Benchmarks, because "would SIMD help" cannot be asked without a baseline

[tannevaled]

Asked whether go-asmgen or go-simd could make go-pdfkit faster. The
honest first answer is that nobody can tell, and here is why:

	test files	benchmarks
render	43	0
reader	22	0
ops	27	0
extract	10	0
opentype	40	0
go-gfx/gfx	55	13

The only measured corner of the stack is gfx, and the only generated SIMD
anywhere in it is gfx/resample — amd64, arm64 and s390x, from go-asmgen.
go-highway is an indirect dependency of both gfx and render: no code of
ours calls it.

Optimising from here would be choosing where to spend effort by guessing, and
the way that fails is well documented in this fleet — a plateau read as
progress, or two errors cancelling inside a proxy measurement.

Five benchmarks, over the paths where the question lives

	what it exercises
PagePaths	gfx/vector's scanline rasteriser — no SIMD today
PageByScale	the same page at 72, 150 and 300 dpi: does cost follow pixels or operators? That is the first thing worth knowing before moving anything into assembly
PageEmpty	the fixed cost of opening, laying out and allocating, so the floor is visible beside the work in every other benchmark
PageColourChanges	many small fills with a colour change between each — where per-operator cost dominates per-pixel cost

⛔ No numbers are claimed here

This machine is at load average 12–19 with twenty peer sessions on it, and
a number taken under load is a number about the load. The benchmarks exist so
the question can be answered on a quiet machine, with -count=6 and
benchstat, rather than one run against one run.

A smoke run (one iteration each, not a measurement) confirms they execute and
hints at the shape: an empty page costs 2.1 ms of the 3.3 ms a hundred
rectangles cost, so the fixed floor is most of a small page. If that holds when
it is measured properly, the first win is not SIMD at all.

Change to existing code

onePage takes a testing.TB rather than a *testing.T. That is the whole of
it. Coverage is untouched at 100 %.

🤖 Generated with Claude Code