[danielplohmann]

Cuts 5.0.0: __version__ goes from 4.11.0 to 5.0.0. ## [Unreleased] becomes ## [v5.0.0] - 2026-10-16 without its empty subsections (Deprecated, Removed, Security). A fresh ## [Unreleased] skeleton goes above it, and the compare links at the foot are updated. No changelog fixes were needed while cutting.

Major because the release adds an architecture, as v4.0.0 did. The version is a release counter, not semver. None of the three compatibility constants (ESCAPER_DOWNWARD_COMPATIBILITY, INTEL_PIC_HASH_ESCAPE_VERSION, CIL_PIC_HASH_ESCAPE_VERSION) moved since v4.11.0.

What 5.0.0 carries

Milestone v5.0.0, 1 closed before this PR:

• Added: a 32-bit ARM backend for A32 and Thumb/T32 code in ELF, PE (ARMNT, Windows CE), Mach-O armv7 and raw buffers, with the instruction set chosen per function from its strongest evidence (feat: add 32-bit ARM and Thumb disassembly backend #364, which folds in fix(arm): keep an IT block's condition in getDetailed() and the exporter #397).
• Changed: SmdaInstruction and SmdaBasicBlock get __slots__, and per-instruction work is trimmed: callgrind -1.7% to -3.4%, report heap -6% to -9%, reports identical (perf(core): optimize multi-architecture memory layout, CFG recovery, and opcode decode hot paths #404). Instructions and blocks no longer take attributes outside their slots.
• Fixed:
  • An uncomputed block hash stays uncomputed across pickle and deepcopy (perf(core): optimize multi-architecture memory layout, CFG recovery, and opcode decode hot paths #404).
  • capstone's lower bound is declared as >=5.0.2, so pip no longer resolves 4.0.2 and fails at import (fix(build): declare capstone>=5.0.2 #405).
• Compatibility: little-endian 32-bit ARM files and raw buffers the probe reads as ARM now return an arm report instead of an error. Big-endian ARM still errors, now naming armeb (feat: add 32-bit ARM and Thumb disassembly backend #364).

Not in this release: #355 (reusable release workflow) and #297 (capstone 6 port).

Checked before opening

• feat: add 32-bit ARM and Thumb disassembly backend #364 at ecdfbb0: lld-fill and scan-vs-blx blocker repros fixed on six clang/lld builds; x86/x64/AArch64 reports byte-identical to master; full suite 2559 passed, 2 skipped.
• perf(core): optimize multi-architecture memory layout, CFG recovery, and opcode decode hot paths #404 merged onto 8f15a1a: 19 ARM reports and 6 non-ARM reports byte-identical to master; ARM pickle, deepcopy and fromDict round-trips unchanged; full suite 2561 passed, 2 skipped.
• fix(build): declare capstone>=5.0.2 #405 merged with master: full suite 2561 passed, 2 skipped on capstone 5.0.2, the new floor.
• testPackageMetadata.py, testReleaseGuard.py and testCommonModels.py pass on this branch. release_guard.py --tag v5.0.0 passes and extracts the 37-line notes section.
• A local python -m build produced a 435 KB sdist with no tests/ and a 474 KB wheel with py.typed and the smda/arm/ package. twine check --strict passed, and the metadata reads Version: 5.0.0 and Requires-Dist: capstone<6,>=5.0.2.

After merge

publish-release.yml and release_guard.py are unchanged since v4.11.0, so no TestPyPI rehearsal is needed. Once CI and the benchmark are green on the merge commit, tag it and push the tag:

git tag -a v5.0.0 -m "SMDA 5.0.0"
git push origin v5.0.0

🤖 Generated with Claude Code

[github-actions]

📊 SMDA Performance Evaluation Benchmark Results

Generated on: 2026-10-16 13:51:58

Summary

Metric	Result
Correctness	✅ PASS — 0 / 155 common file(s) differ
Determinism	✅ PASS — base 3 run(s), PR 3 run(s)
Verdict	inconclusive — within run-to-run noise (±1.7%)
Median paired speedup	-0.35% (95% CI [-0.59%, +0.15%])
Timing noise band	±1.7%

Per-side Timing Context (best of 3/3 runs per file)

Side	Files	Functions	Median best time/file (s)	Sum of best times (s)	Throughput estimate (func/s)
base	155	173540	0.6373	226.06	~768
pr	155	173540	0.6591	226.27	~767

These context rows sum each file's best observed time across repeated runs; they are normalized comparison estimates, not single-run CI wall-clock times.

Paired per-file timing (positive speedup = PR faster):

Statistic	Value
Files compared	155
Median paired speedup	-0.35% (95% CI [-0.59%, +0.15%])
Mean speedup	-0.90% (95% CI [-1.62%, -0.19%])
Std dev / IQR	4.52% / 2.78%
Wilcoxon signed-rank p	0.2108 (n=155)

ℹ️ base and PR are timed in the same job on the same runner, with their passes interleaved base/PR/PR/base/base/PR so that drift over the job cancels rather than landing on whichever side ran last. The noise band above is therefore run-to-run variation on one machine, which is what it is able to measure; a difference inside it is reported as inconclusive. Correctness and determinism are not timing-based and are unaffected either way.

Determinism (self-check across repeated runs)

Side	Runs	Files	Deterministic	Median timing CV
base	3	155	✅	1.7%
pr	3	155	✅	1.2%

Pairwise run matrix (individual run medians, diagnostic)

Diagnostic only: each row compares one raw PR run against one raw base run. The headline verdict above uses paired per-file best-of-runs timings.

Comparison Run	PR Files	Base Files	Common	Function Set Matches	Med PR (s)	Med Base (s)	Med Diff (s)	Speedup %
pr_0 vs base_0	155	155	155	155/155	0.6726s	0.6931s	-0.0205s	+2.96%
pr_0 vs base_1	155	155	155	155/155	0.6726s	0.7042s	-0.0316s	+4.49%
pr_0 vs base_2	155	155	155	155/155	0.6726s	0.6820s	-0.0094s	+1.39%
pr_1 vs base_0	155	155	155	155/155	0.6878s	0.6931s	-0.0053s	+0.76%
pr_1 vs base_1	155	155	155	155/155	0.6878s	0.7042s	-0.0164s	+2.33%
pr_1 vs base_2	155	155	155	155/155	0.6878s	0.6820s	+0.0058s	-0.84%
pr_2 vs base_0	155	155	155	155/155	0.6800s	0.6931s	-0.0132s	+1.90%
pr_2 vs base_1	155	155	155	155/155	0.6800s	0.7042s	-0.0242s	+3.44%
pr_2 vs base_2	155	155	155	155/155	0.6800s	0.6820s	-0.0021s	+0.31%