Sample magnified nearest planes through cached columns and rows

[timzifer]

What and why

Fixes #34.

Nearest, axis-aligned sampling where u does not depend on y (m[1] = m[2] = 0) now:

• Columns once per image: finds the texel column of every device column once and caches it. The cache and the per-pixel path compute it with the same expressions through shared helpers (origin, nearestCol), so a pixel gets the same texel either way.
• Rows once per texel row: when rows repeat (|m[3]| < 1), expands each texel row once and copies it for every device row on it. No float work per pixel.
• Direct compositing: ImageShader.srcRow returns the expanded row, so coverage spans composite straight from it. Opaque runs at alpha 255 copy it into the destination through the existing opaqueSpan path, without spanOver.
• Unit-scale one-bit planes: bytes that are all one colour (0x00/0xff) are filled 8 pixels at a time.

Cache buffers survive ImageShader.Reset, so there are no allocations after warm-up. Sampler is now documented as not safe for concurrent use.

Not included:

• No exported OpaqueShader interface, since that would change the API. Once cera gets this version it can drop its own bitShader and use ImageShader again.
• Stencils: the mask is now read through the same cache, but the 256-entry stencil LUT is still looked up per pixel.

Benchmarks

BenchmarkMagnifiedBits uses the geometries of cera's scan pages; -count 6, compared with benchstat:

case	before	after
2.1×	2587 µs	926 µs	−64 %
8.3×	517 µs	88 µs	−83 %
7.9×	420 µs	78 µs	−81 %
1×	1803 µs	1043 µs	−42 %

Tests

bitplane_test.go:

• Compares cached sampling with the old per-pixel float mapping for RGBA, index and bit planes. Matrices include mirrored, sheared and rotated ones. Spans come in random order and length.
• Tests the unit-scale bit path at every offset.
• Compares direct and opaque compositing against the same shader hidden behind an interface: antialiased edges, clip, alpha 255 and 140.
• Checks that drawing does not allocate.
• Two deliberately injected bugs both made the new tests fail.

cera's test suite passes with this stilus through a local replace.

Checklist

• go test ./... passes
• gofmt and go vet are clean (also built for js/wasm, arm64, 386)
• Tests added or updated for behaviour changes
• Benchmarks attached for performance-relevant changes (benchstat before/after)
• No new allocations on the drawing path after warm-up
• No new dependencies in the core module

🤖 Generated with Claude Code

https://claude.ai/code/session_01JLyJY68DWZFiofJb5BDadh

---

Generated by Claude Code