[DanielLuning]

What changed

Owner decision 2026-10-26: make the layout noise visible first (this PR); the -falign-* experiment is the follow-up defk-w8r.4, measured with this band.

• bench/layout_noise.exs: builds one module of five kernels 8 times with only the code addresses moved. A CC wrapper per build passes -include of a header with an unused function of 0-48 bytes, so every function after it moves by that many bytes (checked with nm -n on the built .so files; the C differs only in the module name in ERL_NIF_INIT, which is data). The first build is also measured under two more labels as an A/A control. Interleaved rounds, the order rotated each round, best of 3 per label per round; it prints per-label medians, the layout range, the same-build range and count_primes' round-bests.
• docs/performance.md: new entry 1.10 with the numbers; 1.7 points to it and gets the rule (a difference inside a kernel's layout range can be layout alone); 1.7's source line no longer calls defk-9p6.24 open.

Results

Quiet machine (other processes 0.65-0.67 cores), 15 rounds, one run:

kernel	layout range (8 builds)	same-build range (3 labels)
count_primes(300_000)	3.6%	0.8%
count_primes_via_call(300_000)	2.9%	0.2%
count_primes_early(1_000_000)	3.6%	2.1%
sum, 1M [f64]	1.9%	0.8%
dot, 2 x 1M [f64]	1.2%	1.2%

The integer loops move with layout well beyond their control (pad 40 made count_primes ~3% faster than pad 0 in 10 of 11 rounds outside a slow stretch); the float loops do not clearly. Rounds 8-11 were 8-10% slower for every build at once; the interleaving spreads that over all builds. A first 9-round run without the control gave the same picture (layout ranges 0.4-4.1%).

How it was checked

mix ci on an M-series Mac in the worktree: 2326 tests passed. No test runs the bench scripts; the script was smoke-run and then run in full.

• Touches a promised surface? No: a bench script and the maintainers' performance record, so no CHANGELOG.md entry.
• Touches test/stability/? No.