[#71] perf: ~2x faster coverage quant (byte-identical) + lossy-input denoise (pigeon 17→12 colours)

[realproject7]

Two optimizations for the Tier 2 coverage engine. Both leave clean (PNG) output byte-identical to 0.4.0 (verified by md5 on logo/gradient/flat).

Perf — ~2× faster coverage quantization, output unchanged

The LAB-bin histogram used np.unique(..., axis=0), whose lexsort dominated runtime on large inputs. Encode each (L,a,b) bin into one integer key (L highest-order; a/b offset into range) and dedup with a 1-D np.unique. The key is monotonic in (L,a,b), so bins/inverse/counts are bit-for-bit identical to the lexsort.

image	0.4.0	0.4.1	output
17-gradient (1024×1536)	4.4s	2.4s	identical
12-sticker	2.6s	1.1s	identical
16-gradient	2.2s	1.1s	identical
pigeon	1.4s	1.0s	identical

Denoise (#71) — cut lossy-input palette inflation

On lossy (JPEG/MPO) sources, compression noise inflates the CIELAB volume so coverage emits near-duplicate colours (pigeon 17 vs the reference's ~10). A stronger flatten paradoxically inflates the palette further (measured), so for lossy inputs the coverage path swaps its default bilateral-flatten for a lighter edge-preserving bilateral (sigma_color=0.02):

• pigeon 17 → 12 colours, outlines stay continuous, SSIM unchanged (0.896 → 0.895), bytes smaller. The JPEG-noise beak artifact is removed.
• Gated strictly to lossy sources — clean PNG/flat inputs never enter it and are byte-identical (the source format is re-attached after load_image's convert() drops it).
• coverage_denoise_lossy=False is the safe-revert.

Note: the denoise bake-off initially favoured an additive bilateral pre-filter, but end-to-end testing showed that is redundant with the existing flatten (no effect in the real pipeline); the working fix is to replace flatten with the lighter bilateral for lossy inputs only.

Verification

• Regression test asserts the gate is load-bearing (clean PNG identical with the option on vs off) + _is_lossy_source classification.
• Full suite green; ruff clean. Comparison render: z-design/.../test-v0.4.0/denoise-proto/pigeon_final_0.4.1.png.

Version 0.4.0 → 0.4.1. Closes #71.