[carsontung666]

Change

Reuse codec allocations and decoded instruction addresses in the default executor. Cached key references accelerate current KVSpace reads, result writes, and PC persistence; small input snapshots are released after each instruction. Preserve cache collision chains across calls/returns and remove redundant frame directory creation.

PC, status, operands, and frame values remain in KVSpace. Reject malformed None-shaped writes without deleting their destination or advancing PC.

Performance

Incremental comparison against previous PR head ddac99ab and SHM backend 703797b9: all 10 repository workloads / 30 frozen scales, 9 alternating baseline/candidate/Python pairs (810 samples), matching outputs. Same Release CLI/layout/frontend, matched -O3 backend flags and allocator, CPU 24, Xeon Gold 6418H, GCC 14.3, Python 3.10.12; fresh stores.

Measurement	Geometric mean speedup
Algorithm execution	1.207x
Whole CLI, including startup/layout	1.021x

Startup/layout and recursive frame work still dominate. This is not a large whole-command gain; Python algorithm timers remain about 85–1670x faster. All samples, medians, hashes, and reproduction command.

Validation

Full SHM and FS tutorials: each 222 passed, 0 failed, 1 extern/WIP skip (FS verified in CI). Runtime 11/11, SHM backend 7/7, frontend 2/2 CTests; 44 old/new frontend/backend compatibility checks. Clang ASan/UBSan runtime tests: 11/11, including malformed writes and live updates through a second handle. Fresh independent agent review passed, including 520 differential reference-write checks.

Native macOS runtime, layout, and runtime-rs builds passed on final head 69309ca; backend macOS build/tests passed on a10d60f.

Write acceleration uses the optional API in kvspace #29 and kvspace-c #42; older libraries retain normal writes. Performance claims apply to SHM; FS receives correctness coverage, and Redis performance is untested.