Mechanical cleanup PR. Moves the inline #[cfg(all(test, feature = "std"))] mod tests { ... } blocks out of seven large source files (one per file) into sibling tests.rs files, and folds in two unmerged Copilot review fixes from PR #20. Zero behavior change — pure relocation + two doc/codegen tweaks.

Driven by a long-running file-size concern: src/sinker/mixed.rs had grown to 13,758 lines (of which ~5,900 were inline tests), and each per-arch SIMD backend was in the 5,000–5,700-line range. IDE navigation, code review, and git blame all suffered. Splitting tests into siblings shrinks the production-facing surface by 36% on average across the seven files without changing what runs in CI.

Mechanics

Rust 2024 edition supports the foo.rs + foo/ submodule layout natively. For each source file <dir>/<stem>.rs:

1. The inline #[cfg(all(test, feature = "std"))] mod tests { ... } block is replaced with the one-line declaration #[cfg(all(test, feature = "std"))] mod tests;.
2. The block body moves verbatim to <dir>/<stem>/tests.rs. use super::*; continues to resolve the same way (the new file is still inside the parent module's namespace).
3. #[cfg_attr(miri, ignore = ...)] and other per-test attributes carry through unchanged.

The wasm backend keeps its target_feature = "simd128" cfg gate (#[cfg(all(test, feature = "std", target_feature = "simd128"))] mod tests;).

What's in this PR

Splits (7 files → 7 sibling test files)

Total: 15,402 lines of test code relocated; production-facing surface dropped by ~36% on average across the seven files. Test count unchanged (475 lib tests still discovered + run on aarch64).

Copilot review fixes from PR #20 (folded in)

Two of the four findings on Copilot review #4176623587 were applied to PR #20 locally but never made it into the merge commit. Folded in here since they're trivial and adjacent:

• Finding feat(0.1.0): row-primitive kernels with SIMD dispatch + Sink API (yuv420p, rgb→hsv, bgr↔rgb) #1 — src/sinker/mixed.rs:2591–2601. MixedSinker<Nv42>::with_rgba doc now lists both failure modes (RgbaBufferTooShort and GeometryOverflow), matching the Nv24 doc above it.
• Finding feat(nv21): NV21 (semi-planar 4:2:0, VU-ordered) via const-generic kernel share with NV12 #3 — src/row/scalar.rs:512–528. expand_rgb_to_rgba_row rewritten to iterate via chunks_exact(3) zipped with chunks_exact_mut(4) instead of [x * 3 + k] indexing — tighter codegen on the Strategy A RGB→RGBA fan-out hot path.

What's deferred (still in the cleanup memory)

The other two PR #20 review findings stay deferred to future cleanup PRs because addressing them properly requires a sweep beyond this PR's scope:

• Finding feat(NV12): NV12(semi-planar 4:2:0) + fallible PixelSink contract #2 — visibility tightening on pub(crate) fn nv12_or_nv21_to_rgb_or_rgba_row_impl<...> and pub(crate) fn nv24_or_nv42_to_rgb_or_rgba_row_impl<...> in src/row/scalar.rs, plus the corresponding SIMD pub(crate) unsafe fn *_to_rgb_or_rgba_row_impl functions across all 5 backends. Each is only called from its own file's wrappers — they should all be fn (module-private). Single uniform sweep, no behavior change. Defer because mixing visibility tightening into this PR would obscure the test-mod relocation diff.
• Finding feat(yuv420p10): 10-bit YUV 4:2:0 planar → u8 + native u16 RGB #4 — extract a small rgba_plane_slice helper to remove the duplicated one_plane_end.checked_mul(4)? + one_plane_start * 4 + slice pattern across 8 MixedSinker<F>::process impls (Yuv420p / Yuv422p / Yuv444p / Nv12 / Nv16 / Nv21 / Nv24 / Nv42), in both the standalone-RGBA branch and the Strategy A expand branch. Parallel rgb_plane_slice (3-channel + scratch fallback) helper should ship at the same time. Worth doing before tranches 5–7 (high-bit-depth RGBA) start landing — those will multiply the duplication.

The cleanup memory tracks both of these for follow-up work.

What's NOT in scope here (intentionally)

• No production-code changes beyond the two tiny Copilot fixes above.
• No test logic changes — pure mechanical relocation. No renames, no skips, no coverage shifts.
• No per-arch test helper consolidation (high_bit_plane_*, interleave_uv_*, etc., duplicated across the 5 backend test files). Already noted in docs/color-conversion-functions.md § Cleanup follow-ups; would belong in a sibling refactor PR.
• No reordering or grouping inside the moved test files — they're verbatim from the source files. Reorganizing is a separate concern.

Verification

• cargo test --lib — 475 passed; 0 failed (same as pre-refactor baseline)
• cargo test --doc — 1 passed (the compile_fail doctest on MixedSinker::<Yuv420p>::with_rgba still rejects MixedSinker::<Yuv440p>::with_rgba)
• cargo check --lib --target wasm32-unknown-unknown — clean
• cargo check --lib --target x86_64-unknown-linux-gnu — clean
• cargo clippy --lib --tests — only one pre-existing warning in src/raw/bayer.rs (unrelated unnecessary_cast)

Test plan

• CI green on test, test-sde-avx512, cross, coverage, clippy, build, miri-* jobs.
• Per-tier coverage matrix exercises every backend's tests via colconv_disable_* rustflags (the relocation should be transparent to the matrix).
• Spot-check the new file layout: src/sinker/mixed/tests.rs, src/row/arch/neon/tests.rs, etc. resolve under both cargo test --lib and IDE jump-to-test.

🤖 Generated with Claude Code