippy -- -D warnings — no warnings
• ruff check + ty check — all clean
• pytest tests/ -v — all 72 tests pass
• SIEVE eviction order verified: unvisited entries evicted first, visited entries get second chance
• Shared memory reads are fully lock-free (no write_lock() in get path)
• Cross-process tests pass (visibility, concurrent writers, hashseed independence)

🤖 Generated with Claude Code