[#67] Tier 2: region-coverage as the default colour engine (clean outlines + economy)

[realproject7]

Summary

Promotes perceptual-coverage quantization (#65) from a narrow gated path to the default colour engine, by adding a connected-component min-AREA cleanup pass. This fixes the gaps real-world v0.3.0 testing surfaced (#69): gradients inside hard-edged shapes, path over-segmentation, and ragged dark outlines.

The mechanism (one knob, no per-category switch)

After coverage quantization, merge by connected-component area, not global colour mass:

• a gradient band is thin-but-long (large area → kept) → smooth gradients survive, even inside shapes;
• an anti-alias sliver / linework fringe is thin-but-short (sub-area → absorbed into its nearest-ΔE neighbour) → ragged dark linework collapses into clean continuous outlines, and flat art stays economical.

This is why the same pass covers #66 (outline continuity) and #69 (gradient-in-shape + economy) at once.

Validated end-to-end through the real tracer

Renders + metrics in z-design/SVGSmith_Convert/test-v0.3.0/ (montage tier2-before-after.png):

image	0.3.0 (current)	0.4.0 Tier 2	reference
logo	747 paths / 257 KB / banded	34 paths / 17 KB / smooth	33 paths
pigeon	ragged outlines + arm blotches / 125 KB	clean continuous outlines / 30 KB	59 KB
16-gradient (no-regress)	SSIM 0.966	0.968, far fewer paths	—
flat corpus	—	no regression (golden suite green)	—

Principle #1 (clean flats) preserved

• The step stays at the Tier 1 value (ΔE76 4.0) — only the area filter + the widened gate are new. Gradient fidelity is not regressed (16-grad 0.966→0.968; 17-grad 0.877≈0.873).
• Flat corpus goldens stay green; the gradient-bars regression test now asserts the clean economical coverage output.
• SAFE-REVERT is one flag: _TIER2_REGION_COVERAGE = False restores pure Tier 1 (narrow full-frame-gradient gate, no region cleanup). No code path removed; the supersample path remains as the coverage fallback when a trace still explodes (>4000 paths).

Follow-ups (non-blocking)

• JPEG pre-quant denoise to bring the pigeon colour count toward the reference (filed separately).
• Vectorized RAG for the region-merge hot spot (perf, not correctness).
• Per-axis px clamp on the area threshold for 4K+ inputs (already added: coverage_region_max_px).

Full suite green (the lone local test_version_is_single_sourced is a worktree egg-info artifact — clean build metadata is 0.4.0; CI regenerates it). No external-tool brand names.

Closes #66, closes #67, closes #69.