chore: cut v4.6.0

[danielplohmann]

Cuts v4.6.0 over what has landed since v4.5.1 (4868e19): #300 (50d3289), #325 (e9dfe5f) and #299 (46adf4b).

Ten engine changes under feat(core) plus a hot-path pass is a minor-bump shape, which is the reason #299 and #300 were re-scoped off v4.5.1 rather than squeezed into it. Nothing else is queued for this release.

Format

Deliberately the current format, not keep-a-changelog. #323 proposed adopting against an empty Unreleased so the first entries it holds are written by their authors at merge time; v4.6.0 cannot be that, because #299's and #300's entries are being written here at release time either way. Adopting after this release means the inaugural section fills itself from the seven PRs currently open, which is the demonstration #323 was after. Four of @r0ny123's additions are also still unsettled, one of which (reusing the semantic-pr-title.yml scope list as the entry prefix vocabulary) would rewrite every heading below.

What this entry does adopt is the compromise: each topic keeps the mechanism and its headline figure with the cost, and links the PR that carries the full measurement, the per-rule attribution and the dead ends. It is ~2,200 words against v4.5.1's ~2,790, so the depth that matters at release time is here and the rest is one click away. This is the first entry in the file to use links at all.

What the entry covers

• Recovery — believe the binary over the bytes (feat(core): function-boundary accuracy from what the image declares #300), as seven short items: the header read, the declared PE exception-table address (419 → 626 on a ReadyToRun image), the prologue-inside-prologue refusal (912 FP removed, 0 TP lost), the endbr64 seeding refusal, the AArch64 noreturn-call boundary, the pre-analysis candidate snapshot, and the bl fall-through seed moving behind RESOLVE_TAILCALLS — including the −580 FP / −53 TP row and why the gate is kept.
• New (default on) (feat(core): function-boundary accuracy from what the image declares #300) — USE_LSDA_LANDING_PADS and USE_ELF_FDE_INTERIOR_GAPS, each with the guard that makes it safe and the cost measured without it (the PLT exemption's 3,457 real functions; the FDE-starts-in-padding 35), plus the combined endbr64 result: 822 false positives to zero.
• Measured — five corpus rows plus the 260-cell matrix; the three bit-identical corpora are named as the control in prose rather than given rows.
• Robustness, Performance (perf: cut analysis hot-path overhead without changing output #299), three defects a report-identity check could not see (perf: cut analysis hot-path overhead without changing output #299, including the updateFunctionGaps() revert, 6,249 → 5,247 on a 32-bit MSVC PE), two moved fixture baselines, Compatibility, new fixtures, Housekeeping (feat(core): function-boundary accuracy from what the image declares #300, chore: correct the figures and the directory lookup #300 shipped #325).

Compatibility

No escaper output changed. The only escaping-related edit memoises escapeBinary results behind a cache keyed on its complete input tuple, and no escaper module is touched, so ESCAPER_DOWNWARD_COMPATIBILITY stays at 4.4.5 and INTEL_PIC_HASH_ESCAPE_VERSION at 4.3.5. No report needs reprocessing.

Provenance of the figures

The contributor's own except the ARM64 Mach-O row and the ReadyToRun result, which were reproduced during the #300 review. The entry says which is which rather than presenting one table as equally checked — the same distinction v4.5.1 made.

Second commit

docs(docs) fixes AGENTS.md:117, which has prescribed a one-line changelog entry since before v4.5.1 while v4.5.1 and v4.6.0 both use the nested **Topic:** list. Independent of #323, which would replace that line again on adoption — raised in the #299/#300 review pass and never ruled on, so it is folded in here rather than sent round as a one-line PR of its own.

Checks run locally

• pytest tests/test* — 2094 passed, 2 skipped, 2596 subtests, in 159s.
• ruff check . and ruff format --check . clean.
• ty check src/smda/ fuzzing/ profiling/ .github/workflows/scripts/ — exit 0, unchanged against master.
• Master's own post-perf: cut analysis hot-path overhead without changing output #299 CI is green on all four workflows, benchmark included (run 34521599914).

🤖 Generated with Claude Code

[github-actions]

📊 SMDA Performance Evaluation Benchmark Results

Generated on: 2026-09-10 20:49:05

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 (±4.4%)
Median paired speedup	-0.76% (95% CI [-1.29%, -0.26%])
Timing noise band	±4.4%

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	173464	0.5263	214.49	~809
pr	155	173464	0.5589	216.85	~800

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	-0.76% (95% CI [-1.29%, -0.26%])
Mean speedup	-1.75% (95% CI [-3.83%, +0.17%])
Std dev / IQR	12.97% / 5.13%
Wilcoxon signed-rank p	0.0449 (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	✅	4.4%
pr	3	155	✅	4.2%

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	0.6560s	0.5841s	+0.0719s	-12.30%
pr_0 vs base_1	155	155	155	155/155	0.6560s	0.7126s	-0.0567s	+7.95%
pr_0 vs base_2	155	155	155	155/155	0.6560s	0.6292s	+0.0268s	-4.25%
pr_1 vs base_0	155	155	155	155/155	0.6517s	0.5841s	+0.0676s	-11.57%
pr_1 vs base_1	155	155	155	155/155	0.6517s	0.7126s	-0.0609s	+8.55%
pr_1 vs base_2	155	155	155	155/155	0.6517s	0.6292s	+0.0225s	-3.58%
pr_2 vs base_0	155	155	155	155/155	0.6227s	0.5841s	+0.0386s	-6.61%
pr_2 vs base_1	155	155	155	155/155	0.6227s	0.7126s	-0.0899s	+12.62%
pr_2 vs base_2	155	155	155	155/155	0.6227s	0.6292s	-0.0065s	+1.03%